{
 "schema_version": "1.3",
 "kind": "hardware-roster-rows",
 "figure_definition": "v3",
 "note": "One record per figure this shelf published for a piece of hardware, with the conditions its page states and proved against that page's own row by tools/hardware_gate.py: 839 of these 1054 were read off a published table by tools/hardware_extract.py and 215 were carried across by hand. A condition is part of a row's identity, never a heading above one: the same card under two caps is two rows. Nothing here is normalised, ranked or averaged, and no ratio is computed — a ratio an exhibit printed is quoted like any other figure.",
 "rows": [
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-34-13-tokens-a-second",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "34.13 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    16821,
    16918
   ],
   "quote": "| `mistral-small3.2:24b`, 24B **dense**, decode | 34.13 tokens a second | 92.77 tokens a second |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX 3090 (24 GB) — a bench instance"
   },
   "conditions_cites": [],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-a-power-cap-actually-costs"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/decode-tokens-a-second-92-77-tokens-a-second",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "92.77 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    16821,
    16918
   ],
   "quote": "| `mistral-small3.2:24b`, 24B **dense**, decode | 34.13 tokens a second | 92.77 tokens a second |",
   "identity_basis": "column-head",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX PRO 6000 Blackwell (96 GB) — the served seat"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-126-12-tokens-a-second",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "126.12 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    16919,
    17006
   ],
   "quote": "| `gemma4:26b`, 26B **MoE**, decode | 126.12 tokens a second | 205.85 tokens a second |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX 3090 (24 GB) — a bench instance"
   },
   "conditions_cites": [],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-a-power-cap-actually-costs"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/decode-tokens-a-second-205-85-tokens-a-second",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "205.85 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    16919,
    17006
   ],
   "quote": "| `gemma4:26b`, 26B **MoE**, decode | 126.12 tokens a second | 205.85 tokens a second |",
   "identity_basis": "column-head",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX PRO 6000 Blackwell (96 GB) — the served seat"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/watts-while-decoding-dense-moe-228-2-w-206-4-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "watts while decoding, dense / MoE",
   "value": "228.2 W / 206.4 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17007,
    17084
   ],
   "quote": "| Watts while decoding, dense / MoE | 228.2 W / 206.4 W | 328.3 W / 182.5 W |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX 3090 (24 GB) — a bench instance"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/watts-while-decoding-dense-moe-328-3-w-182-5-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "watts while decoding, dense / MoE",
   "value": "328.3 W / 182.5 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17007,
    17084
   ],
   "quote": "| Watts while decoding, dense / MoE | 228.2 W / 206.4 W | 328.3 W / 182.5 W |",
   "identity_basis": "column-head",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX PRO 6000 Blackwell (96 GB) — the served seat"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-dense-moe-6-680-j-1-632-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens, dense / MoE",
   "value": "6,680 J / 1,632 J",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17085,
    17163
   ],
   "quote": "| Energy per 1,000 tokens, dense / MoE | 6,680 J / 1,632 J | 3,538 J / 887 J |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX 3090 (24 GB) — a bench instance"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/energy-per-1-000-tokens-dense-moe-3-538-j-887-j",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens, dense / MoE",
   "value": "3,538 J / 887 J",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17085,
    17163
   ],
   "quote": "| Energy per 1,000 tokens, dense / MoE | 6,680 J / 1,632 J | 3,538 J / 887 J |",
   "identity_basis": "column-head",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX PRO 6000 Blackwell (96 GB) — the served seat"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/first-token-on-a-prompt-it-has-already-seen-dense-moe-98-ms-167-ms",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token on a prompt it has already seen, dense / MoE",
   "value": "98 ms / 167 ms",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17164,
    17258
   ],
   "quote": "| First token on a prompt it has already seen, dense / MoE | 98 ms / 167 ms | 28 ms / 162 ms |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX 3090 (24 GB) — a bench instance"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/first-token-on-a-prompt-it-has-already-seen-dense-moe-28-ms-162-ms",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token on a prompt it has already seen, dense / MoE",
   "value": "28 ms / 162 ms",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b, gemma4:26b",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "the small hours of 2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    17164,
    17258
   ],
   "quote": "| First token on a prompt it has already seen, dense / MoE | 98 ms / 167 ms | 28 ms / 162 ms |",
   "identity_basis": "column-head",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16705,
    16806
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16705,
     16806
    ],
    "header": "| Measured | RTX 3090 (24 GB) — a bench instance | RTX PRO 6000 Blackwell (96 GB) — the served seat |",
    "head_cell": "RTX PRO 6000 Blackwell (96 GB) — the served seat"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-34-13",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "34.13",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19571,
    19635
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 250 W | 34.13 | — | 6,680 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-we-measured-on-two-cards-in-one-box"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-against-the-250-w-rung",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "decode against the 250 W rung",
   "value": "—",
   "non_figure": "—",
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19571,
    19635
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 250 W | 34.13 | — | 6,680 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 3,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Vs 250 W"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the baseline rung — the page prints an em dash, not a figure",
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-6-680-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens",
   "value": "6,680 J",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19571,
    19635
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 250 W | 34.13 | — | 6,680 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 4,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Energy per 1,000 tokens"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-48-18",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "48.18",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19636,
    19705
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 300 W | 48.18 | +41.2% | 5,673 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-against-the-250-w-rung-41-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode against the 250 W rung",
   "value": "+41.2%",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19636,
    19705
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 300 W | 48.18 | +41.2% | 5,673 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 3,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Vs 250 W"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-5-673-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens",
   "value": "5,673 J",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19636,
    19705
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 300 W | 48.18 | +41.2% | 5,673 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 4,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Energy per 1,000 tokens"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-50-02",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "50.02",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19706,
    19797
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 350 W (the board's default) | 50.02 | +46.6% | 5,318 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-against-the-250-w-rung-46-6",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode against the 250 W rung",
   "value": "+46.6%",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19706,
    19797
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 350 W (the board's default) | 50.02 | +46.6% | 5,318 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 3,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Vs 250 W"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-5-318-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens",
   "value": "5,318 J",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19706,
    19797
   ],
   "quote": "| `mistral-small3.2:24b` (dense) | 350 W (the board's default) | 50.02 | +46.6% | 5,318 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 4,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Energy per 1,000 tokens"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-126-12",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "126.12",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19798,
    19851
   ],
   "quote": "| `gemma4:26b` (MoE) | 250 W | 126.12 | — | 1,632 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-we-measured-on-two-cards-in-one-box"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-1-632-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens",
   "value": "1,632 J",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19798,
    19851
   ],
   "quote": "| `gemma4:26b` (MoE) | 250 W | 126.12 | — | 1,632 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 4,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Energy per 1,000 tokens"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-136-49",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "136.49",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19852,
    19931
   ],
   "quote": "| `gemma4:26b` (MoE) | 350 W (the board's default) | 136.49 | +8.2% | 1,823 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-against-the-250-w-rung-8-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode against the 250 W rung",
   "value": "+8.2%",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19852,
    19931
   ],
   "quote": "| `gemma4:26b` (MoE) | 350 W (the board's default) | 136.49 | +8.2% | 1,823 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 3,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Vs 250 W"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/energy-per-1-000-tokens-1-823-j",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy per 1,000 tokens",
   "value": "1,823 J",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "350 W (the board's default)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19852,
    19931
   ],
   "quote": "| `gemma4:26b` (MoE) | 350 W (the board's default) | 136.49 | +8.2% | 1,823 J |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 4,
    "header_span": [
     19472,
     19548
    ],
    "header": "| Model | Power cap | Tokens a second | Vs 250 W | Energy per 1,000 tokens |",
    "head_cell": "Energy per 1,000 tokens"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/the-fraction-of-a-122b-model-on-card-and-aggregate-decode-seats-up-39-5",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the fraction of a 122B model on card, and aggregate decode, seats up",
   "value": "39.5%",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:122b, 10 billion active per token, 75.7 GiB",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "a 3090 at 250 W, the workstation card at 420 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    20684,
    20762
   ],
   "quote": "| The workshop's seats up | 39.5% | 38.22 | `q1-split-resident-seatsup.json` |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20598,
     20665
    ],
    "header": "| What the cards could offer | On card | Tokens a second | Source |",
    "head_cell": "On card"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "split across the 3090 and the workstation card; the page's own words",
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-split-seats-up-38-22",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second, split, seats up",
   "value": "38.22",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:122b, 10 billion active per token, 75.7 GiB",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "a 3090 at 250 W, the workstation card at 420 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    20684,
    20762
   ],
   "quote": "| The workshop's seats up | 39.5% | 38.22 | `q1-split-resident-seatsup.json` |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20598,
     20665
    ],
    "header": "| What the cards could offer | On card | Tokens a second | Source |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/decode-tokens-a-second-split-seats-stopped-71-48",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second, split, seats stopped",
   "value": "71.48",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:122b, 10 billion active per token, 75.7 GiB",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "a 3090 at 250 W, the workstation card at 420 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    20763,
    20866
   ],
   "quote": "| The seats stopped, inside one authorised window | 89.6% | 71.48 | `q1-split-resident-seatsoff.json` |",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20598,
     20665
    ],
    "header": "| What the cards could offer | On card | Tokens a second | Source |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/a-122b-model-on-this-card-alone-seats-stopped-refused-does-not-fit",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "a 122B model on this card alone, seats stopped",
   "value": "*refused — does not fit*",
   "non_figure": "refused — does not fit",
   "conditions": {
    "model": "qwen3.5:122b, 10 billion active per token, 75.7 GiB",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    20867,
    20964
   ],
   "quote": "| The 96 GB card alone, seats stopped | 63.0% | *refused — does not fit* | `q1-6000-whole.json` |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB) — the served seat",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16756,
    16804
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20598,
     20665
    ],
    "header": "| What the cards could offer | On card | Tokens a second | Source |",
    "head_cell": "Tokens a second"
   },
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own refusal, carried as a refusal: 63.0% of the model would board and the rest would not",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/writing-speed-single-warm-calls-95-0-94-6-93-5-tokens-a-second",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, single warm calls",
   "value": "95.0 / 94.6 / 93.5 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 / 500 / 450 W",
    "context": "not stated",
    "batch_or_steps": "128 tokens out",
    "concurrency": "not stated",
    "window": "2026-08-27"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-does-it-do-on-our-own-machines",
    "md_sha": "6044b73dbda82eeb2a7e6ddc6afcd95987ef6ec960382d6acee30077759e9452"
   },
   "cite": "what-does-it-do-on-our-own-machines",
   "span": [
    13230,
    13436
   ],
   "quote": "| Writing speed on an RTX PRO 6000 Blackwell (96 GB), single warm calls, 128 tokens out | 95.0 / 94.6 / 93.5 tokens a second | 600 / 500 / 450 W | 2026-08-27 · [what 150 watts buys](/what-150-watts-buys/) |",
   "identity_basis": "row",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB)",
   "identity_cite": "what-does-it-do-on-our-own-machines",
   "identity_span": [
    13230,
    13436
   ],
   "cell": {
    "column": 1,
    "header_span": [
     13118,
     13211
    ],
    "header": "| What was measured, and on which model | Reading | Power cap | When, and where to check it |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/mean-draw-and-peak-temperature-594-6-w-mean-across-its-first-four-minute-leg-87-",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw and peak temperature",
   "value": "594.6 W mean across its first four-minute leg, 87 °C peak across the eight minutes",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "uncapped in that arm",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-does-it-do-on-our-own-machines",
    "md_sha": "6044b73dbda82eeb2a7e6ddc6afcd95987ef6ec960382d6acee30077759e9452"
   },
   "cite": "what-does-it-do-on-our-own-machines",
   "span": [
    13437,
    13709
   ],
   "quote": "| Power and heat under its own load, same card | 594.6 W mean across its first four-minute leg, 87 °C peak across the eight minutes — the hottest arm of the serving bench it was read from | Uncapped in that arm | 2026-08-25 · [what 150 watts buys](/what-150-watts-buys/) |",
   "identity_basis": "section",
   "identity_token": "RTX PRO 6000 Blackwell (96 GB)",
   "identity_cite": "what-does-it-do-on-our-own-machines",
   "identity_span": [
    13232,
    13282
   ],
   "cell": {
    "column": 1,
    "header_span": [
     13118,
     13211
    ],
    "header": "| What was measured, and on which model | Reading | Power cap | When, and where to check it |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/a-short-history-of-mistral/decode-split-with-system-memory-against-cpu-alone-5-13-tokens-a-second-split-aga",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode split with system memory, against CPU alone",
   "value": "5.13 tokens a second split, against 2.71 on that box's processor alone — 1.89×",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "cap not set by that harness",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-10"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-does-it-do-on-our-own-machines",
    "md_sha": "6044b73dbda82eeb2a7e6ddc6afcd95987ef6ec960382d6acee30077759e9452"
   },
   "cite": "what-does-it-do-on-our-own-machines",
   "span": [
    14385,
    14666
   ],
   "quote": "| On a used consumer card too small to hold it, half in RAM | 5.13 tokens a second split, against 2.71 on that box's processor alone — 1.89× | Cap not set by that harness | 2026-09-10 · this workshop's own bench record, `gpu-3080-2026-09-10/RESULTS.md` (not yet a published page) |",
   "identity_basis": "row",
   "identity_token": "a used consumer card too small to hold it",
   "identity_cite": "what-does-it-do-on-our-own-machines",
   "identity_span": [
    14385,
    14666
   ],
   "cell": {
    "column": 1,
    "header_span": [
     13118,
     13211
    ],
    "header": "| What was measured, and on which model | Reading | Power cap | When, and where to check it |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-we-measured-on-two-cards-in-one-box"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/draw-while-decoding-and-the-clock-228-2-w-91-of-cap-clock-pulled-to-495-675-mhz",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "draw while decoding, and the clock",
   "value": "228.2 W — 91% of cap — clock pulled to **495–675 MHz**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b (dense)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19139,
    19246
   ],
   "quote": "The dense model sat at 228.2 W — 91% of cap — clock pulled to **495–675 MHz** across three runs at 64–66 °C",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": null,
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/draw-while-decoding-206-4-w",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "draw while decoding",
   "value": "206.4 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    19287,
    19330
   ],
   "quote": "The MoE drew 206.4 W, 82.6% of the same cap",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": null,
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/aggregate-decode-at-one-two-and-four-streams-127-1-195-8-261-4-tokens-a-second",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "aggregate decode at one, two and four streams",
   "value": "**127.1 / 195.8 / 261.4 tokens a second**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one, two and four simultaneous streams",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    21302,
    21377
   ],
   "quote": "produced **127.1 / 195.8 / 261.4 tokens a second** aggregate at a 250 W cap",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": null,
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/the-embedding-arm-s-spread-against-the-gate-it-was-pre-registered-under-five-bat",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the embedding arm's spread, against the gate it was pre-registered under",
   "value": "five batches spread 34.9% around their median where the gate allowed 15%",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "one ollama build",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-16 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-a-power-cap-actually-costs",
    "md_sha": "6abca9f23bbe871379ff14ca66add4f3ae0acbf57d7b82816e605670981743a8"
   },
   "cite": "what-a-power-cap-actually-costs",
   "span": [
    21611,
    21683
   ],
   "quote": "five batches spread 34.9% around their median where the gate allowed 15%",
   "identity_basis": "page",
   "identity_token": "RTX 3090 (24 GB) — a bench instance",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    16718,
    16753
   ],
   "cell": null,
   "conditions_cites": [
    "what-we-measured-on-two-cards-in-one-box"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the arm voided itself against its own pre-registered gate, and the page's own rule is that a rung rejected by its own gate gets no ratio — so this shelf carries the spread and no ratio from it",
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/dollars-per-million-output-tokens-electricity-only-0-29-per-million",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "dollars per million output tokens, electricity only",
   "value": "**$0.29 per million**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "18.34 cents a kilowatt-hour, the US residential average for June 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "is-it-cheaper-to-run-it-yourself",
    "md_sha": "17782b318f6760ba14d5ae1cbe25ee2919ebe78efa16affd72e2dd0a243aac42"
   },
   "cite": "is-it-cheaper-to-run-it-yourself",
   "span": [
    22401,
    22517
   ],
   "quote": "**5,673 J per 1,000 tokens** (`capladder-mistral-small3.2-24b.json`) → 1.576 kWh per million → **$0.29 per million**",
   "identity_basis": "section",
   "identity_token": "an EVGA XC3 Ultra",
   "identity_cite": "is-it-cheaper-to-run-it-yourself",
   "identity_span": [
    24528,
    24545
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-short-history-of-mistral/dollars-per-million-output-tokens-electricity-only-0-045-per-million",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "dollars per million output tokens, electricity only",
   "value": "**$0.045 per million**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "420 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "18.34 cents a kilowatt-hour, the US residential average for June 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "is-it-cheaper-to-run-it-yourself",
    "md_sha": "17782b318f6760ba14d5ae1cbe25ee2919ebe78efa16affd72e2dd0a243aac42"
   },
   "cite": "is-it-cheaper-to-run-it-yourself",
   "span": [
    22568,
    22665
   ],
   "quote": "**887 J per 1,000 tokens** (`m5-6000-live.json`) → 0.246 kWh per million → **$0.045 per million**",
   "identity_basis": "section",
   "identity_token": "96 GB workstation card",
   "identity_cite": "is-it-cheaper-to-run-it-yourself",
   "identity_span": [
    25008,
    25030
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/dollars-per-token-a-second-at-what-was-paid-11-50-per-token-a-second",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "dollars per token a second, at what was paid",
   "value": "**$11.50 per token a second**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "bought for $1,449.99 in May 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "is-it-cheaper-to-run-it-yourself",
    "md_sha": "17782b318f6760ba14d5ae1cbe25ee2919ebe78efa16affd72e2dd0a243aac42"
   },
   "cite": "is-it-cheaper-to-run-it-yourself",
   "span": [
    24864,
    24919
   ],
   "quote": "**$11.50 per token a second** on the MoE (126.12 tok/s)",
   "identity_basis": "section",
   "identity_token": "an EVGA XC3 Ultra",
   "identity_cite": "is-it-cheaper-to-run-it-yourself",
   "identity_span": [
    24528,
    24545
   ],
   "cell": null,
   "conditions_cites": [
    "what-a-power-cap-actually-costs"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-short-history-of-mistral/dollars-per-token-a-second-at-what-was-paid-30-10",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "dollars per token a second, at what was paid",
   "value": "**$30.10**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "bought for $1,449.99 in May 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "is-it-cheaper-to-run-it-yourself",
    "md_sha": "17782b318f6760ba14d5ae1cbe25ee2919ebe78efa16affd72e2dd0a243aac42"
   },
   "cite": "is-it-cheaper-to-run-it-yourself",
   "span": [
    24924,
    24978
   ],
   "quote": "**$30.10** on the dense Mistral at 300 W (48.18 tok/s)",
   "identity_basis": "section",
   "identity_token": "an EVGA XC3 Ultra",
   "identity_cite": "is-it-cheaper-to-run-it-yourself",
   "identity_span": [
    24528,
    24545
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/a-short-history-of-mistral/decode-tokens-a-second-split-with-system-memory-34-40-tokens-a-second",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, tokens a second, split with system memory",
   "value": "34.40 tokens a second",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    18800,
    18869
   ],
   "quote": "on a 3080 split with system memory, 34.40 tokens a second for the MoE",
   "identity_basis": "row",
   "identity_token": "on a 3080 split with system memory",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    18800,
    18869
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a control, not a like-for-like leg — the page says so in the same sentence, and dates it nowhere",
   "record_published": null
  },
  {
   "id": "rtx-3080/a-short-history-of-mistral/decode-tokens-a-second-split-with-system-memory-5-13",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, tokens a second, split with system memory",
   "value": "5.13",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-10"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-short-history-of-mistral",
    "section": "what-we-measured-on-two-cards-in-one-box",
    "md_sha": "8b54888bf04b1baf4decb0124aedfaf77af73273fc595f4d81a304887c8dbefa"
   },
   "cite": "what-we-measured-on-two-cards-in-one-box",
   "span": [
    18870,
    18904
   ],
   "quote": "against 5.13 for the dense Mistral",
   "identity_basis": "section",
   "identity_token": "on a 3080 split with system memory",
   "identity_cite": "what-we-measured-on-two-cards-in-one-box",
   "identity_span": [
    18800,
    18834
   ],
   "cell": null,
   "conditions_cites": [
    "what-does-it-do-on-our-own-machines"
   ],
   "same_reading_as": {
    "slug": "a-short-history-of-mistral",
    "cite": "what-does-it-do-on-our-own-machines"
   },
   "later_reading_of": null,
   "note": "the same reading this page prints in its own table next door, where it carries the date and the bench record it came from",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-520-8",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "520.8",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 2,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @600`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-428-6",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "428.6",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @450"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @450`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sm-clock-2737",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "SM clock",
   "value": "2737",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 4,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "SM MHz @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `SM MHz @600`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sm-clock-2409",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "SM clock",
   "value": "2409",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 5,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "SM MHz @450"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `SM MHz @450`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-clock-12-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the clock",
   "value": "**−12.0%**",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 6,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Clock"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta between its two capped rungs, this row at 450 W against the same row's 600 W reading — the `SM MHz @600` and `SM MHz @450` cells of this table",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-throughput-1-4",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the throughput",
   "value": "−1.4%",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    7938,
    8015
   ],
   "quote": "| Dense mistral (Q4) | 1 | 520.8 | 428.6 | 2737 | 2409 | **−12.0%** | −1.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 7,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Throughput"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta in tokens a second between its two capped rungs, this row at 450 W against the same row at 600 W. This table prints no tok/s: the concurrency-16 readings behind these deltas are in §The answer, in one paragraph, and the concurrency-1 pair is not printed on this page at all",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-586-5",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "586.5",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 2,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @600`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-448-4",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "448.4",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @450"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @450`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sm-clock-2764",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "SM clock",
   "value": "2764",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 4,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "SM MHz @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `SM MHz @600`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sm-clock-2246",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "SM clock",
   "value": "2246",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 5,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "SM MHz @450"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `SM MHz @450`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-clock-18-7",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the clock",
   "value": "**−18.7%**",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 6,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Clock"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta between its two capped rungs, this row at 450 W against the same row's 600 W reading — the `SM MHz @600` and `SM MHz @450` cells of this table",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-throughput-2-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the throughput",
   "value": "**−2.0%**",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8016,
    8098
   ],
   "quote": "| Dense mistral (Q4) | 16 | 586.5 | 448.4 | 2764 | 2246 | **−18.7%** | **−2.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 7,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Throughput"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta in tokens a second between its two capped rungs, this row at 450 W against the same row at 600 W. This table prints no tok/s: the concurrency-16 readings behind these deltas are in §The answer, in one paragraph, and the concurrency-1 pair is not printed on this page at all",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-412-8",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "412.8",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8099,
    8182
   ],
   "quote": "| MoE gemma (FP8), sustained * | 16 | 412.8 | 419.0 | 2797 | 2797 | 0.0% | −0.09% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 2,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @600`; the vLLM row averages the two interleaved repeats at each cap",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-419-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "419.0",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8099,
    8182
   ],
   "quote": "| MoE gemma (FP8), sustained * | 16 | 412.8 | 419.0 | 2797 | 2797 | 0.0% | −0.09% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Busy W @450"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `busy W @450`; the page prints beside it that the two are inside the arm's own ~10 W spread — the cap did nothing there, and the number says so",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sm-clock-2797",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "SM clock",
   "value": "2797",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8099,
    8182
   ],
   "quote": "| MoE gemma (FP8), sustained * | 16 | 412.8 | 419.0 | 2797 | 2797 | 0.0% | −0.09% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 4,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "SM MHz @600"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `SM MHz @600`",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-clock-0-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the clock",
   "value": "0.0%",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8099,
    8182
   ],
   "quote": "| MoE gemma (FP8), sustained * | 16 | 412.8 | 419.0 | 2797 | 2797 | 0.0% | −0.09% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 6,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Clock"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta between its two capped rungs, this row at 450 W against the same row's 600 W reading — the `SM MHz @600` and `SM MHz @450` cells of this table",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-the-throughput-0-09",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs the throughput",
   "value": "−0.09%",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-mechanism-checked-rather-than-inferred",
    "md_sha": "f8ef6fce4026aeaac0f803741eb8e6acb059aa7b27d15cc64d1b7796fd9c140e"
   },
   "cite": "the-mechanism-checked-rather-than-inferred",
   "span": [
    8099,
    8182
   ],
   "quote": "| MoE gemma (FP8), sustained * | 16 | 412.8 | 419.0 | 2797 | 2797 | 0.0% | −0.09% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 7,
    "header_span": [
     7811,
     7903
    ],
    "header": "| Arm | Conc. | Busy W @600 | Busy W @450 | SM MHz @600 | SM MHz @450 | Clock | Throughput |",
    "head_cell": "Throughput"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own delta in tokens a second between its two capped rungs, this row at 450 W against the same row at 600 W. This table prints no tok/s: the concurrency-16 readings behind these deltas are in §The answer, in one paragraph, and the concurrency-1 pair is not printed on this page at all",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-draw-p50-599-3-w-pegged-with-no-cap-to-blame",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy draw, p50",
   "value": "**599.3 W — pegged with no cap to blame**",
   "non_figure": null,
   "conditions": {
    "model": "flux1-dev (the heavy painter)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15509,
    15612
   ],
   "quote": "| flux1-dev (the heavy painter) | **599.3 W — pegged with no cap to blame** | **+10.2%** | **+19.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "Busy draw @600 (p50)"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-500-w-cap-costs-10-2",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 500 W cap costs",
   "value": "**+10.2%**",
   "non_figure": null,
   "conditions": {
    "model": "flux1-dev (the heavy painter)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "500 W, against the same-day 600 W rung",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15509,
    15612
   ],
   "quote": "| flux1-dev (the heavy painter) | **599.3 W — pegged with no cap to blame** | **+10.2%** | **+19.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "500 W costs"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-19-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs",
   "value": "**+19.0%**",
   "non_figure": null,
   "conditions": {
    "model": "flux1-dev (the heavy painter)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "450 W, against the same-day 600 W rung",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15509,
    15612
   ],
   "quote": "| flux1-dev (the heavy painter) | **599.3 W — pegged with no cap to blame** | **+10.2%** | **+19.0%** |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "450 W costs"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-draw-p50-449-6-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy draw, p50",
   "value": "449.6 W",
   "non_figure": null,
   "conditions": {
    "model": "flux1-schnell",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15613,
    15657
   ],
   "quote": "| flux1-schnell | 449.6 W | +5.9% | +12.1% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "Busy draw @600 (p50)"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-12-1",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs",
   "value": "+12.1%",
   "non_figure": null,
   "conditions": {
    "model": "flux1-schnell",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "450 W, against the same-day 600 W rung",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15613,
    15657
   ],
   "quote": "| flux1-schnell | 449.6 W | +5.9% | +12.1% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "450 W costs"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-draw-p50-401-7-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy draw, p50",
   "value": "401.7 W",
   "non_figure": null,
   "conditions": {
    "model": "flux-2-klein-4b (the limner)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15658,
    15716
   ],
   "quote": "| flux-2-klein-4b (the limner) | 401.7 W | +3.7% | +8.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "Busy draw @600 (p50)"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-8-4",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "what a 450 W cap costs",
   "value": "+8.4%",
   "non_figure": null,
   "conditions": {
    "model": "flux-2-klein-4b (the limner)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "450 W, against the same-day 600 W rung",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15658,
    15716
   ],
   "quote": "| flux-2-klein-4b (the limner) | 401.7 W | +3.7% | +8.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "450 W costs"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-draw-p50-291-4-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy draw, p50",
   "value": "291.4 W",
   "non_figure": null,
   "conditions": {
    "model": "sdxl (albedobase)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15717,
    15766
   ],
   "quote": "| sdxl (albedobase) | 291.4 W | −1.5% * | +0.4% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "Busy draw @600 (p50)"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-draw-p50-236-8-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "busy draw, p50",
   "value": "236.8 W",
   "non_figure": null,
   "conditions": {
    "model": "sdxl (base)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 W",
    "context": "not stated",
    "batch_or_steps": "139 sealed cells, n per row 24–31",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    15767,
    15810
   ],
   "quote": "| sdxl (base) | 236.8 W | −1.1% * | −0.2% |",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15287,
     15490
    ],
    "header": "| Image model (each cost = the median of that lane's paired per-cell ratios — same cell, same seed — against the same-day 600 W rung; n per row 24–31) | Busy draw @600 (p50) | 500 W costs | 450 W costs |",
    "head_cell": "Busy draw @600 (p50)"
   },
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-the-cap-costs-and-buys-the-600-450-w-cap-costs-2-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what the cap costs and buys",
   "value": "the 600 → 450 W cap costs 2.0%",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "256 tokens out",
    "concurrency": "16",
    "window": "a median of 48 requests at each cap, 2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6334,
    6435
   ],
   "quote": "the 600 → 450 W cap costs 2.0%\nof throughput and buys back 138 W, 11 °C, and a 28% jump in efficiency",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/decode-per-stream-94-6-92-7-tok-s-per-stream",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per stream",
   "value": "94.6 → 92.7 tok/s per stream",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "256 tokens out",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6458,
    6486
   ],
   "quote": "94.6 → 92.7 tok/s per stream",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/busy-window-draw-586-5-448-4-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "busy-window draw",
   "value": "586.5 → 448.4 W",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6488,
    6520
   ],
   "quote": "busy-window draw 586.5 → 448.4 W",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/peak-temperature-83-72-c",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "peak temperature",
   "value": "83 → 72 °C",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6522,
    6549
   ],
   "quote": "peak temperature 83 → 72 °C",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/efficiency-0-161-0-206-tok-s-per-watt",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "efficiency",
   "value": "0.161 → 0.206 tok/s per watt",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6551,
    6590
   ],
   "quote": "efficiency 0.161 → 0.206 tok/s per watt",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/time-to-first-token-p50-and-p95-p50-20-6-22-2-ms-and-p95-20-8-23-8-ms",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "time to first token, p50 and p95",
   "value": "p50 20.6 → 22.2 ms and p95 20.8 → 23.8 ms",
   "non_figure": null,
   "conditions": {
    "model": "dense mistral (Q4)",
    "quant": "Q4",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 → 450 W",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "three requests, 2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    6930,
    6971
   ],
   "quote": "p50 20.6 → 22.2 ms and p95 20.8 → 23.8 ms",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/sustained-decode-170-24-170-09-tok-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "sustained decode",
   "value": "170.24 → 170.09 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "MoE gemma (FP8), sustained",
    "quant": "FP8",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "600 → 450 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "averaged over both interleaved repeats, 2026-08-25"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "the-answer-in-one-paragraph",
    "md_sha": "dc42e5b59903e9955018d6a631ece05c5e99085304ff0c857c94806027d909b2"
   },
   "cite": "the-answer-in-one-paragraph",
   "span": [
    7361,
    7399
   ],
   "quote": "sustained decode\n170.24 → 170.09 tok/s",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/what-a-450-w-cap-costs-by-resolution-12-0-at-512-px-19-2-at-768-20-4-at-1024",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what a 450 W cap costs, by resolution",
   "value": "+12.0% at 512 px, +19.2% at 768, +20.4% at 1024",
   "non_figure": null,
   "conditions": {
    "model": "flux1-dev (the heavy painter)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "450 W, against the same-day 600 W rung",
    "context": "not stated",
    "batch_or_steps": "512 px, 768, 1024",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    16030,
    16077
   ],
   "quote": "+12.0% at 512 px, +19.2% at 768, +20.4% at 1024",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the table row above pools them — the page's own words",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/the-clock-under-the-ladder-dragged-from-2392-to-1927-mhz-and-at-450-to-1796-mhz",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the clock under the ladder",
   "value": "dragged from 2392 to 1927 MHz, and at 450 to 1796 MHz",
   "non_figure": null,
   "conditions": {
    "model": "flux1-dev (the heavy painter)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "600 → 500 → 450 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-27, 16:19–18:14Z (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    16341,
    16404
   ],
   "quote": "its clock dragged from 2392 to 1927 MHz, and at 450\nto 1796 MHz",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/decode-the-serving-lane-gemma-read-207-3-207-1-207-2-tok-s-at-600-500-450-w",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, the serving lane",
   "value": "gemma read 207.3 / 207.1 / 207.2 tok/s at 600 / 500 / 450 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 / 500 / 450 W",
    "context": "32768",
    "batch_or_steps": "128 tokens out",
    "concurrency": "1",
    "window": "the median of 6–7 per rung, 2026-08-27"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    17130,
    17189
   ],
   "quote": "gemma read 207.3 /\n207.1 / 207.2 tok/s at 600 / 500 / 450 W",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/what-150-watts-buys/decode-the-serving-lane-mistral-95-0-94-6-93-5-at-the-same-three",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, the serving lane",
   "value": "mistral 95.0 / 94.6 / 93.5 at the same three",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "not stated",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "600 / 500 / 450 W",
    "context": "32768",
    "batch_or_steps": "128 tokens out",
    "concurrency": "1",
    "window": "the median of 6–7 per rung, 2026-08-27"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "what-150-watts-buys",
    "section": "same-dial-other-lane-measured-on-the-production-card-the",
    "md_sha": "494349609a3a9574ae554725b1f88d0731345064a416027ea7a4b41334a84fe3"
   },
   "cite": "same-dial-other-lane-measured-on-the-production-card-the",
   "span": [
    17227,
    17271
   ],
   "quote": "mistral 95.0 /\n94.6 / 93.5 at the same three",
   "identity_basis": "page",
   "identity_token": "96 GB workstation card",
   "identity_cite": "the-moment-this-page-is-about",
   "identity_span": [
    2845,
    2889
   ],
   "cell": null,
   "conditions_cites": [
    "the-moment-this-page-is-about",
    "what-we-ran-and-the-rule-behind-every-number",
    "the-mechanism-checked-rather-than-inferred"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-0-48-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "0.48 s",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B *(our sketch artist)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8699,
    8781
   ],
   "quote": "| FLUX.2 klein-4B *(our sketch artist)* | 0.48 s | 1.65 s | 1.39 s | 3.4× | 2.9× |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-1-65-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "1.65 s",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B *(our sketch artist)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8699,
    8781
   ],
   "quote": "| FLUX.2 klein-4B *(our sketch artist)* | 0.48 s | 1.65 s | 1.39 s | 3.4× | 2.9× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-1-39-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "1.39 s",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B *(our sketch artist)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8699,
    8781
   ],
   "quote": "| FLUX.2 klein-4B *(our sketch artist)* | 0.48 s | 1.65 s | 1.39 s | 3.4× | 2.9× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-3-4",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "3.4×",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B *(our sketch artist)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8699,
    8781
   ],
   "quote": "| FLUX.2 klein-4B *(our sketch artist)* | 0.48 s | 1.65 s | 1.39 s | 3.4× | 2.9× |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-2-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "2.9×",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B *(our sketch artist)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8699,
    8781
   ],
   "quote": "| FLUX.2 klein-4B *(our sketch artist)* | 0.48 s | 1.65 s | 1.39 s | 3.4× | 2.9× |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-1-13-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "1.13 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8782,
    8840
   ],
   "quote": "| z-image-turbo | 1.13 s | 4.44 s | 3.80 s | 3.9× | 3.4× |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-4-44-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "4.44 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8782,
    8840
   ],
   "quote": "| z-image-turbo | 1.13 s | 4.44 s | 3.80 s | 3.9× | 3.4× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-3-80-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "3.80 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8782,
    8840
   ],
   "quote": "| z-image-turbo | 1.13 s | 4.44 s | 3.80 s | 3.9× | 3.4× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-3-9",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "3.9×",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8782,
    8840
   ],
   "quote": "| z-image-turbo | 1.13 s | 4.44 s | 3.80 s | 3.9× | 3.4× |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-3-4",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "3.4×",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8782,
    8840
   ],
   "quote": "| z-image-turbo | 1.13 s | 4.44 s | 3.80 s | 3.9× | 3.4× |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-1-97-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "1.97 s",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8841,
    8897
   ],
   "quote": "| krea2-turbo | 1.97 s | 9.62 s | 5.86 s | 4.9× | 3.0× |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-9-62-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "9.62 s",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8841,
    8897
   ],
   "quote": "| krea2-turbo | 1.97 s | 9.62 s | 5.86 s | 4.9× | 3.0× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-5-86-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "5.86 s",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8841,
    8897
   ],
   "quote": "| krea2-turbo | 1.97 s | 9.62 s | 5.86 s | 4.9× | 3.0× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-4-9",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "4.9×",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8841,
    8897
   ],
   "quote": "| krea2-turbo | 1.97 s | 9.62 s | 5.86 s | 4.9× | 3.0× |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-3-0",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "3.0×",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8841,
    8897
   ],
   "quote": "| krea2-turbo | 1.97 s | 9.62 s | 5.86 s | 4.9× | 3.0× |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-2-53-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "2.53 s",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8898,
    8951
   ],
   "quote": "| hidream-o1-dev | 2.53 s | 7.34 s | 3.45 s | † | † |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-7-34-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "7.34 s",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8898,
    8951
   ],
   "quote": "| hidream-o1-dev | 2.53 s | 7.34 s | 3.45 s | † | † |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-3-45-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "3.45 s",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8898,
    8951
   ],
   "quote": "| hidream-o1-dev | 2.53 s | 7.34 s | 3.45 s | † | † |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "†",
   "non_figure": "†",
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8898,
    8951
   ],
   "quote": "| hidream-o1-dev | 2.53 s | 7.34 s | 3.45 s | † | † |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "†",
   "non_figure": "†",
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8898,
    8951
   ],
   "quote": "| hidream-o1-dev | 2.53 s | 7.34 s | 3.45 s | † | † |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-6-17-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "6.17 s",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8952,
    9014
   ],
   "quote": "| kandinsky5-lite | 6.17 s | 28.95 s | 24.46 s | 4.7× | 4.0× |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-28-95-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "28.95 s",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8952,
    9014
   ],
   "quote": "| kandinsky5-lite | 6.17 s | 28.95 s | 24.46 s | 4.7× | 4.0× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-24-46-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "24.46 s",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8952,
    9014
   ],
   "quote": "| kandinsky5-lite | 6.17 s | 28.95 s | 24.46 s | 4.7× | 4.0× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-4-7",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "4.7×",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8952,
    9014
   ],
   "quote": "| kandinsky5-lite | 6.17 s | 28.95 s | 24.46 s | 4.7× | 4.0× |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-4-0",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "4.0×",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    8952,
    9014
   ],
   "quote": "| kandinsky5-lite | 6.17 s | 28.95 s | 24.46 s | 4.7× | 4.0× |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-7-84-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "7.84 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-base *(undistilled, 768 only)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    9015,
    9100
   ],
   "quote": "| z-image-base *(undistilled, 768 only)* | 7.84 s | 33.01 s | 28.88 s | 4.2× | 3.7× |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 1,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-33-01-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "33.01 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-base *(undistilled, 768 only)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    9015,
    9100
   ],
   "quote": "| z-image-base *(undistilled, 768 only)* | 7.84 s | 33.01 s | 28.88 s | 4.2× | 3.7× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 2,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 3090 24G (same box)"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/median-seconds-per-render-mixed-resolutions-28-88-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "median seconds per render, mixed resolutions",
   "value": "28.88 s",
   "non_figure": null,
   "conditions": {
    "model": "z-image-base *(undistilled, 768 only)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    9015,
    9100
   ],
   "quote": "| z-image-base *(undistilled, 768 only)* | 7.84 s | 33.01 s | 28.88 s | 4.2× | 3.7× |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8570,
    8672
   ],
   "cell": {
    "column": 3,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "RTX 5090 Laptop 24G"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-4-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "4.2×",
   "non_figure": null,
   "conditions": {
    "model": "z-image-base *(undistilled, 768 only)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    9015,
    9100
   ],
   "quote": "| z-image-base *(undistilled, 768 only)* | 7.84 s | 33.01 s | 28.88 s | 4.2× | 3.7× |",
   "identity_basis": "section",
   "identity_token": "RTX 3090 24G (same box)",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8591,
    8614
   ],
   "cell": {
    "column": 4,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "3090 × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/the-exhibit-s-own-multiplier-against-the-96-gb-card-3-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own multiplier against the 96 GB card",
   "value": "3.7×",
   "non_figure": null,
   "conditions": {
    "model": "z-image-base *(undistilled, 768 only)*",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "identical mixed resolution-and-step sets on every rig",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-table-that-answers-the-question",
    "md_sha": "361504aefa815d23bee50a606b6ad7b7b8e2e2515831e074b14381f0f7b484c2"
   },
   "cite": "the-table-that-answers-the-question",
   "span": [
    9015,
    9100
   ],
   "quote": "| z-image-base *(undistilled, 768 only)* | 7.84 s | 33.01 s | 28.88 s | 4.2× | 3.7× |",
   "identity_basis": "section",
   "identity_token": "RTX 5090 Laptop 24G",
   "identity_cite": "the-table-that-answers-the-question",
   "identity_span": [
    8617,
    8636
   ],
   "cell": {
    "column": 5,
    "header_span": [
     8570,
     8672
    ],
    "header": "| Model | 96G card | RTX 3090 24G (same box) | RTX 5090 Laptop 24G | 3090 × slower | Laptop × slower |",
    "head_cell": "Laptop × slower"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-0-26-s-0-26-0-40",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "0.26 s (0.26–0.40)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512×512, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9892,
    9987
   ],
   "quote": "| 512×512 *(the live default)* | 0.26 s (0.26–0.40) | 0.88 s (0.88–1.19) | 0.71 s (0.69–1.27) |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-0-88-s-0-88-1-19",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "0.88 s (0.88–1.19)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512×512, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9892,
    9987
   ],
   "quote": "| 512×512 *(the live default)* | 0.26 s (0.26–0.40) | 0.88 s (0.88–1.19) | 0.71 s (0.69–1.27) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 3090"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-0-71-s-0-69-1-27",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "0.71 s (0.69–1.27)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512×512, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9892,
    9987
   ],
   "quote": "| 512×512 *(the live default)* | 0.26 s (0.26–0.40) | 0.88 s (0.88–1.19) | 0.71 s (0.69–1.27) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 3,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 5090 Laptop"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-0-48-s-0-47-0-61",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "0.48 s (0.47–0.61)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768×768, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9988,
    10062
   ],
   "quote": "| 768×768 | 0.48 s (0.47–0.61) | 1.65 s (1.64–1.96) | 1.37 s (1.36–2.02) |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-1-65-s-1-64-1-96",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "1.65 s (1.64–1.96)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768×768, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9988,
    10062
   ],
   "quote": "| 768×768 | 0.48 s (0.47–0.61) | 1.65 s (1.64–1.96) | 1.37 s (1.36–2.02) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 3090"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-1-37-s-1-36-2-02",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "1.37 s (1.36–2.02)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768×768, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    9988,
    10062
   ],
   "quote": "| 768×768 | 0.48 s (0.47–0.61) | 1.65 s (1.64–1.96) | 1.37 s (1.36–2.02) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 3,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 5090 Laptop"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-0-82-s-0-80-0-96",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "0.82 s (0.80–0.96)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1024×1024, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "August 23rd"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    10063,
    10139
   ],
   "quote": "| 1024×1024 | 0.82 s (0.80–0.96) | 2.86 s (2.84–3.17) | 2.40 s (2.35–2.98) |",
   "identity_basis": "column-head",
   "identity_token": "96G card",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "96G card"
   },
   "conditions_cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-2-86-s-2-84-3-17",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "2.86 s (2.84–3.17)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1024×1024, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    10063,
    10139
   ],
   "quote": "| 1024×1024 | 0.82 s (0.80–0.96) | 2.86 s (2.84–3.17) | 2.40 s (2.35–2.98) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 3090",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 3090"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/seconds-per-render-median-min-max-2-40-s-2-35-2-98",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median (min–max)",
   "value": "2.40 s (2.35–2.98)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1024×1024, steps held at 4, six renders per cell",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "the-sketch-ladder-the-number-the-game-lives-on",
    "md_sha": "40103c158d96a0289b4125f652fca78f13298396984a1190a9405ea172b0125d"
   },
   "cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "span": [
    10063,
    10139
   ],
   "quote": "| 1024×1024 | 0.82 s (0.80–0.96) | 2.86 s (2.84–3.17) | 2.40 s (2.35–2.98) |",
   "identity_basis": "column-head",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-sketch-ladder-the-number-the-game-lives-on",
   "identity_span": [
    9818,
    9873
   ],
   "cell": {
    "column": 3,
    "header_span": [
     9818,
     9873
    ],
    "header": "| Sketch size | 96G card | RTX 3090 | RTX 5090 Laptop |",
    "head_cell": "RTX 5090 Laptop"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-rig-your-friend-already-owns/seconds-per-render-against-the-96-gb-card-at-512-px-the-laptop-beats-the-96-gb-c",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "seconds per render against the 96 GB card, at 512 px",
   "value": "the *laptop beats the 96 GB card outright* (1.90 s vs 2.49 s)",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "at 512",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-rig-your-friend-already-owns",
    "section": "three-findings-we-didnt-order",
    "md_sha": "b1d2eac319529b8a9501ee117f62817f7ddb582266cf54454f2aa7ec3f0ad788"
   },
   "cite": "three-findings-we-didnt-order",
   "span": [
    11217,
    11278
   ],
   "quote": "the *laptop beats the 96 GB card outright* (1.90 s vs 2.49 s)",
   "identity_basis": "page",
   "identity_token": "an **RTX 5090 Laptop GPU** (24 GB — the laptop part, not the desktop card) in a gaming laptop",
   "identity_cite": "what-we-ran-and-the-rule-behind-every-number",
   "identity_span": [
    3909,
    4002
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-13-6-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "13.6 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11934,
    11996
   ],
   "quote": "| qwen3.5:9b | 9B dense | 13.6 tok/s | 7.5 tok/s | 9.1 tok/s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-7-5-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "7.5 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11934,
    11996
   ],
   "quote": "| qwen3.5:9b | 9B dense | 13.6 tok/s | 7.5 tok/s | 9.1 tok/s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-9-1-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "9.1 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11934,
    11996
   ],
   "quote": "| qwen3.5:9b | 9B dense | 13.6 tok/s | 7.5 tok/s | 9.1 tok/s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-11-0-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "11.0 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11997,
    12060
   ],
   "quote": "| gemma4:12b | 12B dense | 11.0 tok/s | 5.0 tok/s | 6.0 tok/s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-5-0-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "5.0 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11997,
    12060
   ],
   "quote": "| gemma4:12b | 12B dense | 11.0 tok/s | 5.0 tok/s | 6.0 tok/s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-6-0-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "6.0 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    11997,
    12060
   ],
   "quote": "| gemma4:12b | 12B dense | 11.0 tok/s | 5.0 tok/s | 6.0 tok/s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-11-1-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "11.1 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12061,
    12122
   ],
   "quote": "| phi4:14b | 14B dense | 11.1 tok/s | 4.4 tok/s | 5.2 tok/s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-4-4-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "4.4 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12061,
    12122
   ],
   "quote": "| phi4:14b | 14B dense | 11.1 tok/s | 4.4 tok/s | 5.2 tok/s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-5-2-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "5.2 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12061,
    12122
   ],
   "quote": "| phi4:14b | 14B dense | 11.1 tok/s | 4.4 tok/s | 5.2 tok/s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-7-1-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "7.1 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12123,
    12195
   ],
   "quote": "| mistral-small3.2:24b | 24B dense | 7.1 tok/s | 2.7 tok/s | 3.3 tok/s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-2-7-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "2.7 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12123,
    12195
   ],
   "quote": "| mistral-small3.2:24b | 24B dense | 7.1 tok/s | 2.7 tok/s | 3.3 tok/s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-3-3-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "3.3 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12123,
    12195
   ],
   "quote": "| mistral-small3.2:24b | 24B dense | 7.1 tok/s | 2.7 tok/s | 3.3 tok/s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-24-7-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**24.7 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12196,
    12292
   ],
   "quote": "| **gemma4:26b** | **26B MoE (~4B active)** | **24.7 tok/s** | **13.3 tok/s** | **12.7 tok/s** |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-13-3-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**13.3 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12196,
    12292
   ],
   "quote": "| **gemma4:26b** | **26B MoE (~4B active)** | **24.7 tok/s** | **13.3 tok/s** | **12.7 tok/s** |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-12-7-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**12.7 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12196,
    12292
   ],
   "quote": "| **gemma4:26b** | **26B MoE (~4B active)** | **24.7 tok/s** | **13.3 tok/s** | **12.7 tok/s** |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/decode-median-of-five-scored-runs-31-5-tok-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**31.5 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12293,
    12409
   ],
   "quote": "| **nemotron-3.5-lightning:30b-a3b** | **30B MoE (~3B active)** | **31.5 tok/s** | **15.4 tok/s** | **14.6 tok/s** |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/decode-median-of-five-scored-runs-15-4-tok-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**15.4 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12293,
    12409
   ],
   "quote": "| **nemotron-3.5-lightning:30b-a3b** | **30B MoE (~3B active)** | **31.5 tok/s** | **15.4 tok/s** | **14.6 tok/s** |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-one-table-to-remember",
   "identity_span": [
    11853,
    11911
   ],
   "cell": {
    "column": 3,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-median-of-five-scored-runs-14-6-tok-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, median of five scored runs",
   "value": "**14.6 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-one-table-to-remember",
    "md_sha": "c37ccf0f18652cb8214e05d7f204aa1be8b516feac662dffdf1b91e59403e169"
   },
   "cite": "the-one-table-to-remember",
   "span": [
    12293,
    12409
   ],
   "quote": "| **nemotron-3.5-lightning:30b-a3b** | **30B MoE (~3B active)** | **31.5 tok/s** | **15.4 tok/s** | **14.6 tok/s** |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     11853,
     11911
    ],
    "header": "| Model (all Q4_K_M) | Class | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-4-9-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "4.9 s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "537 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15661,
    15707
   ],
   "quote": "| qwen3.5:9b | 537 | 4.9 s | 14.1 s | 16.0 s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-14-1-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "14.1 s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "537 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15661,
    15707
   ],
   "quote": "| qwen3.5:9b | 537 | 4.9 s | 14.1 s | 16.0 s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-16-0-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "16.0 s",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.5:9b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "537 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15661,
    15707
   ],
   "quote": "| qwen3.5:9b | 537 | 4.9 s | 14.1 s | 16.0 s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-6-9-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "6.9 s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15708,
    15754
   ],
   "quote": "| gemma4:12b | 536 | 6.9 s | 22.2 s | 25.6 s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-22-2-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "22.2 s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15708,
    15754
   ],
   "quote": "| gemma4:12b | 536 | 6.9 s | 22.2 s | 25.6 s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-25-6-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "25.6 s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15708,
    15754
   ],
   "quote": "| gemma4:12b | 536 | 6.9 s | 22.2 s | 25.6 s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-8-1-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "8.1 s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "505 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15755,
    15799
   ],
   "quote": "| phi4:14b | 505 | 8.1 s | 26.4 s | 35.4 s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-26-4-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "26.4 s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "505 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15755,
    15799
   ],
   "quote": "| phi4:14b | 505 | 8.1 s | 26.4 s | 35.4 s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-35-4-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "35.4 s",
   "non_figure": null,
   "conditions": {
    "model": "phi4:14b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "505 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15755,
    15799
   ],
   "quote": "| phi4:14b | 505 | 8.1 s | 26.4 s | 35.4 s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-3-7-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "**3.7 s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15800,
    15863
   ],
   "quote": "| gemma4:26b (MoE) | 536 | **3.7 s** | **8.5 s** | **11.4 s** |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-8-5-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "**8.5 s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15800,
    15863
   ],
   "quote": "| gemma4:26b (MoE) | 536 | **3.7 s** | **8.5 s** | **11.4 s** |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-11-4-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "**11.4 s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "536 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15800,
    15863
   ],
   "quote": "| gemma4:26b (MoE) | 536 | **3.7 s** | **8.5 s** | **11.4 s** |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-4-0-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "4.0 s",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "532 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15864,
    15935
   ],
   "quote": "| nemotron-3.5-lightning:30b-a3b (MoE) | 532 | 4.0 s | 9.9 s | 14.1 s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-9-9-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "9.9 s",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "532 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15864,
    15935
   ],
   "quote": "| nemotron-3.5-lightning:30b-a3b (MoE) | 532 | 4.0 s | 9.9 s | 14.1 s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-14-1-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "14.1 s",
   "non_figure": null,
   "conditions": {
    "model": "nemotron-3.5-lightning:30b-a3b (MoE)",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "532 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15864,
    15935
   ],
   "quote": "| nemotron-3.5-lightning:30b-a3b (MoE) | 532 | 4.0 s | 9.9 s | 14.1 s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-24-6-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "24.6 s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "1,023 prompt tokens, its 32 physical cores",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15936,
    15999
   ],
   "quote": "| mistral-small3.2:24b | **1,023** | 24.6 s | 82.4 s | 97.7 s |",
   "identity_basis": "page",
   "identity_token": "**the server** — 128G of ECC RAM, a server-class processor.",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9211,
    9270
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Server"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-82-4-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "82.4 s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "1,023 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15936,
    15999
   ],
   "quote": "| mistral-small3.2:24b | **1,023** | 24.6 s | 82.4 s | 97.7 s |",
   "identity_basis": "column-head",
   "identity_token": "Rig B †",
   "identity_cite": "the-wait-prefill-is-the-real-cpu-tax",
   "identity_span": [
    15572,
    15638
   ],
   "cell": {
    "column": 3,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Rig B †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/time-to-first-token-median-of-five-cold-prefix-runs-97-7-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to first token, median of five cold-prefix runs",
   "value": "97.7 s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "Ollama",
    "runtime_version": "0.32.15",
    "cap": "not stated",
    "context": "a 4,096-token context",
    "batch_or_steps": "1,023 prompt tokens, Ollama's own default of 4 threads",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "the-wait-prefill-is-the-real-cpu-tax",
    "md_sha": "516f426865f5d9e3e494d9abd8fb7b3de3a3c98b1ca9785d4b6d7d5f9760a1d5"
   },
   "cite": "the-wait-prefill-is-the-real-cpu-tax",
   "span": [
    15936,
    15999
   ],
   "quote": "| mistral-small3.2:24b | **1,023** | 24.6 s | 82.4 s | 97.7 s |",
   "identity_basis": "page",
   "identity_token": "a mini PC with 32G RAM and a 12-core Intel processor",
   "identity_cite": "the-question-our-gpu-benches-never-ask",
   "identity_span": [
    9345,
    9413
   ],
   "cell": {
    "column": 4,
    "header_span": [
     15572,
     15638
    ],
    "header": "| Model (all Q4_K_M) | Prompt tokens | Server | Rig B † | Mini † |",
    "head_cell": "Mini †"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-server/two-hours-on-battery/memory-bandwidth-best-triad-100-7-gb-s",
   "identity": "the-server",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory bandwidth, best triad",
   "value": "**100.7 GB/s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "a STREAM triad, gcc -O3 -march=native -fopenmp",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "a 2,289 MiB working set, best of 9 timed reps",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC), 17:57–19:46Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
    "md_sha": "a3e9487d60ca8797eb9440aba567064b429ac679edbae27673917c6eb9fdd7b9"
   },
   "cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "span": [
    10490,
    10552
   ],
   "quote": "| The server (128G ECC) | **100.7 GB/s** | 100.4 GB/s | 0.3% |",
   "identity_basis": "row",
   "identity_token": "the server (128G ECC)",
   "identity_cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "identity_span": [
    10490,
    10552
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10434,
     10471
    ],
    "header": "| Box | Best triad | Worst | Spread |",
    "head_cell": "Best triad"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/two-hours-on-battery/memory-bandwidth-best-triad-59-3-gb-s",
   "identity": "rig-b-the-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory bandwidth, best triad",
   "value": "**59.3 GB/s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "a STREAM triad, gcc -O3 -march=native -fopenmp",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "a 2,289 MiB working set, best of 9 timed reps",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC), 17:57–19:46Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
    "md_sha": "a3e9487d60ca8797eb9440aba567064b429ac679edbae27673917c6eb9fdd7b9"
   },
   "cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "span": [
    10553,
    10605
   ],
   "quote": "| Rig B laptop | **59.3 GB/s** | 49.2 GB/s | 20.5% |",
   "identity_basis": "row",
   "identity_token": "Rig B laptop",
   "identity_cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "identity_span": [
    10553,
    10605
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10434,
     10471
    ],
    "header": "| Box | Best triad | Worst | Spread |",
    "head_cell": "Best triad"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/memory-bandwidth-best-triad-51-1-gb-s",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory bandwidth, best triad",
   "value": "**51.1 GB/s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "a STREAM triad, gcc -O3 -march=native -fopenmp",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "a 2,289 MiB working set, best of 9 timed reps",
    "concurrency": "not stated",
    "window": "2026-08-25 (UTC), 17:57–19:46Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
    "md_sha": "a3e9487d60ca8797eb9440aba567064b429ac679edbae27673917c6eb9fdd7b9"
   },
   "cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "span": [
    10606,
    10659
   ],
   "quote": "| The mini (32G) | **51.1 GB/s** | 50.9 GB/s | 0.5% |",
   "identity_basis": "row",
   "identity_token": "the mini (32G)",
   "identity_cite": "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth",
   "identity_span": [
    10606,
    10659
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10434,
     10471
    ],
    "header": "| Box | Best triad | Worst | Spread |",
    "head_cell": "Best triad"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/draw-at-the-wall-by-the-ups-s-own-meter-0-w-shown-the-box-s-5-w-real-draw-is-bel",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall, by the UPS's own meter",
   "value": "**0 W shown** (the box's ~5 W real draw is below the UPS meter's floor)",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:38:31–02:41"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4652,
    4803
   ],
   "quote": "| Idle, on mains | 02:38:31–02:41 | OL | 0% | **0 W shown** (the box's ~5 W real draw is below the UPS meter's floor) | 100% | **27,000 s = 450 min** |",
   "identity_basis": "section",
   "identity_token": "this bench's mini rig",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3615,
    3636
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Draw"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this column: whole-box wall draw by the UPS's own meter — the reading is the mini PC's and the instrument is the battery unit",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/battery-charge-by-the-ups-s-own-meter-100",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "battery charge, by the UPS's own meter",
   "value": "100%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:38:31–02:41"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4652,
    4803
   ],
   "quote": "| Idle, on mains | 02:38:31–02:41 | OL | 0% | **0 W shown** (the box's ~5 W real draw is below the UPS meter's floor) | 100% | **27,000 s = 450 min** |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 5,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Battery"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-ups-s-own-runtime-estimate-27-000-s-450-min",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "the UPS's own runtime estimate",
   "value": "**27,000 s = 450 min**",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:38:31–02:41"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4652,
    4803
   ],
   "quote": "| Idle, on mains | 02:38:31–02:41 | OL | 0% | **0 W shown** (the box's ~5 W real draw is below the UPS meter's floor) | 100% | **27,000 s = 450 min** |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 6,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Runtime estimate"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/load-by-the-ups-s-own-meter-8",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "load, by the UPS's own meter",
   "value": "8%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:41–02:56:56"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4804,
    4900
   ],
   "quote": "| Generating, on mains | 02:41–02:56:56 | OL | 8% | 60–82 W, mean **75.6 W** | 100% | ~7,575 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 3,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Load"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/draw-at-the-wall-by-the-ups-s-own-meter-60-82-w-mean-75-6-w",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall, by the UPS's own meter",
   "value": "60–82 W, mean **75.6 W**",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:41–02:56:56"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4804,
    4900
   ],
   "quote": "| Generating, on mains | 02:41–02:56:56 | OL | 8% | 60–82 W, mean **75.6 W** | 100% | ~7,575 s |",
   "identity_basis": "section",
   "identity_token": "this bench's mini rig",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3615,
    3636
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Draw"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this column: whole-box wall draw by the UPS's own meter — the reading is the mini PC's and the instrument is the battery unit",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/battery-charge-by-the-ups-s-own-meter-100-2",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "battery charge, by the UPS's own meter",
   "value": "100%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:41–02:56:56"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4804,
    4900
   ],
   "quote": "| Generating, on mains | 02:41–02:56:56 | OL | 8% | 60–82 W, mean **75.6 W** | 100% | ~7,575 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 5,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Battery"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-ups-s-own-runtime-estimate-7-575-s",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "the UPS's own runtime estimate",
   "value": "~7,575 s",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:41–02:56:56"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4804,
    4900
   ],
   "quote": "| Generating, on mains | 02:41–02:56:56 | OL | 8% | 60–82 W, mean **75.6 W** | 100% | ~7,575 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 6,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Runtime estimate"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/load-by-the-ups-s-own-meter-8-9",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "load, by the UPS's own meter",
   "value": "8–9%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:56:56–03:03:00"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4901,
    5070
   ],
   "quote": "| **generating, ON BATTERY** | **02:56:56–03:03:00** | OB DISCHRG | 8–9% | 74–81 W, mean **78.2 W** | 100→96% | 6,186–7,573 s (the estimator hunting between generates) |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 3,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Load"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/draw-at-the-wall-by-the-ups-s-own-meter-74-81-w-mean-78-2-w",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall, by the UPS's own meter",
   "value": "74–81 W, mean **78.2 W**",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:56:56–03:03:00"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4901,
    5070
   ],
   "quote": "| **generating, ON BATTERY** | **02:56:56–03:03:00** | OB DISCHRG | 8–9% | 74–81 W, mean **78.2 W** | 100→96% | 6,186–7,573 s (the estimator hunting between generates) |",
   "identity_basis": "section",
   "identity_token": "this bench's mini rig",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3615,
    3636
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Draw"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this column: whole-box wall draw by the UPS's own meter — the reading is the mini PC's and the instrument is the battery unit",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/battery-charge-by-the-ups-s-own-meter-100-96",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "battery charge, by the UPS's own meter",
   "value": "100→96%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:56:56–03:03:00"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4901,
    5070
   ],
   "quote": "| **generating, ON BATTERY** | **02:56:56–03:03:00** | OB DISCHRG | 8–9% | 74–81 W, mean **78.2 W** | 100→96% | 6,186–7,573 s (the estimator hunting between generates) |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 5,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Battery"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-ups-s-own-runtime-estimate-6-186-7-573-s-the-estimator-hunting-between-gener",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "the UPS's own runtime estimate",
   "value": "6,186–7,573 s (the estimator hunting between generates)",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "02:56:56–03:03:00"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    4901,
    5070
   ],
   "quote": "| **generating, ON BATTERY** | **02:56:56–03:03:00** | OB DISCHRG | 8–9% | 74–81 W, mean **78.2 W** | 100→96% | 6,186–7,573 s (the estimator hunting between generates) |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 6,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Runtime estimate"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/load-by-the-ups-s-own-meter-8-2",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "load, by the UPS's own meter",
   "value": "8%",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "03:03–03:05:02"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5071,
    5181
   ],
   "quote": "| Generating, recovery | 03:03–03:05:02 | OL CHRG | 8% | 75–76 W | 96%, charging | Rebuilding: 7,301→7,314 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 3,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Load"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/draw-at-the-wall-by-the-ups-s-own-meter-75-76-w",
   "identity": "the-mini-pc",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall, by the UPS's own meter",
   "value": "75–76 W",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "03:03–03:05:02"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5071,
    5181
   ],
   "quote": "| Generating, recovery | 03:03–03:05:02 | OL CHRG | 8% | 75–76 W | 96%, charging | Rebuilding: 7,301→7,314 s |",
   "identity_basis": "section",
   "identity_token": "this bench's mini rig",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3615,
    3636
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Draw"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this column: whole-box wall draw by the UPS's own meter — the reading is the mini PC's and the instrument is the battery unit",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/battery-charge-by-the-ups-s-own-meter-96-charging",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "battery charge, by the UPS's own meter",
   "value": "96%, charging",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "03:03–03:05:02"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5071,
    5181
   ],
   "quote": "| Generating, recovery | 03:03–03:05:02 | OL CHRG | 8% | 75–76 W | 96%, charging | Rebuilding: 7,301→7,314 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 5,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Battery"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-ups-s-own-runtime-estimate-rebuilding-7-301-7-314-s",
   "identity": "cyberpower-pr1500",
   "road": "page-table",
   "kind": "figure",
   "metric": "the UPS's own runtime estimate",
   "value": "Rebuilding: 7,301→7,314 s",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "03:03–03:05:02"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5071,
    5181
   ],
   "quote": "| Generating, recovery | 03:03–03:05:02 | OL CHRG | 8% | 75–76 W | 96%, charging | Rebuilding: 7,301→7,314 s |",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": {
    "column": 6,
    "header_span": [
     4541,
     4621
    ],
    "header": "| Phase | Window (UTC) | UPS status | Load | Draw | Battery | Runtime estimate |",
    "head_cell": "Runtime estimate"
   },
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/decode-tokens-a-second-137",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "137",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter mixture-of-experts writer",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one question at a time",
    "window": "24 August 2026, 85 receipts"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9454,
    9529
   ],
   "quote": "| A 26-billion-parameter mixture-of-experts writer, tokens a second | 137 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/decode-tokens-a-second-53",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "53",
   "non_figure": null,
   "conditions": {
    "model": "a dense 24-billion-parameter model",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one question at a time",
    "window": "24 August 2026, 85 receipts"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9530,
    9590
   ],
   "quote": "| A dense 24-billion-parameter model, tokens a second | 53 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/time-to-the-first-token-0-08-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to the first token",
   "value": "0.08 s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "a 40-token prompt",
    "concurrency": "one question at a time",
    "window": "24 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9591,
    9646
   ],
   "quote": "| Time to the first token, a 40-token prompt | 0.08 s |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/time-to-the-first-token-0-30-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to the first token",
   "value": "0.30 s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "a 400-token prompt",
    "concurrency": "one question at a time",
    "window": "24 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9647,
    9703
   ],
   "quote": "| Time to the first token, a 400-token prompt | 0.30 s |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/time-to-the-first-token-1-01-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to the first token",
   "value": "1.01 s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "a 2,000-token prompt",
    "concurrency": "one question at a time",
    "window": "24 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9704,
    9762
   ],
   "quote": "| Time to the first token, a 2,000-token prompt | 1.01 s |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/prompt-reading-tokens-a-second-about-3-600",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt reading, tokens a second",
   "value": "About 3,600",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "the model server of that day",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "2,667 tokens in 0.73 s",
    "concurrency": "not stated",
    "window": "24 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9763,
    9837
   ],
   "quote": "| Prompt reading, tokens a second (2,667 tokens in 0.73 s) | About 3,600 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/aggregate-decode-eight-questions-at-once-328",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "aggregate decode, eight questions at once",
   "value": "328",
   "non_figure": null,
   "conditions": {
    "model": "the same 26-billion writer at 4-bit",
    "quant": "4-bit",
    "runtime": "a batching runtime",
    "runtime_version": "not stated",
    "cap": "measured before this workshop capped its cards",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "eight questions at once",
    "window": "22 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    9838,
    9982
   ],
   "quote": "| The same 26-billion writer at 4-bit through a batching runtime, eight questions at once, tokens a second in aggregate (22 August 2026) | 328 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9413,
     9442
    ],
    "header": "| Words, one card | Reading |",
    "head_cell": "Reading"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-0-88",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "0.88",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B, the sketch model",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512, 4 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    10889,
    10943
   ],
   "quote": "| FLUX.2 klein 4B, the sketch model | 512 | 4 | 0.88 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-1-65",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "1.65",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768, 4 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    10944,
    10980
   ],
   "quote": "| FLUX.2 klein 4B | 768 | 4 | 1.65 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-2-86",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "2.86",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1,024, 4 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    10981,
    11019
   ],
   "quote": "| FLUX.2 klein 4B | 1,024 | 4 | 2.86 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-2-10",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "2.10",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512, 8 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11020,
    11054
   ],
   "quote": "| z-image-turbo | 512 | 8 | 2.10 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-4-44",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "4.44",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768, 8 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11055,
    11089
   ],
   "quote": "| z-image-turbo | 768 | 8 | 4.44 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-8-08",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "8.08",
   "non_figure": null,
   "conditions": {
    "model": "z-image-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1,024, 8 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11090,
    11126
   ],
   "quote": "| z-image-turbo | 1,024 | 8 | 8.08 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-5-34",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "5.34",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512, 8 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11127,
    11159
   ],
   "quote": "| krea2-turbo | 512 | 8 | 5.34 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-16-09",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "16.09",
   "non_figure": null,
   "conditions": {
    "model": "krea2-turbo",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1,024, 8 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11160,
    11195
   ],
   "quote": "| krea2-turbo | 1,024 | 8 | 16.09 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-6-19",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "6.19",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512, 32 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11196,
    11232
   ],
   "quote": "| hidream-o1-dev | 512 | 32 | 6.19 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-11-62",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "11.62",
   "non_figure": null,
   "conditions": {
    "model": "hidream-o1-dev",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1,024, 32 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11233,
    11272
   ],
   "quote": "| hidream-o1-dev | 1,024 | 32 | 11.62 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-11-43",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "11.43",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512, 32 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11273,
    11311
   ],
   "quote": "| kandinsky5-lite | 512 | 32 | 11.43 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/seconds-per-render-median-of-its-cell-49-89",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per render, median of its cell",
   "value": "49.89",
   "non_figure": null,
   "conditions": {
    "model": "kandinsky5-lite",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "1,024, 32 steps",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11312,
    11352
   ],
   "quote": "| kandinsky5-lite | 1,024 | 32 | 49.89 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10817,
     10867
    ],
    "header": "| Image model, one card | Size | Steps | Seconds |",
    "head_cell": "Seconds"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/draw-at-the-wall-91-w",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall",
   "value": "91 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "each card at 280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "measured, the day's median, 9 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    20888,
    21031
   ],
   "quote": "| Idle, the box awake and both cards doing nothing (measured, the day's median, 9 September 2026) | 91 W | 2.18 kWh · $0.40 | 66 kWh · $12.10 |",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20822,
     20866
    ],
    "header": "| Posture | At the meter | A day | A month |",
    "head_cell": "At the meter"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/a-day-and-a-month-at-18-34-kwh-2-18-kwh-0-40",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "a day, and a month, at 18.34 ¢/kWh",
   "value": "2.18 kWh · $0.40",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "each card at 280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "measured, the day's median, 9 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    20888,
    21031
   ],
   "quote": "| Idle, the box awake and both cards doing nothing (measured, the day's median, 9 September 2026) | 91 W | 2.18 kWh · $0.40 | 66 kWh · $12.10 |",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20822,
     20866
    ],
    "header": "| Posture | At the meter | A day | A month |",
    "head_cell": "A day"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "an addendum in this section re-read the same box over 2026-09-12 to 2026-09-14 and moved the idle month from $12.10 to $12.28; both readings are dated and both stand",
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/draw-at-the-wall-233-w-average",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall",
   "value": "233 W average",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W caps for six hours, idle the other eighteen",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    21032,
    21184
   ],
   "quote": "| A working day: both cards at their 280 W caps for six hours, idle the other eighteen (derived) | 233 W average | 5.60 kWh · $1.03 | 168 kWh · $30.81 |",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20822,
     20866
    ],
    "header": "| Posture | At the meter | A day | A month |",
    "head_cell": "At the meter"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page labels this row `(derived)` in the cell itself: an estimate made in good faith from the caps, a ballpark, not a reading",
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/draw-at-the-wall-660-w",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw at the wall",
   "value": "660 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "both cards at their caps all day",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    21185,
    21290
   ],
   "quote": "| Both cards at their caps all day, the ceiling (derived) | 660 W | 15.8 kWh · $2.90 | 475 kWh · $87.12 |",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20822,
     20866
    ],
    "header": "| Posture | At the meter | A day | A month |",
    "head_cell": "At the meter"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page labels this row `(derived)` in the cell itself: an estimate made in good faith from the caps, a ballpark, not a reading",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-1-449-99",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, one card",
   "value": "$1,449.99",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "paid by this workshop, May 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15686,
    15773
   ],
   "quote": "| Paid by this workshop, May 2026: an EVGA XC3 Ultra, renewed | $1,449.99 | $2,899.98 |",
   "identity_basis": "row",
   "identity_token": "an EVGA XC3 Ultra",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    15686,
    15773
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "One card"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/price-the-two-cards-2-899-98",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the two cards",
   "value": "$2,899.98",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "paid by this workshop, May 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15686,
    15773
   ],
   "quote": "| Paid by this workshop, May 2026: an EVGA XC3 Ultra, renewed | $1,449.99 | $2,899.98 |",
   "identity_basis": "page",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "The pair"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this column as the one-card reading doubled, and reads the pair against the pair in its own prose: *the cheapest new pair of any brand is 26 per cent dearer than this pair cost renewed*",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-1-849-99",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, one card",
   "value": "$1,849.99",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "13 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15774,
    15845
   ],
   "quote": "| The same renewed listing, 13 September 2026 | $1,849.99 | $3,699.98 |",
   "identity_basis": "section",
   "identity_token": "the cheapest new RTX 3090 of any brand",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    15848,
    15886
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "One card"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/price-the-two-cards-3-699-98",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the two cards",
   "value": "$3,699.98",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "13 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15774,
    15845
   ],
   "quote": "| The same renewed listing, 13 September 2026 | $1,849.99 | $3,699.98 |",
   "identity_basis": "page",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "The pair"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this column as the one-card reading doubled, and reads the pair against the pair in its own prose: *the cheapest new pair of any brand is 26 per cent dearer than this pair cost renewed*",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-1-829-99",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, one card",
   "value": "$1,829.99",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "13 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15846,
    15931
   ],
   "quote": "| The cheapest new RTX 3090 of any brand, 13 September 2026 | $1,829.99 | $3,659.98 |",
   "identity_basis": "row",
   "identity_token": "the cheapest new RTX 3090 of any brand",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    15846,
    15931
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "One card"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/price-the-two-cards-3-659-98",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the two cards",
   "value": "$3,659.98",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "13 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15846,
    15931
   ],
   "quote": "| The cheapest new RTX 3090 of any brand, 13 September 2026 | $1,829.99 | $3,659.98 |",
   "identity_basis": "page",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "The pair"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this column as the one-card reading doubled, and reads the pair against the pair in its own prose: *the cheapest new pair of any brand is 26 per cent dearer than this pair cost renewed*",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-1-094-44",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, one card",
   "value": "$1,094.44",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "August 2026's average completed sale"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15932,
    16013
   ],
   "quote": "| A used RTX 3090, August 2026's average completed sale | $1,094.44 | $2,188.88 |",
   "identity_basis": "row",
   "identity_token": "a used RTX 3090",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    15932,
    16013
   ],
   "cell": {
    "column": 1,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "One card"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/price-the-two-cards-2-188-88",
   "identity": "the-two-3090-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the two cards",
   "value": "$2,188.88",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "August 2026's average completed sale"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    15932,
    16013
   ],
   "quote": "| A used RTX 3090, August 2026's average completed sale | $1,094.44 | $2,188.88 |",
   "identity_basis": "page",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     15632,
     15669
    ],
    "header": "| The reading | One card | The pair |",
    "head_cell": "The pair"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this column as the one-card reading doubled, and reads the pair against the pair in its own prose: *the cheapest new pair of any brand is 26 per cent dearer than this pair cost renewed*",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/draw-at-the-wall-320-w",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "draw at the wall",
   "value": "320 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "the night of 26 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    10417,
    10500
   ],
   "quote": "the card reading 320 W at the wall that night, above the 280 W this page recommends",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    10148,
    10187
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this beside the render table to say the card was NOT at the 280 W cap it recommends that night, and that a capped card would draw each of those pictures a few per cent slower, unmeasured",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/what-the-house-cap-costs-3-per-cent-of-writing-speed-and-6-per-cent-of-prompt-re",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what the house cap costs",
   "value": "3 per cent of writing speed and 6 per cent of prompt reading",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W, 80 per cent of the card's 350 W rating",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "23 August 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-kind-of-performance-to-expect",
    "md_sha": "efcf797578f2f8dbeda0661d71b9cf45c5d292959f802cd896b4bae83756066f"
   },
   "cite": "what-kind-of-performance-to-expect",
   "span": [
    11919,
    12018
   ],
   "quote": "measured the cost on 23 August 2026 at 3 per cent of writing speed and 6 per cent of prompt reading",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-kind-of-performance-to-expect",
   "identity_span": [
    9102,
    9125
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/one-card-idle-by-its-own-reading-about-24-w-each-by-the-second-card-s-own-readin",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one card idle, by its own reading",
   "value": "about 24 W each by the second card's own reading over the same day",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "9 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    21590,
    21656
   ],
   "quote": "about 24 W each by the second card's own reading over the same day",
   "identity_basis": "page",
   "identity_token": "Two EVGA GeForce RTX 3090 XC3 Ultras, renewed, 24 GB of graphics memory each, 48 GB between them",
   "identity_cite": "nothing-in-this-box-is-server-grade",
   "identity_span": [
    4148,
    4244
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/draw-at-the-wall-re-read-over-three-days-93-w",
   "identity": "the-two-3090-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "draw at the wall, re-read over three days",
   "value": "**93 W**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "each card at 280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-12 00:00 to 2026-09-15 00:00 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    22800,
    22888
   ],
   "quote": "the box idled at **93 W**, two watts above the 91 W of 9 September that the table prints",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/draw-at-the-wall-91-w",
   "note": "91 W is 9 September's median; this is the median of each of the three clean days 2026-09-12 to 2026-09-14",
   "record_published": null
  },
  {
   "id": "the-two-3090-rig/home-inference-and-diffusion-rig-cost/the-idle-month-at-18-34-cents-a-kilowatt-hour-re-read-12-10-to-12-28",
   "identity": "the-two-3090-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the idle month at 18.34 cents a kilowatt-hour, re-read",
   "value": "**$12.10 to $12.28**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "each card at 280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-12 00:00 to 2026-09-15 00:00 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-rig-costs-to-run",
    "md_sha": "89f600da9a22aa666d3722ab3a15c78cc9249c228f374e32fb611ac98f2514dc"
   },
   "cite": "what-the-rig-costs-to-run",
   "span": [
    22980,
    23047
   ],
   "quote": "that moves the idle month from **$12.10 to $12.28**, eighteen cents",
   "identity_basis": "section",
   "identity_token": "the darkroom's box",
   "identity_cite": "what-the-rig-costs-to-run",
   "identity_span": [
    20217,
    20235
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the same month the table prints at $12.10, moved by the three-day re-reading above; the page prints both ends of the move and this row carries its own characters",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/wall-clock-three-composers-27-minutes-48-seconds",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "wall-clock, three composers",
   "value": "27 minutes 48 seconds",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5, LoRA adapters, three composers",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "training for three composers"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "what-the-card-is-like-for-training",
    "md_sha": "f8037841a5258bee1c798bb46479c4f9926256e503f0bd435839787e68d622f3"
   },
   "cite": "what-the-card-is-like-for-training",
   "span": [
    13475,
    13496
   ],
   "quote": "27 minutes 48 seconds",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "what-the-card-is-like-for-training",
   "identity_span": [
    12888,
    12896
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-used-908-00",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, one card, used",
   "value": "$908.00",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "the past year's average completed sale"
   },
   "figure_definition": "v3",
   "status": "superseded",
   "superseded_by": {
    "slug": "home-inference-and-diffusion-rig-cost",
    "cite": "the-cards-in-may-and-in-september",
    "span": [
     17484,
     17685
    ],
    "quote": "An operator ruled it inaccurate on 2026-09-17 — a trailing twelve-month average of a market that has since climbed is not a price a reader can pay — and it is no longer a price of record on this shelf."
   },
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    17130,
    17226
   ],
   "quote": "*a used RTX 3090, the past year's average completed sale — $908.00 one card, $1,816.00 the pair*",
   "identity_basis": "row",
   "identity_token": "a used RTX 3090",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    17130,
    17226
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "WITHDRAWN BY ITS OWN EXHIBIT on 2026-09-18: the page quotes the row it used to print and says who withdrew the figure and why. The two totals derived from it, $2,256.87 and $3,164.87, went with it",
   "record_published": null
  },
  {
   "id": "rtx-3090/home-inference-and-diffusion-rig-cost/price-one-card-renewed-1-879-99",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, one card, renewed",
   "value": "**$1,879.99**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read by hand at the store page on 17 September 2026 between 15:00 and 15:30 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "home-inference-and-diffusion-rig-cost",
    "section": "the-cards-in-may-and-in-september",
    "md_sha": "4cc63bac75df32f87e3e64376b60ae08b9db8a2af11725fc65db1a631c5d4b1b"
   },
   "cite": "the-cards-in-may-and-in-september",
   "span": [
    18516,
    18609
   ],
   "quote": "It asked **$1,879.99**, $30.00 above the $1,849.99 it read on 13 September in the table above",
   "identity_basis": "section",
   "identity_token": "EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)",
   "identity_cite": "the-cards-in-may-and-in-september",
   "identity_span": [
    18246,
    18363
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the replacement price of record, read by hand because that retailer serves this workshop's fetcher no price; exhibit fifty publishes the same reading of the same listing",
   "record_published": null
  },
  {
   "id": "printlab-retired-drawing-machine/how-the-print-lab-works/one-plate-about-1-66-s",
   "identity": "printlab-retired-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one plate",
   "value": "about 1.66 s",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512 × 512, thirty-two steps",
    "concurrency": "not stated",
    "window": "median of twelve warm renders, 2026-09-04 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    46561,
    47094
   ],
   "quote": "| One plate, 512 × 512, thirty-two steps, **on the machine retired from drawing on 2026-09-08** (2026-09-04, UTC) | About 1.66 s (median of twelve warm renders at the bench, 2026-09-04 UTC; the slowest was 1.68 s). The same twelve prompts on the same machine at the old eight-pass setting measured 0.63 s, so thirty-two passes cost 2.63 times that — not four times, because a render carries a fixed cost that does not move with passes. That ratio is the ruler the block below still uses; the card it was measured on no longer draws |",
   "identity_basis": "row",
   "identity_token": "the machine retired from drawing on 2026-09-08",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    46561,
    47094
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "printlab-retired-drawing-machine/how-the-print-lab-works/one-plate-0-63-s",
   "identity": "printlab-retired-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one plate",
   "value": "0.63 s",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512 × 512, the old eight-pass setting",
    "concurrency": "not stated",
    "window": "the same twelve prompts, 2026-09-04 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    46561,
    47094
   ],
   "quote": "| One plate, 512 × 512, thirty-two steps, **on the machine retired from drawing on 2026-09-08** (2026-09-04, UTC) | About 1.66 s (median of twelve warm renders at the bench, 2026-09-04 UTC; the slowest was 1.68 s). The same twelve prompts on the same machine at the old eight-pass setting measured 0.63 s, so thirty-two passes cost 2.63 times that — not four times, because a render carries a fixed cost that does not move with passes. That ratio is the ruler the block below still uses; the card it was measured on no longer draws |",
   "identity_basis": "row",
   "identity_token": "the machine retired from drawing on 2026-09-08",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    46561,
    47094
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "printlab-retired-drawing-machine/how-the-print-lab-works/the-exhibit-s-own-ratio-thirty-two-passes-against-eight-2-63-times",
   "identity": "printlab-retired-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, thirty-two passes against eight",
   "value": "2.63 times",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "thirty-two passes, the old eight-pass setting",
    "concurrency": "not stated",
    "window": "2026-09-04 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    46561,
    47094
   ],
   "quote": "| One plate, 512 × 512, thirty-two steps, **on the machine retired from drawing on 2026-09-08** (2026-09-04, UTC) | About 1.66 s (median of twelve warm renders at the bench, 2026-09-04 UTC; the slowest was 1.68 s). The same twelve prompts on the same machine at the old eight-pass setting measured 0.63 s, so thirty-two passes cost 2.63 times that — not four times, because a render carries a fixed cost that does not move with passes. That ratio is the ruler the block below still uses; the card it was measured on no longer draws |",
   "identity_basis": "row",
   "identity_token": "the machine retired from drawing on 2026-09-08",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    46561,
    47094
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this ratio and says why it is not four times — a render carries a fixed cost that does not move with passes",
   "record_published": null
  },
  {
   "id": "printlab-drawing-machine/how-the-print-lab-works/one-plate-warm-about-1-6-s-a-plate-warm-1-56-1-70-s-over-four-renders",
   "identity": "printlab-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one plate, warm",
   "value": "about 1.6 s a plate warm (1.56–1.70 s over four renders)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "the old eight-pass setting",
    "concurrency": "not stated",
    "window": "2026-09-01"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    47095,
    47795
   ],
   "quote": "| The same plate on **the machine that draws today**, at the old eight-pass setting (2026-09-01) | **an eight-pass reading; the thirty-two-pass reading is in the block below.** The machine has two cards and at that sitting they did not match: one measured about 1.6 s a plate warm (1.56–1.70 s over four renders), the other about 6.8 s for the same work (median of five), and the block below says why. The lab draws on the second of those — that is the card it is pinned to — so at eight passes a picture cost about ten seconds of card time (9.86 s median for the three drawn as one batch, measured on that same card); the first plate after a cold start on that card measured 27 s by the same clock |",
   "identity_basis": "row",
   "identity_token": "the machine that draws today",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    47095,
    47795
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the first of the machine's two cards at that sitting",
   "record_published": null
  },
  {
   "id": "printlab-drawing-machine/how-the-print-lab-works/one-plate-about-6-8-s-for-the-same-work-median-of-five",
   "identity": "printlab-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one plate",
   "value": "about 6.8 s for the same work (median of five)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "the old eight-pass setting",
    "concurrency": "not stated",
    "window": "2026-09-01"
   },
   "figure_definition": "v3",
   "status": "superseded",
   "superseded_by": {
    "slug": "how-the-print-lab-works",
    "cite": "numbers-so-it-isnt-just-talk",
    "span": [
     53238,
     53401
    ],
    "quote": "which is also why this card measured 6.8 s a plate against its twin's 1.6 s at the first sitting on 2026-09-01: it was never the slower card, it was the fuller one"
   },
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    47095,
    47795
   ],
   "quote": "| The same plate on **the machine that draws today**, at the old eight-pass setting (2026-09-01) | **an eight-pass reading; the thirty-two-pass reading is in the block below.** The machine has two cards and at that sitting they did not match: one measured about 1.6 s a plate warm (1.56–1.70 s over four renders), the other about 6.8 s for the same work (median of five), and the block below says why. The lab draws on the second of those — that is the card it is pinned to — so at eight passes a picture cost about ten seconds of card time (9.86 s median for the three drawn as one batch, measured on that same card); the first plate after a cold start on that card measured 27 s by the same clock |",
   "identity_basis": "row",
   "identity_token": "the machine that draws today",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    47095,
    47795
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "WITHDRAWN BY ITS OWN EXHIBIT: the card was not slower, it was fuller. The correction is quoted beneath the row",
   "record_published": null
  },
  {
   "id": "printlab-drawing-machine/how-the-print-lab-works/a-whole-picture-three-plates-measured-not-derived-18-5-s-a-picture",
   "identity": "printlab-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "a whole picture, three plates",
   "value": "**measured, not derived: 18.5 s a picture**",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "three plates drawn as one batch, 512 by 512, thirty-two passes",
    "concurrency": "not stated",
    "window": "2026-09-08 17:43:32Z to 17:47:28Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    48090,
    48309
   ],
   "quote": "| A whole picture — three plates — on the card that draws today | **measured, not derived: 18.5 s a picture** — the dated block below carries the window, the twelve readings' spread, the cold first draw and the charge |",
   "identity_basis": "row",
   "identity_token": "the card that draws today",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    48090,
    48309
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "printlab-drawing-machine/how-the-print-lab-works/a-whole-picture-with-two-other-models-resident-46-8-s-a-picture",
   "identity": "printlab-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "a whole picture, with two other models resident",
   "value": "46.8 s a picture",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "three plates drawn as one batch, thirty-two passes",
    "concurrency": "two other models resident beside the drawing tool",
    "window": "four hours earlier, 2026-09-08"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    53105,
    53234
   ],
   "quote": "measured 46.8 s a picture — two and a half times slower — because a card that is full streams its weights instead of holding them",
   "identity_basis": "section",
   "identity_token": "the one card that draws for the print lab",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    52122,
    52163
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "recon T-19: residency is the condition, and the page retracts an earlier reading on exactly this ground",
   "record_published": null
  },
  {
   "id": "printlab-drawing-machine/how-the-print-lab-works/a-picture-drawn-as-one-batch-9-86-s-median-for-the-three-drawn-as-one-batch",
   "identity": "printlab-drawing-machine",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "a picture, drawn as one batch",
   "value": "9.86 s median for the three drawn as one batch",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "three plates drawn as one batch, eight passes",
    "concurrency": "not stated",
    "window": "2026-09-01"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    47095,
    47795
   ],
   "quote": "| The same plate on **the machine that draws today**, at the old eight-pass setting (2026-09-01) | **an eight-pass reading; the thirty-two-pass reading is in the block below.** The machine has two cards and at that sitting they did not match: one measured about 1.6 s a plate warm (1.56–1.70 s over four renders), the other about 6.8 s for the same work (median of five), and the block below says why. The lab draws on the second of those — that is the card it is pinned to — so at eight passes a picture cost about ten seconds of card time (9.86 s median for the three drawn as one batch, measured on that same card); the first plate after a cold start on that card measured 27 s by the same clock |",
   "identity_basis": "row",
   "identity_token": "the machine that draws today",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    47095,
    47795
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rig-b-the-laptop/how-the-print-lab-works/the-whole-board-as-one-sprite-144-cells-5-191-824-bytes-5-2-mb-1-12-s",
   "identity": "rig-b-the-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the whole board as one sprite, 144 cells",
   "value": "**5,191,824 bytes (5.2 MB), 1.12 s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "144 cells, 128 pixels a cell",
    "concurrency": "not stated",
    "window": "2026-09-09 15:45:32Z to 15:46:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "how-the-print-lab-works",
    "section": "numbers-so-it-isnt-just-talk",
    "md_sha": "c1c736f852132b230b47758546fa66655cfd05245ba9f867888a66b162d34780"
   },
   "cite": "numbers-so-it-isnt-just-talk",
   "span": [
    49729,
    50036
   ],
   "quote": "| The whole board as one sprite, 144 cells | 5,330,428 bytes (5.3 MB), 1.12 s to encode, on a full board built from the first set's twelve first-light renders used twelve times each (2026-09-02); **5,191,824 bytes (5.2 MB), 1.12 s**, on the set that hangs, the same way (2026-09-09 — the dated note below) |",
   "identity_basis": "section",
   "identity_token": "the workshop's laptop",
   "identity_cite": "numbers-so-it-isnt-just-talk",
   "identity_span": [
    54665,
    54686
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-typist-and-the-developer/writing-speed-207-7-tok-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed",
   "value": "**207.7 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "the chat seat",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "alone",
    "window": "the whole run took 35 seconds, on 2026-08-26"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "one-card-both-jobs-at-once",
    "md_sha": "ca760b1b9efa845a816ddbefb162b06079c0a2b03afe8ab249aeeceefd5ff812"
   },
   "cite": "one-card-both-jobs-at-once",
   "span": [
    6057,
    6124
   ],
   "quote": "| Tokens, alone | Writing speed | **207.7 tok/s** | 207.5 – 207.7 |",
   "identity_basis": "section",
   "identity_token": "the 96 GB class",
   "identity_cite": "one-card-both-jobs-at-once",
   "identity_span": [
    5081,
    5096
   ],
   "cell": {
    "column": 2,
    "header_span": [
     6002,
     6038
    ],
    "header": "| Row | Measured | Median | Spread |",
    "head_cell": "Median"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-typist-and-the-developer/writing-speed-spread-207-5-207-7",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, spread",
   "value": "207.5 – 207.7",
   "non_figure": null,
   "conditions": {
    "model": "the chat seat",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "alone",
    "window": "the whole run took 35 seconds, on 2026-08-26"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "one-card-both-jobs-at-once",
    "md_sha": "ca760b1b9efa845a816ddbefb162b06079c0a2b03afe8ab249aeeceefd5ff812"
   },
   "cite": "one-card-both-jobs-at-once",
   "span": [
    6057,
    6124
   ],
   "quote": "| Tokens, alone | Writing speed | **207.7 tok/s** | 207.5 – 207.7 |",
   "identity_basis": "section",
   "identity_token": "the 96 GB class",
   "identity_cite": "one-card-both-jobs-at-once",
   "identity_span": [
    5081,
    5096
   ],
   "cell": {
    "column": 3,
    "header_span": [
     6002,
     6038
    ],
    "header": "| Row | Measured | Median | Spread |",
    "head_cell": "Spread"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-typist-and-the-developer/seconds-per-sketch-1-54-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per sketch",
   "value": "**1.54 s**",
   "non_figure": null,
   "conditions": {
    "model": "the game's own sketch graph",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "sixteen-pass",
    "concurrency": "alone",
    "window": "the whole run took 35 seconds, on 2026-08-26"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "one-card-both-jobs-at-once",
    "md_sha": "ca760b1b9efa845a816ddbefb162b06079c0a2b03afe8ab249aeeceefd5ff812"
   },
   "cite": "one-card-both-jobs-at-once",
   "span": [
    6125,
    6192
   ],
   "quote": "| Pictures, alone | Seconds per sketch | **1.54 s** | 1.54 – 1.67 |",
   "identity_basis": "section",
   "identity_token": "the 96 GB class",
   "identity_cite": "one-card-both-jobs-at-once",
   "identity_span": [
    5081,
    5096
   ],
   "cell": {
    "column": 2,
    "header_span": [
     6002,
     6038
    ],
    "header": "| Row | Measured | Median | Spread |",
    "head_cell": "Median"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-typist-and-the-developer/writing-speed-51-6-tok-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed",
   "value": "**51.6 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "the chat seat",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "while pictures develop",
    "window": "the whole run took 35 seconds, on 2026-08-26"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "one-card-both-jobs-at-once",
    "md_sha": "ca760b1b9efa845a816ddbefb162b06079c0a2b03afe8ab249aeeceefd5ff812"
   },
   "cite": "one-card-both-jobs-at-once",
   "span": [
    6193,
    6278
   ],
   "quote": "| Tokens, **while pictures develop** | Writing speed | **51.6 tok/s** | 40.7 – 75.6 |",
   "identity_basis": "section",
   "identity_token": "the 96 GB class",
   "identity_cite": "one-card-both-jobs-at-once",
   "identity_span": [
    5081,
    5096
   ],
   "cell": {
    "column": 2,
    "header_span": [
     6002,
     6038
    ],
    "header": "| Row | Measured | Median | Spread |",
    "head_cell": "Median"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-typist-and-the-developer/seconds-per-sketch-1-82-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per sketch",
   "value": "**1.82 s**",
   "non_figure": null,
   "conditions": {
    "model": "the game's own sketch graph",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "sixteen-pass",
    "concurrency": "while tokens stream",
    "window": "the whole run took 35 seconds, on 2026-08-26"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "one-card-both-jobs-at-once",
    "md_sha": "ca760b1b9efa845a816ddbefb162b06079c0a2b03afe8ab249aeeceefd5ff812"
   },
   "cite": "one-card-both-jobs-at-once",
   "span": [
    6279,
    6364
   ],
   "quote": "| Pictures, **while tokens stream** | Seconds per sketch | **1.82 s** | 1.80 – 1.91 |",
   "identity_basis": "section",
   "identity_token": "the 96 GB class",
   "identity_cite": "one-card-both-jobs-at-once",
   "identity_span": [
    5081,
    5096
   ],
   "cell": {
    "column": 2,
    "header_span": [
     6002,
     6038
    ],
    "header": "| Row | Measured | Median | Spread |",
    "head_cell": "Median"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-ceiling-is-not-the-corpus/seconds-per-optimiser-step-5-73",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per optimiser step",
   "value": "5.73",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "331 tracks five times over, 415 small corrections in all",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-ceiling-is-not-the-corpus",
    "section": "forty-minutes-on-a-card-bought-over-five-years-ago",
    "md_sha": "52096f57cc48d72deecad5ef569c9b9fe6db895d9b610232e1ad866f78a390bd"
   },
   "cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "span": [
    12654,
    12691
   ],
   "quote": "| seconds per optimiser step | 5.73 |",
   "identity_basis": "section",
   "identity_token": "an **RTX 3090 purchased over five years ago**",
   "identity_cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "identity_span": [
    12021,
    12066
   ],
   "cell": {
    "column": 1,
    "header_span": [
     12542,
     12564
    ],
    "header": "| the run | measured |",
    "head_cell": "measured"
   },
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-ceiling-is-not-the-corpus/peak-memory-the-trainer-s-own-allocator-5-7-gib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "peak memory, the trainer's own allocator",
   "value": "5.7 GiB",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "331 tracks five times over, 415 small corrections in all",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-ceiling-is-not-the-corpus",
    "section": "forty-minutes-on-a-card-bought-over-five-years-ago",
    "md_sha": "52096f57cc48d72deecad5ef569c9b9fe6db895d9b610232e1ad866f78a390bd"
   },
   "cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "span": [
    12692,
    12802
   ],
   "quote": "| peak memory, the trainer's own allocator | 5.7 GiB (the card's own draw reads higher; both are in the kit) |",
   "identity_basis": "section",
   "identity_token": "an **RTX 3090 purchased over five years ago**",
   "identity_cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "identity_span": [
    12021,
    12066
   ],
   "cell": {
    "column": 1,
    "header_span": [
     12542,
     12564
    ],
    "header": "| the run | measured |",
    "head_cell": "measured"
   },
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-free-speed-wasnt-free/production-median-decode-138-6-tok-s-over-1-982-live-rulings",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "production median decode",
   "value": "**138.6 tok/s over 1,982 live rulings**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.9",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-15"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-free-speed-wasnt-free",
    "section": "the-day-the-box-got-faster",
    "md_sha": "c84eb8845e628e3656718788da4b08fe297ab36c14d7700f4d73f3ec212d3398"
   },
   "cite": "the-day-the-box-got-faster",
   "span": [
    2743,
    2819
   ],
   "quote": "a production median of **138.6 tok/s over 1,982 live rulings** on 2026-08-15",
   "identity_basis": "section",
   "identity_token": "a single 96 GB-class workstation card",
   "identity_cite": "the-day-the-box-got-faster",
   "identity_span": [
    2595,
    2632
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-free-speed-wasnt-free/sealed-bench-median-decode-139-55-n-6-range-137-75-150-49",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "sealed bench median decode",
   "value": "**139.55** (n=6, range 137.75–150.49)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.9",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-14"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-free-speed-wasnt-free",
    "section": "the-day-the-box-got-faster",
    "md_sha": "c84eb8845e628e3656718788da4b08fe297ab36c14d7700f4d73f3ec212d3398"
   },
   "cite": "the-day-the-box-got-faster",
   "span": [
    2874,
    2947
   ],
   "quote": "a sealed bench run on 2026-08-14 at **139.55** (n=6, range 137.75–150.49)",
   "identity_basis": "section",
   "identity_token": "a single 96 GB-class workstation card",
   "identity_cite": "the-day-the-box-got-faster",
   "identity_span": [
    2595,
    2632
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-free-speed-wasnt-free/decode-after-the-upgrade-202-214-tok-s",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode after the upgrade",
   "value": "**202–214 tok/s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "the service restarted at 2026-08-16T01:36:27Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-free-speed-wasnt-free",
    "section": "the-day-the-box-got-faster",
    "md_sha": "c84eb8845e628e3656718788da4b08fe297ab36c14d7700f4d73f3ec212d3398"
   },
   "cite": "the-day-the-box-got-faster",
   "span": [
    3580,
    3646
   ],
   "quote": "the same model on the same card started printing **202–214 tok/s**",
   "identity_basis": "section",
   "identity_token": "a single 96 GB-class workstation card",
   "identity_cite": "the-day-the-box-got-faster",
   "identity_span": [
    2595,
    2632
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-free-speed-wasnt-free/decode-at-a-deeper-context-tier-124-975",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode at a deeper context tier",
   "value": "124.975",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.9",
    "cap": "not stated",
    "context": "a much deeper context tier",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-free-speed-wasnt-free",
    "section": "the-day-the-box-got-faster",
    "md_sha": "c84eb8845e628e3656718788da4b08fe297ab36c14d7700f4d73f3ec212d3398"
   },
   "cite": "the-day-the-box-got-faster",
   "span": [
    3173,
    3264
   ],
   "quote": "one bench reading sits lower — 124.975 — because it was taken at a much deeper context tier",
   "identity_basis": "section",
   "identity_token": "a single 96 GB-class workstation card",
   "identity_cite": "the-day-the-box-got-faster",
   "identity_span": [
    2595,
    2632
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-24g-class/the-typist-and-the-developer/painter-wake-up-about-fifty-four-seconds",
   "identity": "the-24g-class",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "painter wake-up",
   "value": "about fifty-four seconds",
   "non_figure": null,
   "conditions": {
    "model": "the same twelve-billion model in a heavier build",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "The July notes"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-typist-and-the-developer",
    "section": "the-character-who-moved-one-door-over",
    "md_sha": "b41516d2090657334093ef26db003a7cd8da08defd496c6619372f49f5fc9e2d"
   },
   "cite": "the-character-who-moved-one-door-over",
   "span": [
    9754,
    9778
   ],
   "quote": "about fifty-four seconds",
   "identity_basis": "section",
   "identity_token": "a 24 GB rig",
   "identity_cite": "the-character-who-moved-one-door-over",
   "identity_span": [
    8867,
    8878
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-24g-class/fifty-seven-milliseconds/seconds-per-step-fitted-a-step-costs-56-9-milliseconds",
   "identity": "the-24g-class",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "seconds per step, fitted",
   "value": "**a step costs 56.9 milliseconds**",
   "non_figure": null,
   "conditions": {
    "model": "AlbedoBase XL v2.1",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512×512, dpmpp_2m/karras, seed 424242",
    "concurrency": "not stated",
    "window": "measured 2026-07-10 by the archive's own clock"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "fifty-seven-milliseconds",
    "section": "finding-one-a-step-is-a-fixed-price-good",
    "md_sha": "b1ece6e9316e29cd26ba31f9502a8c47a985fdd78f2f67aab012fe424ea93048"
   },
   "cite": "finding-one-a-step-is-a-fixed-price-good",
   "span": [
    3422,
    3484
   ],
   "quote": "**a step costs 56.9 milliseconds** at 512² on the 24G VRAM rig",
   "identity_basis": "row",
   "identity_token": "the 24G VRAM rig",
   "identity_cite": "finding-one-a-step-is-a-fixed-price-good",
   "identity_span": [
    3422,
    3484
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-24g-class/fifty-seven-milliseconds/the-fitted-line-seconds-0-238-0-0569-steps",
   "identity": "the-24g-class",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the fitted line",
   "value": "`seconds = 0.238 + 0.0569 × steps`",
   "non_figure": null,
   "conditions": {
    "model": "AlbedoBase XL v2.1",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "512×512, per-rung medians",
    "concurrency": "not stated",
    "window": "measured 2026-07-10 by the archive's own clock"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "fifty-seven-milliseconds",
    "section": "finding-one-a-step-is-a-fixed-price-good",
    "md_sha": "b1ece6e9316e29cd26ba31f9502a8c47a985fdd78f2f67aab012fe424ea93048"
   },
   "cite": "finding-one-a-step-is-a-fixed-price-good",
   "span": [
    3203,
    3249
   ],
   "quote": "the line is `seconds = 0.238 + 0.0569 × steps`",
   "identity_basis": "section",
   "identity_token": "the 24G VRAM rig",
   "identity_cite": "finding-one-a-step-is-a-fixed-price-good",
   "identity_span": [
    3422,
    3484
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/fifty-seven-milliseconds/seconds-per-render-at-four-step-counts-0-77-1-08-1-51-1-89-and-3-52-seconds",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "seconds per render at four step counts",
   "value": "0.77, 1.08, 1.51, 1.89 and 3.52 seconds",
   "non_figure": null,
   "conditions": {
    "model": "the shipped painter",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "768², at 4, 8, 12, 16 and 32",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "fifty-seven-milliseconds",
    "section": "finding-four-the-argument-with-our-own-settings-and-what-it",
    "md_sha": "829a290df95bdd6e68e7f7e48445fdce6ad3eee6eecae295eb59b9863703830b"
   },
   "cite": "finding-four-the-argument-with-our-own-settings-and-what-it",
   "span": [
    11323,
    11395
   ],
   "quote": "the shipped painter’s rungs read 0.77, 1.08, 1.51, 1.89 and 3.52 seconds",
   "identity_basis": "section",
   "identity_token": "the 96G VRAM workstation",
   "identity_cite": "finding-four-the-argument-with-our-own-settings-and-what-it",
   "identity_span": [
    11258,
    11282
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this section dates nothing — the page's first run is dated 2026-07-10 and this ladder is the second, undated where it sits",
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-new-kid/resident-vram-three-precisions-17-5-gb-29-2-gb-53-1-gb",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "resident VRAM, three precisions",
   "value": "17.5 GB | 29.2 GB | 53.1 GB",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.8:27b",
    "quant": "all three precisions",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32768",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-new-kid",
    "section": "the-other-precisions",
    "md_sha": "200773ebbefedf9a9c621b4a2a86db1b6af2c133f4d3f139c84b72a1b5a01edc"
   },
   "cite": "the-other-precisions",
   "span": [
    42707,
    42734
   ],
   "quote": "17.5 GB | 29.2 GB | 53.1 GB",
   "identity_basis": "section",
   "identity_token": "a 96G-VRAM box",
   "identity_cite": "the-other-precisions",
   "identity_span": [
    41072,
    41086
   ],
   "cell": null,
   "conditions_cites": [
    "what-it-ran-on"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/ten-minutes-with-living-artists/seconds-per-update-1-515-seconds-per-update",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "seconds per update",
   "value": "**1.515 seconds per update**",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "400 in all",
    "concurrency": "not stated",
    "window": "10 minutes 8 seconds"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ten-minutes-with-living-artists",
    "section": "ten-minutes-on-the-big-card",
    "md_sha": "13a68fb352d184c06e57098fe385a1ec5a98ad57b4b9cfbbf9fb036c8bcf127b"
   },
   "cite": "ten-minutes-on-the-big-card",
   "span": [
    13599,
    13627
   ],
   "quote": "**1.515 seconds per update**",
   "identity_basis": "section",
   "identity_token": "the lab's workstation-class card — 96 gigabytes of memory",
   "identity_cite": "ten-minutes-on-the-big-card",
   "identity_span": [
    12572,
    12629
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks",
    "where-part-two-left-off"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/ten-minutes-with-living-artists/peak-memory-two-instruments-5-7-gigabytes-inside-the-trainer-s-own-allocator-6-7",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "peak memory, two instruments",
   "value": "5.7 gigabytes inside the trainer's own allocator, 6.7 by subtracting the card's occupancy before the run from its peak during it",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "400 in all",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ten-minutes-with-living-artists",
    "section": "ten-minutes-on-the-big-card",
    "md_sha": "13a68fb352d184c06e57098fe385a1ec5a98ad57b4b9cfbbf9fb036c8bcf127b"
   },
   "cite": "ten-minutes-on-the-big-card",
   "span": [
    13823,
    13951
   ],
   "quote": "5.7 gigabytes inside the\ntrainer's own allocator, 6.7 by subtracting the card's occupancy before the run from its peak\nduring it",
   "identity_basis": "section",
   "identity_token": "the lab's workstation-class card — 96 gigabytes of memory",
   "identity_cite": "ten-minutes-on-the-big-card",
   "identity_span": [
    12572,
    12629
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks",
    "where-part-two-left-off"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/six-worlds-painted-large/seconds-per-painting-41-0-to-42-2-seconds-per-painting-by-the-render-engine-s-ow",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "seconds per painting",
   "value": "41.0 to 42.2 seconds per painting by the render engine's own clock",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.1-dev",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "2048×1280, 40 steps",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "six-worlds-painted-large",
    "section": "for-the-technically-curious",
    "md_sha": "8f1197ccb18550e4919a83282e398d38c323bd5ac36e11c6a1993b5bff79bbe6"
   },
   "cite": "for-the-technically-curious",
   "span": [
    8307,
    8373
   ],
   "quote": "41.0 to 42.2 seconds per painting by the render engine's own clock",
   "identity_basis": "section",
   "identity_token": "our 96 GB workstation-class card",
   "identity_cite": "for-the-technically-curious",
   "identity_span": [
    8218,
    8250
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/cove-voice-head-to-head/decode-wall-clock-under-load-37-68-tok-s-median-55",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, wall-clock, under load",
   "value": "37–68 tok/s (median 55)",
   "non_figure": null,
   "conditions": {
    "model": "qwen3.6:27b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "taken while the full chair-trials bench shared this card",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "cove-voice-head-to-head",
    "section": "same-question-three-voices",
    "md_sha": "55f7d17c6589a7e816c2c12f548411d2d9672fb962eba7dc056fcfeda075d89b"
   },
   "cite": "same-question-three-voices",
   "span": [
    18527,
    18562
   ],
   "quote": "qwen3.6:27b 37–68 tok/s (median 55)",
   "identity_basis": "page",
   "identity_token": "the 96G workstation",
   "identity_cite": "the-rig-disclosed-both-ways",
   "identity_span": [
    7829,
    7895
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/cove-voice-head-to-head/decode-wall-clock-under-load-64-102-tok-s-median-94",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, wall-clock, under load",
   "value": "64–102 tok/s (median 94)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:12b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "taken while the full chair-trials bench shared this card",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "cove-voice-head-to-head",
    "section": "same-question-three-voices",
    "md_sha": "55f7d17c6589a7e816c2c12f548411d2d9672fb962eba7dc056fcfeda075d89b"
   },
   "cite": "same-question-three-voices",
   "span": [
    18565,
    18600
   ],
   "quote": "gemma4:12b 64–102 tok/s (median 94)",
   "identity_basis": "page",
   "identity_token": "the 96G workstation",
   "identity_cite": "the-rig-disclosed-both-ways",
   "identity_span": [
    7829,
    7895
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/cove-voice-head-to-head/decode-wall-clock-under-load-50-139-tok-s-median-75",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode, wall-clock, under load",
   "value": "50–139 tok/s (median 75)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "taken while the full chair-trials bench shared this card",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "cove-voice-head-to-head",
    "section": "same-question-three-voices",
    "md_sha": "55f7d17c6589a7e816c2c12f548411d2d9672fb962eba7dc056fcfeda075d89b"
   },
   "cite": "same-question-three-voices",
   "span": [
    18603,
    18638
   ],
   "quote": "gemma4:26b 50–139 tok/s (median 75)",
   "identity_basis": "page",
   "identity_token": "the 96G workstation",
   "identity_cite": "the-rig-disclosed-both-ways",
   "identity_span": [
    7829,
    7895
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/the-ceiling-is-not-the-corpus/the-exhibit-s-own-cross-card-ratio-the-previous-adapter-trained-at-1-5-seconds-p",
   "identity": "rtx-pro-6000-blackwell",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own cross-card ratio",
   "value": "the previous adapter trained at 1.5 seconds per step on a 96 GB workstation card, so this is roughly 3.8 times slower on hardware anyone can buy used",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-ceiling-is-not-the-corpus",
    "section": "forty-minutes-on-a-card-bought-over-five-years-ago",
    "md_sha": "52096f57cc48d72deecad5ef569c9b9fe6db895d9b610232e1ad866f78a390bd"
   },
   "cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "span": [
    13639,
    13788
   ],
   "quote": "the previous adapter trained at\n1.5 seconds per step on a 96 GB workstation card, so this is roughly 3.8 times slower on hardware\nanyone can buy used",
   "identity_basis": "row",
   "identity_token": "a 96 GB workstation card",
   "identity_cite": "forty-minutes-on-a-card-bought-over-five-years-ago",
   "identity_span": [
    13639,
    13788
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the exhibit prints this comparison; the roster computes none",
   "record_published": null
  },
  {
   "id": "rtx-3090/chopin-in-five-minutes/wall-clock-4-minutes-49-seconds-by-the-trainer-s-own-clock",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "wall clock",
   "value": "**4 minutes 49 seconds** by the trainer's own clock",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "250 watts",
    "context": "not stated",
    "batch_or_steps": "ten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1)",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "chopin-in-five-minutes",
    "section": "five-minutes-of-training-honestly-accounted",
    "md_sha": "0ab15547aee1dd9b3c49eba86e2a95b69e83f5fbb19f2357049306961758b5c5"
   },
   "cite": "five-minutes-of-training-honestly-accounted",
   "span": [
    6410,
    6461
   ],
   "quote": "**4 minutes 49 seconds** by the trainer's own clock",
   "identity_basis": "section",
   "identity_token": "A five-year-old RTX 3090 — 24 gigabytes of memory, power-capped at 250 watts",
   "identity_cite": "five-minutes-of-training-honestly-accounted",
   "identity_span": [
    5928,
    6004
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/chopin-in-five-minutes/peak-memory-two-instruments-5-7-gigabytes-by-the-trainer-s-own-bookkeeping-the-c",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "peak memory, two instruments",
   "value": "5.7 gigabytes by the trainer's own bookkeeping; the card itself showed 6.8 in use",
   "non_figure": null,
   "conditions": {
    "model": "ACE-Step 1.5",
    "quant": "not stated",
    "runtime": "Side-Step",
    "runtime_version": "not stated",
    "cap": "250 watts",
    "context": "not stated",
    "batch_or_steps": "ten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1)",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "chopin-in-five-minutes",
    "section": "five-minutes-of-training-honestly-accounted",
    "md_sha": "0ab15547aee1dd9b3c49eba86e2a95b69e83f5fbb19f2357049306961758b5c5"
   },
   "cite": "five-minutes-of-training-honestly-accounted",
   "span": [
    6561,
    6642
   ],
   "quote": "5.7 gigabytes by the\ntrainer's own bookkeeping; the card itself showed 6.8 in use",
   "identity_basis": "section",
   "identity_token": "A five-year-old RTX 3090 — 24 gigabytes of memory, power-capped at 250 watts",
   "identity_cite": "five-minutes-of-training-honestly-accounted",
   "identity_span": [
    5928,
    6004
   ],
   "cell": null,
   "conditions_cites": [
    "who-ran-this-and-thanks"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-retiring-8g-laptop/the-move/usable-vram-8-gb-on-paper-and-closer-to-6-gb-usable-once-the-operating-system-an",
   "identity": "the-retiring-8g-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "usable VRAM",
   "value": "8 GB on paper, and closer to 6 GB usable once the operating system and display took their cut",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-move",
    "section": "the-machine",
    "md_sha": "24e8b5c3fdbc1ca2c094ebf009268bc6874c2e0902617c4bb7d0ef303157149a"
   },
   "cite": "the-machine",
   "span": [
    347,
    505
   ],
   "quote": "The retiring server is a gaming laptop with an 8 GB mobile GPU — 8 GB on paper, and closer to 6 GB usable once the operating system and display took their cut",
   "identity_basis": "row",
   "identity_token": "a gaming laptop with an 8 GB mobile GPU — 8 GB on paper, and closer to 6 GB usable",
   "identity_cite": "the-machine",
   "identity_span": [
    347,
    505
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-8g-class/voice-trials/loaded-and-whether-it-fits-loaded-5-9-gb-warm-turn-5s",
   "identity": "the-8g-class",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "loaded, and whether it fits",
   "value": "**loaded: 5.9 GB · warm turn: ~5s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma2:9b (+q8 KV)",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "voice-trials",
    "section": "era-one-the-8g-vram-rig-nine-models",
    "md_sha": "13a914aa16e0a9421fc4aa98a2dffafb80b9f5a91d5450d45e0ef7c5a300fa29"
   },
   "cite": "era-one-the-8g-vram-rig-nine-models",
   "span": [
    2402,
    2458
   ],
   "quote": "**gemma2:9b (+q8 KV) · loaded: 5.9 GB · warm turn: ~5s**",
   "identity_basis": "section",
   "identity_token": "the 8G VRAM rig",
   "identity_cite": "era-one-the-8g-vram-rig-nine-models",
   "identity_span": [
    2333,
    2348
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-8g-class/voice-trials/loaded-and-whether-it-fits-loaded-3-3-gb-warm-turn-5s",
   "identity": "the-8g-class",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "loaded, and whether it fits",
   "value": "**loaded: 3.3 GB · warm turn: ~5s**",
   "non_figure": null,
   "conditions": {
    "model": "gemma3:4b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "voice-trials",
    "section": "era-one-the-8g-vram-rig-nine-models",
    "md_sha": "13a914aa16e0a9421fc4aa98a2dffafb80b9f5a91d5450d45e0ef7c5a300fa29"
   },
   "cite": "era-one-the-8g-vram-rig-nine-models",
   "span": [
    2626,
    2673
   ],
   "quote": "**gemma3:4b · loaded: 3.3 GB · warm turn: ~5s**",
   "identity_basis": "section",
   "identity_token": "the 8G VRAM rig",
   "identity_cite": "era-one-the-8g-vram-rig-nine-models",
   "identity_span": [
    2333,
    2348
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-mini-pc/two-hours-on-battery/decode-on-battery-and-how-long-it-lasts-12-tokens-second-for-about-two-hours-on-",
   "identity": "the-mini-pc",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode on battery, and how long it lasts",
   "value": "~12 tokens/second for about two hours on battery — and sit ready to answer for seven and a half",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-31, 02:38:31–03:05:02Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    3359,
    3479
   ],
   "quote": "will run a 26 B model at\n~12 tokens/second for about two hours on battery — and sit ready to answer for seven and a half",
   "identity_basis": "section",
   "identity_token": "this bench's mini rig",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3615,
    3636
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-ups-s-own-runtime-estimate-7-300-s-122-minutes",
   "identity": "cyberpower-pr1500",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the UPS's own runtime estimate",
   "value": "**~7,300 s ≈ 122 minutes**",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-31, 02:38:31–03:05:02Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5237,
    5330
   ],
   "quote": "the UPS's own estimator settled around\n**~7,300 s ≈ 122 minutes** under continuous generation",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "one of the page's two independent readings of how long it would last; the other is the charge-burn arithmetic below, and the page's own conclusion is that two different instruments give the same answer",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/how-long-it-would-last-by-the-charge-burn-arithmetic-130-minutes",
   "identity": "cyberpower-pr1500",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "how long it would last, by the charge-burn arithmetic",
   "value": "**~130 minutes**",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-31, 02:38:31–03:05:02Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5370,
    5461
   ],
   "quote": "(4 percentage points in 6 min 04 s, run to the 10% low-battery floor) says **~130 minutes**",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's second reading, from the charge it actually burned rather than from the unit's own estimator",
   "record_published": null
  },
  {
   "id": "cyberpower-pr1500/two-hours-on-battery/the-inverter-s-own-tax-on-battery-draw-above-on-mains-3-w",
   "identity": "cyberpower-pr1500",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the inverter's own tax — on-battery draw above on-mains",
   "value": "~3 W",
   "non_figure": null,
   "conditions": {
    "model": "a 26-billion-parameter model",
    "quant": "4-bit",
    "runtime": "Ollama",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "32 K context pinned",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-08-31, 02:38:31–03:05:02Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-hours-on-battery",
    "section": "two-hours-at-12-tok-s-the-test-that-named-the-page",
    "md_sha": "77ff8b3ac9a915749617346e552ff636cb7d4edf5ddf2b158e1ac6eda83299aa"
   },
   "cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "span": [
    5565,
    5657
   ],
   "quote": "on-battery draw ran ~3 W above on-mains for the same work — the inverter's own tax,\nmeasured",
   "identity_basis": "section",
   "identity_token": "a CyberPower PR1500 — 1500 VA on the label, 900 W by its own firmware's nominal, a consumer rack/tower UPS",
   "identity_cite": "two-hours-at-12-tok-s-the-test-that-named-the-page",
   "identity_span": [
    3919,
    4025
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-our-gpu-benches-never-ask",
    "the-one-table-to-remember",
    "the-wait-prefill-is-the-real-cpu-tax",
    "set-the-thread-count-the-one-free-lever-and-the-exception",
    "read-this-before-the-table-cpu-speed-is-a-memory-bandwidth"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-114-8",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "114.8",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12634,
    12693
   ],
   "quote": "| Mixture, at 131,072 | 114.8 | 115.4 | 125.6 | **136.1** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 1,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-115-4",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "115.4",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12634,
    12693
   ],
   "quote": "| Mixture, at 131,072 | 114.8 | 115.4 | 125.6 | **136.1** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 2,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-tokens-a-second-125-6",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "125.6",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12634,
    12693
   ],
   "quote": "| Mixture, at 131,072 | 114.8 | 115.4 | 125.6 | **136.1** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 3,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-tokens-a-second-136-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**136.1**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12634,
    12693
   ],
   "quote": "| Mixture, at 131,072 | 114.8 | 115.4 | 125.6 | **136.1** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 4,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-44-6",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**44.6**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "49,152",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12694,
    12819
   ],
   "quote": "| Dense, at 49,152 | **44.6** | **44.7** | 33.1 | *not run — its 300 W dense arm is a single point at 65,536, not a ladder* |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 1,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-44-7",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**44.7**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "49,152",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12694,
    12819
   ],
   "quote": "| Dense, at 49,152 | **44.6** | **44.7** | 33.1 | *not run — its 300 W dense arm is a single point at 65,536, not a ladder* |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 2,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-tokens-a-second-33-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "33.1",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "49,152",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12694,
    12819
   ],
   "quote": "| Dense, at 49,152 | **44.6** | **44.7** | 33.1 | *not run — its 300 W dense arm is a single point at 65,536, not a ladder* |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 3,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-does-not-fit",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "decode, tokens a second",
   "value": "Does not fit",
   "non_figure": "does not fit",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12820,
    12888
   ],
   "quote": "| Dense, at 65,536 | Does not fit | Does not fit | 33.3 | **47.8** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 1,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "20 GB between two cards does not hold the dense model at this window; the pair's own ceiling is the 49,152 row above",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-tokens-a-second-does-not-fit-2",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "decode, tokens a second",
   "value": "Does not fit",
   "non_figure": "does not fit",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12820,
    12888
   ],
   "quote": "| Dense, at 65,536 | Does not fit | Does not fit | 33.3 | **47.8** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 2,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-tokens-a-second-33-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "33.3",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12820,
    12888
   ],
   "quote": "| Dense, at 65,536 | Does not fit | Does not fit | 33.3 | **47.8** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 3,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-tokens-a-second-47-8",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**47.8**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12820,
    12888
   ],
   "quote": "| Dense, at 65,536 | Does not fit | Does not fit | 33.3 | **47.8** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    12524,
    12607
   ],
   "cell": {
    "column": 4,
    "header_span": [
     12524,
     12607
    ],
    "header": "| Model | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/largest-window-held-whole-placement-left-to-the-server-131-072-whole",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, placement left to the server",
   "value": "131,072 whole",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16250,
    16351
   ],
   "quote": "| Mixture, placement left to the server | **never held whole at any window tested** | 131,072 whole |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-131-072-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "131,072 whole, 19,531 MiB used",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16352,
    16452
   ],
   "quote": "| Mixture, the layer count named | 131,072 whole, 19,531 MiB used | 131,072 whole, 20,129 MiB used |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-131-072-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "131,072 whole, 20,129 MiB used",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16352,
    16452
   ],
   "quote": "| Mixture, the layer count named | 131,072 whole, 19,531 MiB used | 131,072 whole, 20,129 MiB used |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/residency-under-the-server-s-own-placement-never-measured-as-a-ladder-at-a-49-15",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "residency under the server's own placement",
   "value": "Never measured as a ladder; at a 49,152 window the planner kept **73.7 %** on the cards",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16453,
    16597
   ],
   "quote": "| Dense, placement left to the server | Never measured as a ladder; at a 49,152 window the planner kept **73.7 %** on the cards | 65,536 whole |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/residency-under-the-server-s-own-placement-65-536-whole",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "residency under the server's own placement",
   "value": "65,536 whole",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16453,
    16597
   ],
   "quote": "| Dense, placement left to the server | Never measured as a ladder; at a 49,152 window the planner kept **73.7 %** on the cards | 65,536 whole |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-49-152-wh",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "49,152 whole, 18,909 MiB used",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16598,
    16698
   ],
   "quote": "| Dense, the layer count named | 49,152 whole, 18,909 MiB used | **98,304 whole**, 23,469 MiB used |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-98-304-wh",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "**98,304 whole**, 23,469 MiB used",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16598,
    16698
   ],
   "quote": "| Dense, the layer count named | 49,152 whole, 18,909 MiB used | **98,304 whole**, 23,469 MiB used |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-and-residency-out-of-the-box-the-server-s-own-placement-59-5-tok-s-75-6-o",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode and residency out of the box, the server's own placement",
   "value": "**59.5 tok/s**, 75.6 % on the cards",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16699,
    16820
   ],
   "quote": "| Mixture at a 4,096 window, out of the box | **59.5 tok/s**, 75.6 % on the cards | **129.4 tok/s**, 100 % on the cards |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 1,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-and-residency-out-of-the-box-the-server-s-own-placement-129-4-tok-s-100-o",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode and residency out of the box, the server's own placement",
   "value": "**129.4 tok/s**, 100 % on the cards",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    16699,
    16820
   ],
   "quote": "| Mixture at a 4,096 window, out of the box | **59.5 tok/s**, 75.6 % on the cards | **129.4 tok/s**, 100 % on the cards |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "how-much-can-each-one-hold",
   "identity_span": [
    16154,
    16235
   ],
   "cell": {
    "column": 2,
    "header_span": [
     16154,
     16235
    ],
    "header": "| What was asked of the server | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/reading-98-024-tokens-seconds-28-36-s",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "reading 98,024 tokens, seconds",
   "value": "**28.36 s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    18918,
    18967
   ],
   "quote": "| Reading 98,024 tokens | **28.36 s** | 45.12 s |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 1,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/reading-98-024-tokens-seconds-45-12-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "reading 98,024 tokens, seconds",
   "value": "45.12 s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    18918,
    18967
   ],
   "quote": "| Reading 98,024 tokens | **28.36 s** | 45.12 s |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/time-to-the-first-token-seconds-30-04-s",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to the first token, seconds",
   "value": "**30.04 s**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    18968,
    19015
   ],
   "quote": "| First token arrives | **30.04 s** | 46.64 s |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 1,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/time-to-the-first-token-seconds-46-64-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "time to the first token, seconds",
   "value": "46.64 s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    18968,
    19015
   ],
   "quote": "| First token arrives | **30.04 s** | 46.64 s |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-once-the-window-is-full-tokens-a-second-61-5-tok-s",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode once the window is full, tokens a second",
   "value": "61.5 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    19016,
    19076
   ],
   "quote": "| Writing, once the window is full | 61.5 tok/s | **72.8** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 1,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-once-the-window-is-full-tokens-a-second-72-8",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode once the window is full, tokens a second",
   "value": "**72.8**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    19016,
    19076
   ],
   "quote": "| Writing, once the window is full | 61.5 tok/s | **72.8** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/energy-at-depth-joules-per-1-000-tokens-5-837",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy at depth, joules per 1,000 tokens",
   "value": "5,837",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    19077,
    19141
   ],
   "quote": "| Energy at depth, joules per 1,000 tokens | 5,837 | **3,902** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 1,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/energy-at-depth-joules-per-1-000-tokens-3-902",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "energy at depth, joules per 1,000 tokens",
   "value": "**3,902**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    19077,
    19141
   ],
   "quote": "| Energy at depth, joules per 1,000 tokens | 5,837 | **3,902** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-about-a-long-document",
   "identity_span": [
    18861,
    18901
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18861,
     18901
    ],
    "header": "| | Two 3080s, 300 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/total-tokens-a-second-two-callers-189-5",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "total tokens a second, two callers",
   "value": "**189.5**",
   "non_figure": null,
   "conditions": {
    "model": "one 26-billion model split across both cards",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21542,
    21617
   ],
   "quote": "| One 26-billion model split across both cards | **189.5** | 94.8 | 356 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Total tokens a second"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tokens-a-second-each-caller-two-callers-94-8",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second each caller, two callers",
   "value": "94.8",
   "non_figure": null,
   "conditions": {
    "model": "one 26-billion model split across both cards",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21542,
    21617
   ],
   "quote": "| One 26-billion model split across both cards | **189.5** | 94.8 | 356 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Each caller"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-356-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "356 W",
   "non_figure": null,
   "conditions": {
    "model": "one 26-billion model split across both cards",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21542,
    21617
   ],
   "quote": "| One 26-billion model split across both cards | **189.5** | 94.8 | 356 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/total-tokens-a-second-two-callers-118-2",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "total tokens a second, two callers",
   "value": "118.2",
   "non_figure": null,
   "conditions": {
    "model": "two copies of a 12-billion model, one per card",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21618,
    21691
   ],
   "quote": "| Two copies of a 12-billion model, one per card | 118.2 | 59.1 | 542 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Total tokens a second"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tokens-a-second-each-caller-two-callers-59-1",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second each caller, two callers",
   "value": "59.1",
   "non_figure": null,
   "conditions": {
    "model": "two copies of a 12-billion model, one per card",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21618,
    21691
   ],
   "quote": "| Two copies of a 12-billion model, one per card | 118.2 | 59.1 | 542 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Each caller"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-542-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "542 W",
   "non_figure": null,
   "conditions": {
    "model": "two copies of a 12-billion model, one per card",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21618,
    21691
   ],
   "quote": "| Two copies of a 12-billion model, one per card | 118.2 | 59.1 | 542 W |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21428,
     21520
    ],
    "header": "| Serving two callers on the pair, 300 W | Total tokens a second | Each caller | Mean draw |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-124-4",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "124.4",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22614,
    22655
   ],
   "quote": "| 1 | 124.4 | 124.0 | 126.7 | **134.8** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 1,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-124-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "124.0",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22614,
    22655
   ],
   "quote": "| 1 | 124.4 | 124.0 | 126.7 | **134.8** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-126-7",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "126.7",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22614,
    22655
   ],
   "quote": "| 1 | 124.4 | 124.0 | 126.7 | **134.8** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-134-8",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**134.8**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22614,
    22655
   ],
   "quote": "| 1 | 124.4 | 124.0 | 126.7 | **134.8** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-188-9",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "188.9",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22656,
    22697
   ],
   "quote": "| 2 | 188.9 | 189.5 | 195.5 | **210.4** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 1,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-189-5",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "189.5",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22656,
    22697
   ],
   "quote": "| 2 | 188.9 | 189.5 | 195.5 | **210.4** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-195-5",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "195.5",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22656,
    22697
   ],
   "quote": "| 2 | 188.9 | 189.5 | 195.5 | **210.4** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-210-4",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**210.4**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22656,
    22697
   ],
   "quote": "| 2 | 188.9 | 189.5 | 195.5 | **210.4** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-268-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**268.0**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22698,
    22743
   ],
   "quote": "| 4 | **268.0** | 267.6 | 265.2 | **288.4** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 1,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-267-6",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "267.6",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22698,
    22743
   ],
   "quote": "| 4 | **268.0** | 267.6 | 265.2 | **288.4** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-265-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "265.2",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22698,
    22743
   ],
   "quote": "| 4 | **268.0** | 267.6 | 265.2 | **288.4** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-summed-across-conversations-tokens-a-second-288-4",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**288.4**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22698,
    22743
   ],
   "quote": "| 4 | **268.0** | 267.6 | 265.2 | **288.4** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-per-conversation-tokens-a-second-67-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode per conversation, tokens a second",
   "value": "67.0",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22744,
    22801
   ],
   "quote": "| Per conversation at 4 | 67.0 | 66.9 | 66.3 | **72.1** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 1,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-per-conversation-tokens-a-second-66-9",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode per conversation, tokens a second",
   "value": "66.9",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22744,
    22801
   ],
   "quote": "| Per conversation at 4 | 67.0 | 66.9 | 66.3 | **72.1** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "Two 3080s, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-per-conversation-tokens-a-second-66-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode per conversation, tokens a second",
   "value": "66.3",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22744,
    22801
   ],
   "quote": "| Per conversation at 4 | 67.0 | 66.9 | 66.3 | **72.1** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-per-conversation-tokens-a-second-72-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode per conversation, tokens a second",
   "value": "**72.1**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    22744,
    22801
   ],
   "quote": "| Per conversation at 4 | 67.0 | 66.9 | 66.3 | **72.1** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "the-full-concurrency-ladder",
   "identity_span": [
    22488,
    22587
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22488,
     22587
    ],
    "header": "| Conversations at once | Two 3080s, 250 W | Two 3080s, 300 W | One 3090, 250 W | One 3090, 300 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/residency-under-the-server-s-own-placement-73-7",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "residency under the server's own placement",
   "value": "73.7 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "1 parallel slot",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23789,
    23870
   ],
   "quote": "| Two 3080s, 1 parallel slot | 73.7 % | **450 ms / 486 ms** | 782 ms / 1,098 ms |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23789,
    23870
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "Model on the cards"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/a-gate-shaped-call-median-p95-450-ms-486-ms",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "**450 ms / 486 ms**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "1 caller · 1 parallel slot",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23789,
    23870
   ],
   "quote": "| Two 3080s, 1 parallel slot | 73.7 % | **450 ms / 486 ms** | 782 ms / 1,098 ms |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23789,
    23870
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "1 caller, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/a-gate-shaped-call-median-p95-782-ms-1-098-ms",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "782 ms / 1,098 ms",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "4 callers · 1 parallel slot",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23789,
    23870
   ],
   "quote": "| Two 3080s, 1 parallel slot | 73.7 % | **450 ms / 486 ms** | 782 ms / 1,098 ms |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23789,
    23870
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "4 callers, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/residency-under-the-server-s-own-placement-41-7",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "residency under the server's own placement",
   "value": "41.7 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23871,
    23955
   ],
   "quote": "| Two 3080s, 4 parallel slots | 41.7 % | 709 ms / 712 ms | 1,076 ms / **7,599 ms** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23871,
    23955
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "Model on the cards"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/a-gate-shaped-call-median-p95-709-ms-712-ms",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "709 ms / 712 ms",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "1 caller · 4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23871,
    23955
   ],
   "quote": "| Two 3080s, 4 parallel slots | 41.7 % | 709 ms / 712 ms | 1,076 ms / **7,599 ms** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23871,
    23955
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "1 caller, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/a-gate-shaped-call-median-p95-1-076-ms-7-599-ms",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "1,076 ms / **7,599 ms**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "4 callers · 4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23871,
    23955
   ],
   "quote": "| Two 3080s, 4 parallel slots | 41.7 % | 709 ms / 712 ms | 1,076 ms / **7,599 ms** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23871,
    23955
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "4 callers, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/residency-under-the-server-s-own-placement-56-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "residency under the server's own placement",
   "value": "56.2 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23956,
    24039
   ],
   "quote": "| One 3090, 4 parallel slots | 56.2 % | 594 ms / 683 ms | 1,179 ms / **6,011 ms** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23956,
    24039
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "Model on the cards"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/a-gate-shaped-call-median-p95-594-ms-683-ms",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "594 ms / 683 ms",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "1 caller · 4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23956,
    24039
   ],
   "quote": "| One 3090, 4 parallel slots | 56.2 % | 594 ms / 683 ms | 1,179 ms / **6,011 ms** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23956,
    24039
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "1 caller, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/a-gate-shaped-call-median-p95-1-179-ms-6-011-ms",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, median / p95",
   "value": "1,179 ms / **6,011 ms**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "eight tokens out",
    "concurrency": "4 callers · 4 parallel slots",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-slot-trap-if-you-serve-other-people",
    "md_sha": "7cfcd585452024eeea4ad0e8d71be64af64a64ddc0257b8c7ff1047d994c4fb2"
   },
   "cite": "the-slot-trap-if-you-serve-other-people",
   "span": [
    23956,
    24039
   ],
   "quote": "| One 3090, 4 parallel slots | 56.2 % | 594 ms / 683 ms | 1,179 ms / **6,011 ms** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "the-slot-trap-if-you-serve-other-people",
   "identity_span": [
    23956,
    24039
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23674,
     23770
    ],
    "header": "| The server was given | Model on the cards | 1 caller, median / p95 | 4 callers, median / p95 |",
    "head_cell": "4 callers, median / p95"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-out-of-memory-holds-65-536-inst",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "**out of memory** — holds 65,536 instead, at 132.6 tok/s",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`, unquantised",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    24913,
    25024
   ],
   "quote": "| `f16`, unquantised | **out of memory** — holds 65,536 instead, at 132.6 tok/s | **133.6 tok/s**, 20,995 MiB |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-much-can-each-one-hold"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-133-6-tok-s-20-995-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "**133.6 tok/s**, 20,995 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`, unquantised",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    24913,
    25024
   ],
   "quote": "| `f16`, unquantised | **out of memory** — holds 65,536 instead, at 132.6 tok/s | **133.6 tok/s**, 20,995 MiB |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-much-can-each-one-hold"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-115-4-tok-s-19-531-mib",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "115.4 tok/s, 19,531 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q8_0`, the setting both rigs are measured at",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    25025,
    25126
   ],
   "quote": "| `q8_0`, the setting both rigs are measured at | 115.4 tok/s, 19,531 MiB | 125.6 tok/s, 20,129 MiB |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-125-6-tok-s-20-129-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "125.6 tok/s, 20,129 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q8_0`, the setting both rigs are measured at",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    25025,
    25126
   ],
   "quote": "| `q8_0`, the setting both rigs are measured at | 115.4 tok/s, 19,531 MiB | 125.6 tok/s, 20,129 MiB |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-113-8-tok-s-18-815-mib",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "113.8 tok/s, 18,815 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    25127,
    25189
   ],
   "quote": "| `q4_0` | 113.8 tok/s, 18,815 MiB | 124.7 tok/s, 19,415 MiB |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "Two 3080s (20,480 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-at-the-full-window-and-the-memory-it-used-124-7-tok-s-19-415-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "124.7 tok/s, 19,415 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    25127,
    25189
   ],
   "quote": "| `q4_0` | 113.8 tok/s, 18,815 MiB | 124.7 tok/s, 19,415 MiB |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "what-the-cache-type-costs",
   "identity_span": [
    24835,
    24898
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24835,
     24898
    ],
    "header": "| Cache type | Two 3080s (20,480 MiB) | One 3090 (24,576 MiB) |",
    "head_cell": "One 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x16-slot-2-074-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x16 slot",
   "value": "2.074 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26723,
    26797
   ],
   "quote": "| `klein-4b-8st` | 2.074 s | 2.639 s | **1.683 s** | 1.560 s | **1.23×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x16 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x4-slot-2-639-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x4 slot",
   "value": "2.639 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26723,
    26797
   ],
   "quote": "| `klein-4b-8st` | 2.074 s | 2.639 s | **1.683 s** | 1.560 s | **1.23×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x4 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-1-683-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "**1.683 s**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26723,
    26797
   ],
   "quote": "| `klein-4b-8st` | 2.074 s | 2.639 s | **1.683 s** | 1.560 s | **1.23×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-1-560-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "1.560 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26723,
    26797
   ],
   "quote": "| `klein-4b-8st` | 2.074 s | 2.639 s | **1.683 s** | 1.560 s | **1.23×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-ratio-the-single-card-against-one-card-of-the-pair-1-23",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own ratio, the single card against one card of the pair",
   "value": "**1.23×**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "matched at 250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26723,
    26797
   ],
   "quote": "| `klein-4b-8st` | 2.074 s | 2.639 s | **1.683 s** | 1.560 s | **1.23×** |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "Matched at 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x16-slot-2-067-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x16 slot",
   "value": "2.067 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st-pixel4`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26798,
    26879
   ],
   "quote": "| `klein-4b-8st-pixel4` | 2.067 s | 2.643 s | **1.721 s** | 1.579 s | **1.20×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x16 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x4-slot-2-643-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x4 slot",
   "value": "2.643 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st-pixel4`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26798,
    26879
   ],
   "quote": "| `klein-4b-8st-pixel4` | 2.067 s | 2.643 s | **1.721 s** | 1.579 s | **1.20×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x4 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-1-721-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "**1.721 s**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st-pixel4`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26798,
    26879
   ],
   "quote": "| `klein-4b-8st-pixel4` | 2.067 s | 2.643 s | **1.721 s** | 1.579 s | **1.20×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-1-579-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "1.579 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st-pixel4`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26798,
    26879
   ],
   "quote": "| `klein-4b-8st-pixel4` | 2.067 s | 2.643 s | **1.721 s** | 1.579 s | **1.20×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-ratio-the-single-card-against-one-card-of-the-pair-1-20",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own ratio, the single card against one card of the pair",
   "value": "**1.20×**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-8st-pixel4`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "matched at 250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26798,
    26879
   ],
   "quote": "| `klein-4b-8st-pixel4` | 2.067 s | 2.643 s | **1.721 s** | 1.579 s | **1.20×** |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "Matched at 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x16-slot-7-278-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x16 slot",
   "value": "7.278 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-32st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26880,
    26955
   ],
   "quote": "| `klein-4b-32st` | 7.278 s | 7.821 s | **6.297 s** | 5.827 s | **1.16×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x16 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-a-finished-image-median-one-card-in-the-x4-slot-7-821-s",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, one card in the x4 slot",
   "value": "7.821 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-32st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26880,
    26955
   ],
   "quote": "| `klein-4b-32st` | 7.278 s | 7.821 s | **6.297 s** | 5.827 s | **1.16×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3080, x4 slot, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-6-297-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "**6.297 s**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-32st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26880,
    26955
   ],
   "quote": "| `klein-4b-32st` | 7.278 s | 7.821 s | **6.297 s** | 5.827 s | **1.16×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-a-finished-image-median-the-single-card-5-827-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median, the single card",
   "value": "5.827 s",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-32st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "17 September 2026, 17:11 and 17:24 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26880,
    26955
   ],
   "quote": "| `klein-4b-32st` | 7.278 s | 7.821 s | **6.297 s** | 5.827 s | **1.16×** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26564,
    26692
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "One 3090, 300 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-ratio-the-single-card-against-one-card-of-the-pair-1-16",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own ratio, the single card against one card of the pair",
   "value": "**1.16×**",
   "non_figure": null,
   "conditions": {
    "model": "`klein-4b-32st`",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "matched at 250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    26880,
    26955
   ],
   "quote": "| `klein-4b-32st` | 7.278 s | 7.821 s | **6.297 s** | 5.827 s | **1.16×** |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26564,
     26692
    ],
    "header": "| Graph, batch 3 | One 3080, x16 slot, 250 W | One 3080, x4 slot, 250 W | One 3090, 250 W | One 3090, 300 W | Matched at 250 W |",
    "head_cell": "Matched at 250 W"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/images-drawn-in-ten-minutes-158",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "158",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28955,
    28987
   ],
   "quote": "| Images drawn | 158 | **269** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 1,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3080, x16 slot"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/images-drawn-in-ten-minutes-269",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "**269**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28955,
    28987
   ],
   "quote": "| Images drawn | 158 | **269** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/seconds-an-image-median-3-737",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, median",
   "value": "3.737",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28988,
    29036
   ],
   "quote": "| Seconds an image, median | 3.737 | **2.156** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 1,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3080, x16 slot"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/seconds-an-image-median-2-156",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, median",
   "value": "**2.156**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28988,
    29036
   ],
   "quote": "| Seconds an image, median | 3.737 | **2.156** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/core-temperature-plateau-77-c-at-5-6-min",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "77 °C at 5.6 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29037,
    29107
   ],
   "quote": "| Core temperature plateau | 77 °C at 5.6 min | **65 °C at 0.5 min** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 1,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3080, x16 slot"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temperature-plateau-65-c-at-0-5-min",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "**65 °C at 0.5 min**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29037,
    29107
   ],
   "quote": "| Core temperature plateau | 77 °C at 5.6 min | **65 °C at 0.5 min** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/core-temperature-peak-82-c",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "82 °C",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29108,
    29136
   ],
   "quote": "| Peak | 82 °C | **67 °C** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 1,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3080, x16 slot"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temperature-peak-67-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "**67 °C**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29108,
    29136
   ],
   "quote": "| Peak | 82 °C | **67 °C** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/fan-maximum-92",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan, maximum",
   "value": "92 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29137,
    29171
   ],
   "quote": "| Fan, maximum | 92 % | **58 %** |",
   "identity_basis": "column-head",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 1,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3080, x16 slot"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/fan-maximum-58",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan, maximum",
   "value": "**58 %**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    29137,
    29171
   ],
   "quote": "| Fan, maximum | 92 % | **58 %** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28903,
    28938
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28903,
     28938
    ],
    "header": "| | One 3080, x16 slot | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once",
    "how-to-check-our-work"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tok-s-114-8",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "114.8",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30827,
    30909
   ],
   "quote": "| Two 3080s, 250 W, mixture | 114.8 | 290 W | 2,531 | 46 · 42 °C | **0 % · 0 %** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30827,
    30909
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-290-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "290 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30827,
    30909
   ],
   "quote": "| Two 3080s, 250 W, mixture | 114.8 | 290 W | 2,531 | 46 · 42 °C | **0 % · 0 %** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30827,
    30909
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/joules-per-1-000-tokens-2-531",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "2,531",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30827,
    30909
   ],
   "quote": "| Two 3080s, 250 W, mixture | 114.8 | 290 W | 2,531 | 46 · 42 °C | **0 % · 0 %** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30827,
    30909
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/core-temp-max-46-42-c",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "46 · 42 °C",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30827,
    30909
   ],
   "quote": "| Two 3080s, 250 W, mixture | 114.8 | 290 W | 2,531 | 46 · 42 °C | **0 % · 0 %** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30827,
    30909
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/fan-max-0-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "fan max",
   "value": "**0 % · 0 %**",
   "non_figure": "0 %",
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30827,
    30909
   ],
   "quote": "| Two 3080s, 250 W, mixture | 114.8 | 290 W | 2,531 | 46 · 42 °C | **0 % · 0 %** |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30827,
    30909
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tok-s-115-4",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "115.4",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30910,
    30989
   ],
   "quote": "| Two 3080s, 300 W, mixture | 115.4 | 346 W | 2,994 | 59 · 58 °C | 64 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30910,
    30989
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-346-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "346 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30910,
    30989
   ],
   "quote": "| Two 3080s, 300 W, mixture | 115.4 | 346 W | 2,994 | 59 · 58 °C | 64 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30910,
    30989
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/joules-per-1-000-tokens-2-994",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "2,994",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30910,
    30989
   ],
   "quote": "| Two 3080s, 300 W, mixture | 115.4 | 346 W | 2,994 | 59 · 58 °C | 64 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30910,
    30989
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/core-temp-max-59-58-c",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "59 · 58 °C",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30910,
    30989
   ],
   "quote": "| Two 3080s, 300 W, mixture | 115.4 | 346 W | 2,994 | 59 · 58 °C | 64 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30910,
    30989
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/fan-max-64-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "64 % · 0 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30910,
    30989
   ],
   "quote": "| Two 3080s, 300 W, mixture | 115.4 | 346 W | 2,994 | 59 · 58 °C | 64 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30910,
    30989
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tok-s-44-6",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "44.6",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30990,
    31067
   ],
   "quote": "| Two 3080s, 250 W, dense | 44.6 | 466 W | 10,442 | 58 · 52 °C | 62 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30990,
    31067
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-466-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "466 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30990,
    31067
   ],
   "quote": "| Two 3080s, 250 W, dense | 44.6 | 466 W | 10,442 | 58 · 52 °C | 62 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30990,
    31067
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/joules-per-1-000-tokens-10-442",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "10,442",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30990,
    31067
   ],
   "quote": "| Two 3080s, 250 W, dense | 44.6 | 466 W | 10,442 | 58 · 52 °C | 62 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30990,
    31067
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/core-temp-max-58-52-c",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "58 · 52 °C",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30990,
    31067
   ],
   "quote": "| Two 3080s, 250 W, dense | 44.6 | 466 W | 10,442 | 58 · 52 °C | 62 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30990,
    31067
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/fan-max-62-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "62 % · 0 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    30990,
    31067
   ],
   "quote": "| Two 3080s, 250 W, dense | 44.6 | 466 W | 10,442 | 58 · 52 °C | 62 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    30990,
    31067
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/tok-s-44-7",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "44.7",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31068,
    31145
   ],
   "quote": "| Two 3080s, 300 W, dense | 44.7 | 509 W | 11,385 | 69 · 58 °C | 74 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31068,
    31145
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/mean-draw-509-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "509 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31068,
    31145
   ],
   "quote": "| Two 3080s, 300 W, dense | 44.7 | 509 W | 11,385 | 69 · 58 °C | 74 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31068,
    31145
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/joules-per-1-000-tokens-11-385",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "11,385",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31068,
    31145
   ],
   "quote": "| Two 3080s, 300 W, dense | 44.7 | 509 W | 11,385 | 69 · 58 °C | 74 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31068,
    31145
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/core-temp-max-69-58-c",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "69 · 58 °C",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31068,
    31145
   ],
   "quote": "| Two 3080s, 300 W, dense | 44.7 | 509 W | 11,385 | 69 · 58 °C | 74 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31068,
    31145
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/fan-max-74-0",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "74 % · 0 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31068,
    31145
   ],
   "quote": "| Two 3080s, 300 W, dense | 44.7 | 509 W | 11,385 | 69 · 58 °C | 74 % · 0 % |",
   "identity_basis": "row",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31068,
    31145
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "per card, the re-padded Gigabyte board first and the stock EVGA board second — the page's own column note",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/tok-s-125-6",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "125.6",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31146,
    31213
   ],
   "quote": "| One 3090, 250 W, mixture | 125.6 | 230 W | 1,832 | 63 °C | 42 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31146,
    31213
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/mean-draw-230-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "230 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31146,
    31213
   ],
   "quote": "| One 3090, 250 W, mixture | 125.6 | 230 W | 1,832 | 63 °C | 42 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31146,
    31213
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/joules-per-1-000-tokens-1-832",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,832",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31146,
    31213
   ],
   "quote": "| One 3090, 250 W, mixture | 125.6 | 230 W | 1,832 | 63 °C | 42 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31146,
    31213
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temp-max-63-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "63 °C",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31146,
    31213
   ],
   "quote": "| One 3090, 250 W, mixture | 125.6 | 230 W | 1,832 | 63 °C | 42 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31146,
    31213
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/fan-max-42",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "42 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31146,
    31213
   ],
   "quote": "| One 3090, 250 W, mixture | 125.6 | 230 W | 1,832 | 63 °C | 42 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31146,
    31213
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/tok-s-136-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "136.1",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31214,
    31284
   ],
   "quote": "| One 3090, 300 W, mixture | 136.1 | 242 W | 1,772 | 47 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31214,
    31284
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/mean-draw-242-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "242 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31214,
    31284
   ],
   "quote": "| One 3090, 300 W, mixture | 136.1 | 242 W | 1,772 | 47 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31214,
    31284
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/joules-per-1-000-tokens-1-772",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,772",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31214,
    31284
   ],
   "quote": "| One 3090, 300 W, mixture | 136.1 | 242 W | 1,772 | 47 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31214,
    31284
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temp-max-47-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "47 °C",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31214,
    31284
   ],
   "quote": "| One 3090, 300 W, mixture | 136.1 | 242 W | 1,772 | 47 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31214,
    31284
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/fan-max-0",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "fan max",
   "value": "**0 %**",
   "non_figure": "0 %",
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31214,
    31284
   ],
   "quote": "| One 3090, 300 W, mixture | 136.1 | 242 W | 1,772 | 47 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31214,
    31284
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/tok-s-33-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "33.3",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31285,
    31349
   ],
   "quote": "| One 3090, 250 W, dense | 33.3 | 235 W | 7,076 | 63 °C | 51 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31285,
    31349
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/mean-draw-235-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "235 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31285,
    31349
   ],
   "quote": "| One 3090, 250 W, dense | 33.3 | 235 W | 7,076 | 63 °C | 51 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31285,
    31349
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/joules-per-1-000-tokens-7-076",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "7,076",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31285,
    31349
   ],
   "quote": "| One 3090, 250 W, dense | 33.3 | 235 W | 7,076 | 63 °C | 51 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31285,
    31349
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temp-max-63-c-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "63 °C",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31285,
    31349
   ],
   "quote": "| One 3090, 250 W, dense | 33.3 | 235 W | 7,076 | 63 °C | 51 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31285,
    31349
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/fan-max-51",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "51 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31285,
    31349
   ],
   "quote": "| One 3090, 250 W, dense | 33.3 | 235 W | 7,076 | 63 °C | 51 % |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31285,
    31349
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/tok-s-47-8",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tok/s",
   "value": "47.8",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31350,
    31417
   ],
   "quote": "| One 3090, 300 W, dense | 47.8 | 290 W | 6,059 | 57 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31350,
    31417
   ],
   "cell": {
    "column": 1,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/mean-draw-290-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "290 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31350,
    31417
   ],
   "quote": "| One 3090, 300 W, dense | 47.8 | 290 W | 6,059 | 57 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31350,
    31417
   ],
   "cell": {
    "column": 2,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/joules-per-1-000-tokens-6-059",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,059",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31350,
    31417
   ],
   "quote": "| One 3090, 300 W, dense | 47.8 | 290 W | 6,059 | 57 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31350,
    31417
   ],
   "cell": {
    "column": 3,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Joules per 1,000 tokens"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-temp-max-57-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temp max",
   "value": "57 °C",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31350,
    31417
   ],
   "quote": "| One 3090, 300 W, dense | 47.8 | 290 W | 6,059 | 57 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31350,
    31417
   ],
   "cell": {
    "column": 4,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Core temp max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/fan-max-0-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "fan max",
   "value": "**0 %**",
   "non_figure": "0 %",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31350,
    31417
   ],
   "quote": "| One 3090, 300 W, dense | 47.8 | 290 W | 6,059 | 57 °C | **0 %** |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    31350,
    31417
   ],
   "cell": {
    "column": 5,
    "header_span": [
     30708,
     30797
    ],
    "header": "| Rig and model | tok/s | Mean draw | Joules per 1,000 tokens | Core temp max | Fan max |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/card-board-power-cards-empty-44-0-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "card board power, cards empty",
   "value": "44.0 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33070,
    33103
   ],
   "quote": "| Cards empty | 44.0 W | 25.4 W |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 1,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "Two 3080s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/card-board-power-cards-empty-25-4-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card board power, cards empty",
   "value": "25.4 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17 03:15 to 03:52 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33070,
    33103
   ],
   "quote": "| Cards empty | 44.0 W | 25.4 W |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 2,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/card-board-power-the-model-resident-85-7-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "card board power, the model resident",
   "value": "**85.7 W**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33104,
    33152
   ],
   "quote": "| The model resident | **85.7 W** | **53.1 W** |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 1,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "Two 3080s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/card-board-power-the-model-resident-53-1-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card board power, the model resident",
   "value": "**53.1 W**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17 03:15 to 03:52 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33104,
    33152
   ],
   "quote": "| The model resident | **85.7 W** | **53.1 W** |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 2,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/what-the-model-itself-costs-to-hold-41-7-w",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "what the model itself costs to hold",
   "value": "41.7 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33153,
    33210
   ],
   "quote": "| What the model itself costs to hold | 41.7 W | 27.7 W |",
   "identity_basis": "column-head",
   "identity_token": "two 3080s",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 1,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "Two 3080s"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/what-the-model-itself-costs-to-hold-27-7-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "what the model itself costs to hold",
   "value": "27.7 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17 03:15 to 03:52 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33153,
    33210
   ],
   "quote": "| What the model itself costs to hold | 41.7 W | 27.7 W |",
   "identity_basis": "column-head",
   "identity_token": "one 3090",
   "identity_cite": "power-heat-and-the-fifty-watts",
   "identity_span": [
    32998,
    33053
   ],
   "cell": {
    "column": 2,
    "header_span": [
     32998,
     33053
    ],
    "header": "| At 250 W, nothing generating | Two 3080s | One 3090 |",
    "head_cell": "One 3090"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "intel-core-i9-10850k/two-used-3080s-priced/price-the-processor-238-00",
   "identity": "intel-core-i9-10850k",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the processor",
   "value": "$238.00",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    35427,
    35535
   ],
   "quote": "| Processor | Newegg, \"Intel Core i9-10850K Comet Lake 10-Core (CM8070104608302)\", new, in stock | $238.00 |",
   "identity_basis": "row",
   "identity_token": "Intel Core i9-10850K Comet Lake 10-Core (CM8070104608302)",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    35427,
    35535
   ],
   "cell": {
    "column": 2,
    "header_span": [
     35250,
     35291
    ],
    "header": "| Line | The listing, as titled | Price |",
    "head_cell": "Price"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/price-one-card-the-day-s-ask-375-00-each",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, one card, the day's ask",
   "value": "$375.00 each",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    35536,
    35679
   ],
   "quote": "| Two used GeForce RTX 3080 10 GB, any make | Jawa, \"Dell RTX 3080 (10GB) OEM\", Used · Like New, **19 available**, $375.00 each | **$750.00** |",
   "identity_basis": "row",
   "identity_token": "two used GeForce RTX 3080 10 GB, any make",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    35536,
    35679
   ],
   "cell": {
    "column": 1,
    "header_span": [
     35250,
     35291
    ],
    "header": "| Line | The listing, as titled | Price |",
    "head_cell": "The listing, as titled"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/price-the-two-cards-750-00",
   "identity": "rtx-3080",
   "road": "page-table",
   "kind": "figure",
   "metric": "price, the two cards",
   "value": "**$750.00**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    35536,
    35679
   ],
   "quote": "| Two used GeForce RTX 3080 10 GB, any make | Jawa, \"Dell RTX 3080 (10GB) OEM\", Used · Like New, **19 available**, $375.00 each | **$750.00** |",
   "identity_basis": "row",
   "identity_token": "two used GeForce RTX 3080 10 GB, any make",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    35536,
    35679
   ],
   "cell": {
    "column": 2,
    "header_span": [
     35250,
     35291
    ],
    "header": "| Line | The listing, as titled | Price |",
    "head_cell": "Price"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "priced by MODEL and not by board, because that is what the bench measured and what the marketplace sells — the page's own reason",
   "record_published": true
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-build-total-1-578-80",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "the build, total",
   "value": "$1,578.80",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    36931,
    36979
   ],
   "quote": "| As built — 48 GB and the HX1000i | $1,578.80 |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     36898,
     36919
    ],
    "header": "| The build | Total |",
    "head_cell": "Total"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-build-total-1-483-80",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "the build, total",
   "value": "$1,483.80",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    36980,
    37017
   ],
   "quote": "| 48 GB and the RM1000e | $1,483.80 |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     36898,
     36919
    ],
    "header": "| The build | Total |",
    "head_cell": "Total"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-build-total-1-488-81",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "the build, total",
   "value": "$1,488.81",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    37018,
    37055
   ],
   "quote": "| 32 GB and the HX1000i | $1,488.81 |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     36898,
     36919
    ],
    "header": "| The build | Total |",
    "head_cell": "Total"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-build-total-1-393-81",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "the build, total",
   "value": "**$1,393.81**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    37056,
    37097
   ],
   "quote": "| 32 GB and the RM1000e | **$1,393.81** |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     36898,
     36919
    ],
    "header": "| The build | Total |",
    "head_cell": "Total"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cheapest combination the page totals",
   "record_published": true
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-build-total-1-080-83",
   "identity": "the-two-3080-rig",
   "road": "page-table",
   "kind": "figure",
   "metric": "the build, total",
   "value": "**$1,080.83**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    37098,
    37158
   ],
   "quote": "| Board, processor and the two cards alone | **$1,080.83** |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 1,
    "header_span": [
     36898,
     36919
    ],
    "header": "| The build | Total |",
    "head_cell": "Total"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "what a thousand dollars buys, and nothing else: no case, no drive, no operating system, and no shipping — the page's own list of what is still to add",
   "record_published": true
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-multiplier-out-of-the-box-against-the-pair-one-3090-2-17",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own multiplier, out of the box, against the pair",
   "value": "**one 3090, 2.17×**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "at-a-glance",
    "md_sha": "a8cd38f63739fb9eb3d1ba69009ce021722e84fa518934a6cb227d54fc58e936"
   },
   "cite": "at-a-glance",
   "span": [
    2488,
    2839
   ],
   "quote": "| **out of the box, before you set anything** — mixture at a 4,096 window | *not run at this cap* | 59.5 · 75.6 % on the cards | 129.4 · all of it on the cards | *not run at this cap* | **one 3090, 2.17×** — until the layer count is named, and then the pair doubles to 124.7. Each rig was measured at one cap only here, and they are not the same cap |",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "at-a-glance",
   "identity_span": [
    2488,
    2839
   ],
   "cell": {
    "column": 5,
    "header_span": [
     2361,
     2457
    ],
    "header": "| | Two 3080s, 250 W each | Two 3080s, 300 W each | One 3090, 250 W | One 3090, 300 W | Winner |",
    "head_cell": "Winner"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is `not stated` because the page states it is not one cap: the cell's own words are that each rig was measured at one cap only here, and they are not the same cap",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/decode-out-of-the-box-once-the-layer-count-is-named-124-7",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode out of the box once the layer count is named",
   "value": "124.7",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "at-a-glance",
    "md_sha": "a8cd38f63739fb9eb3d1ba69009ce021722e84fa518934a6cb227d54fc58e936"
   },
   "cite": "at-a-glance",
   "span": [
    2488,
    2839
   ],
   "quote": "| **out of the box, before you set anything** — mixture at a 4,096 window | *not run at this cap* | 59.5 · 75.6 % on the cards | 129.4 · all of it on the cards | *not run at this cap* | **one 3090, 2.17×** — until the layer count is named, and then the pair doubles to 124.7. Each rig was measured at one cap only here, and they are not the same cap |",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": {
    "column": 5,
    "header_span": [
     2361,
     2457
    ],
    "header": "| | Two 3080s, 250 W each | Two 3080s, 300 W each | One 3090, 250 W | One 3090, 300 W | Winner |",
    "head_cell": "Winner"
   },
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the pair's own doubling, printed in the winner cell and nowhere else on the page; the same cell says the two rigs' out-of-the-box arms did not run at the same cap",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-margin-over-the-pair-mixture-18-0-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the pair, mixture",
   "value": "18.0 per cent",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    12997,
    13025
   ],
   "quote": "the mixture by 18.0 per cent",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-margin-over-the-pair-dense-the-dense-model-by-7-0",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the pair, dense",
   "value": "the dense model by 7.0",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    13026,
    13052
   ],
   "quote": "and the dense model by 7.0",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own sentence gives the unit once for both halves: *the mixture by 18.0 per cent and the dense model by 7.0*",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-margin-over-the-pair-mixture-9-4-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the pair, mixture",
   "value": "9.4 per cent",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    13063,
    13105
   ],
   "quote": "it still takes the mixture by 9.4 per cent",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-exhibit-s-own-margin-over-the-single-card-dense-34-9-per-cent",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the single card, dense",
   "value": "**34.9 per cent**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "49,152",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    13112,
    13202
   ],
   "quote": "loses the dense model by **34.9 per cent**, judged at the same 49,152 window on both sides",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page states this as the single card's loss; it is the pair's one decode win and the page calls it its turning point",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/what-fifty-watts-are-worth-dense-43-8-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what fifty watts are worth, dense",
   "value": "43.8 per cent",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    13204,
    13265
   ],
   "quote": "Fifty watts are worth 43.8 per cent to one 3090 on that model",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "how-fast-does-each-one-write",
   "identity_span": [
    13204,
    13265
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/what-fifty-watts-are-worth-dense-0-2-per-cent",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what fifty watts are worth, dense",
   "value": "0.2 per cent",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-fast-does-each-one-write",
    "md_sha": "0b7fe0a4e729f0dd0ea01286bf81b3ec542b7e7feec61226b36260fd2d58ee5d"
   },
   "cite": "how-fast-does-each-one-write",
   "span": [
    13270,
    13294
   ],
   "quote": "0.2 per cent to the pair",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "power-heat-and-the-fifty-watts"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/memory-left-unused-under-the-server-s-own-placement-about-5-gb-of-the-pair-s-20-",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "memory left unused under the server's own placement",
   "value": "about 5 GB of the pair's 20 unused",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    17022,
    17076
   ],
   "quote": "left about 5 GB of the pair's 20 unused while doing it",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/decode-with-an-unquantised-cache-tokens-a-second-133-6",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode with an unquantised cache, tokens a second",
   "value": "**133.6**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "the lower cap",
    "context": "131,072-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    18123,
    18175
   ],
   "quote": "writes at **133.6** tokens a second at the lower cap",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "what-the-cache-type-costs"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/what-an-unquantised-cache-needs-at-the-full-window-20-995-mib-which-is-515-more-",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what an unquantised cache needs at the full window",
   "value": "**20,995 MiB, which is 515 more than two 10 GB cards have between them**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    18232,
    18304
   ],
   "quote": "**20,995 MiB, which is 515 more than two 10 GB cards have between them**",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "what-the-cache-type-costs"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/largest-window-an-unquantised-cache-leaves-the-pair-holds-65-536",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "largest window an unquantised cache leaves the pair",
   "value": "holds 65,536",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "how-much-can-each-one-hold",
    "md_sha": "7ece92bd5e9d6596f432733e5fe55c21bf4d11256f009aebda2086bd08f8180f"
   },
   "cite": "how-much-can-each-one-hold",
   "span": [
    18309,
    18388
   ],
   "quote": "on the pair that setting refuses to load at that window at all and holds 65,536",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "what-the-cache-type-costs"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-exhibit-s-own-ratio-on-the-read-1-59",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio on the read",
   "value": "1.59×",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "One 131,072-token window with about 98,000 tokens really in it",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-about-a-long-document",
    "md_sha": "f6e2aad38631542e411b117182efd03bbfb9c025f00d693d6ec64cbc5caf2817"
   },
   "cite": "what-about-a-long-document",
   "span": [
    19145,
    19219
   ],
   "quote": "This is the pair's one clear win, and it is a real one: 1.59× on the read.",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-exhibit-s-own-ratio-one-big-model-against-two-copies-60-3-per-cent",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, one big model against two copies",
   "value": "60.3 per cent",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "two callers",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21695,
    21726
   ],
   "quote": "The split wins by 60.3 per cent",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/residency-of-a-14-billion-model-on-one-card-92-1-per-cent-resident",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "residency of a 14-billion model on one card",
   "value": "92.1 per cent resident",
   "non_figure": null,
   "conditions": {
    "model": "a 14-billion model",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "Measured 2026-09-17 01:17 to 01:22 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "two-copies-or-one-bigger-model",
    "md_sha": "a924238f12bc4d349b03fb1d9a999f37091a27f9e247250ec2d6971702b899cc"
   },
   "cite": "two-copies-or-one-bigger-model",
   "span": [
    21820,
    21919
   ],
   "quote": "a 14-billion model does not fit a 10 GB card once its window is reserved, at 92.1 per cent resident",
   "identity_basis": "page",
   "identity_token": "two GeForce RTX 3080 10 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "one card of the pair, not the pair: the page's own sentence is about what a 10 GB card holds",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-margin-over-the-pair-one-caller-and-four-7-8-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the pair, one caller and four",
   "value": "7.8 per cent",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096 window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "the-full-concurrency-ladder",
    "md_sha": "744b0c10d6ecba1d8d80711d71476084dd10120b5ec82fcf661fd10452b21f1c"
   },
   "cite": "the-full-concurrency-ladder",
   "span": [
    23051,
    23128
   ],
   "quote": "the gap it opens at 300 W, 7.8 per cent, is the same gap it has at one caller",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/what-compressing-the-cache-to-8-bits-costs-about-6-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what compressing the cache to 8 bits costs",
   "value": "about **6 per cent**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q8_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "its full 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-the-cache-type-costs",
    "md_sha": "1b702c8056db2e41c06feeab0df8477c3c4b2db11337697ffad51ae13930801b"
   },
   "cite": "what-the-cache-type-costs",
   "span": [
    25394,
    25473
   ],
   "quote": "Compressing it to 8 bits costs the single card about **6 per cent** of its rate",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/core-clock-while-drawing-median-1-860-mhz-median",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core clock while drawing, median",
   "value": "1,860 MHz median",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    27833,
    27849
   ],
   "quote": "1,860 MHz median",
   "identity_basis": "section",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26583,
    26608
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/core-clock-while-drawing-median-1-575",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core clock while drawing, median",
   "value": "1,575",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    27850,
    27888
   ],
   "quote": "against 1,575 on the card that beat it",
   "identity_basis": "section",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26638,
    26653
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/one-picture-with-the-text-encoder-moved-to-the-second-card-from-3-462-s-to-2-261",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "one picture with the text encoder moved to the second card",
   "value": "from 3.462 s to 2.261 s — **1.53×**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    27972,
    28021
   ],
   "quote": "cut a picture from 3.462 s to 2.261 s — **1.53×**",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the same bytes and the same picture — the page's own words, and it says this is more than the whole generational gap above",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/two-pictures-at-once-images-a-minute-24-96-images-a-minute-against-16-37-for-one",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "two pictures at once, images a minute",
   "value": "**24.96 images a minute against 16.37 for one of them alone, 1.52×**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28369,
    28437
   ],
   "quote": "**24.96 images a minute against 16.37 for one of them alone, 1.52×**",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/images-a-minute-sustained-over-ten-minutes-26-87-a-minute",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "images a minute, sustained over ten minutes",
   "value": "**26.87 a minute**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28469,
    28564
   ],
   "quote": "the **26.87 a minute** one 3090 sustained over ten minutes on the same graph, at half the power",
   "identity_basis": "row",
   "identity_token": "one 3090",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    28469,
    28564
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/what-three-pictures-at-a-time-do-to-the-narrow-slot-from-1-99-times-the-x16-boar",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what three pictures at a time do to the narrow slot",
   "value": "from 1.99 times the x16 board's seconds down to 1.27",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "three pictures at a time",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "renders-one-3080-against-one-3090-seconds-per-image",
    "md_sha": "5e0f3c1285769cc7c7778964a933fb9b8d7fec17c3cfb00a460ed7f7477c13e9"
   },
   "cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "span": [
    28713,
    28765
   ],
   "quote": "from 1.99 times the x16 board's seconds down to 1.27",
   "identity_basis": "section",
   "identity_token": "one 3080",
   "identity_cite": "renders-one-3080-against-one-3090-seconds-per-image",
   "identity_span": [
    26611,
    26635
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words: on two submissions each, which is thin and is not leant on",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/what-fifty-watts-buy-on-the-pair-0-5-per-cent-on-the-mixture-and-0-2-on-the-dens",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what fifty watts buy on the pair",
   "value": "0.5 per cent on the mixture and 0.2 on the dense model",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them · 300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31461,
    31515
   ],
   "quote": "0.5 per cent on the mixture and 0.2 on the dense model",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": "the page's own heading for this reading is that on the pair, fifty watts buy nothing",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/what-fifty-watts-cost-the-pair-in-energy-18-and-9-per-cent-more-energy-per-thous",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what fifty watts cost the pair in energy",
   "value": "18 and 9 per cent more energy per thousand tokens",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them · 300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31517,
    31570
   ],
   "quote": "for 18 and 9 per cent more energy per thousand tokens",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/what-fifty-watts-are-worth-to-the-single-card-43-8-per-cent-on-the-dense-model-a",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what fifty watts are worth to the single card",
   "value": "43.8 per cent on the dense model and 8.4 on the mixture",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W · 300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31593,
    31681
   ],
   "quote": "the same fifty watts are worth 43.8 per cent** on the dense model and 8.4 on the mixture",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": {
    "slug": "two-used-3080s-priced",
    "cite": "how-fast-does-each-one-write"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/temperature-on-the-quietest-arm-with-neither-fan-turning-40-to-46-c",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "temperature on the quietest arm, with neither fan turning",
   "value": "40 to 46 °C",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W each — 500 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    31939,
    32015
   ],
   "quote": "**neither fan turned at all**, at 40 to 46 °C, writing 114.8 tokens a second",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page calls this the most surprising measurement on it: two five-year-old gaming cards writing in silence",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-energy-saving-against-the-pair-40-8-per-cent-less-per-thousand",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own energy saving against the pair",
   "value": "40.8 per cent less per thousand tokens on the mixture at 300 W and 46.8 per cent less on the dense model",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33964,
    34076
   ],
   "quote": "it uses 40.8 per cent less per thousand tokens on the mixture at 300 W and 46.8 per cent less on the dense model",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 24 GB",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    10163,
    10228
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/the-exhibit-s-own-energy-cost-against-the-single-card-the-pair-spends-69-0-and-8",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own energy cost against the single card",
   "value": "the pair spends 69.0 and 87.9 per cent more for the same thousand tokens",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W each — 600 W between them",
    "context": "each rig's headline window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    34106,
    34178
   ],
   "quote": "the pair spends 69.0 and 87.9 per cent more for the same thousand tokens",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/what-the-pair-costs-over-the-single-card-to-sit-there-holding-a-model-32-6-w-mor",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what the pair costs over the single card to sit there holding a model",
   "value": "**32.6 W more than the single card**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33212,
    33299
   ],
   "quote": "The pair costs **32.6 W more than the single card** to sit there holding the same model",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/left-on-continuously-a-day-and-a-year-0-78-kwh-a-day-and-286-kwh-a-year",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "left on continuously, a day and a year",
   "value": "**0.78 kWh a day and 286 kWh a year**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33330,
    33367
   ],
   "quote": "**0.78 kWh a day and 286 kWh a year**",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the difference between the two rigs held idle, not either rig's own consumption",
   "record_published": null
  },
  {
   "id": "the-two-3080-rig/two-used-3080s-priced/left-on-continuously-in-money-about-52-a-year-or-4-37-a-month",
   "identity": "the-two-3080-rig",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "left on continuously, in money",
   "value": "**about $52 a year, or $4.37 a month**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W each — 500 W between them",
    "context": "not stated",
    "batch_or_steps": "nothing generating",
    "concurrency": "not stated",
    "window": "2026-09-17, 01:43 to 01:50"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "power-heat-and-the-fifty-watts",
    "md_sha": "1b6d0128fdb49c436788ab4a18d862d2d1a4c63619907c92abb47f50c948a25b"
   },
   "cite": "power-heat-and-the-fifty-watts",
   "span": [
    33561,
    33599
   ],
   "quote": "**about $52 a year, or $4.37 a month**",
   "identity_basis": "page",
   "identity_token": "two 3080s",
   "identity_cite": "the-four-rigs-named-once",
   "identity_span": [
    9986,
    10082
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-how-the-control-was-run",
    "the-four-rigs-named-once"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "at 18.34 cents a kilowatt-hour, the United States residential average for June 2026 as the page names it — and it is the DIFFERENCE between the two rigs held idle, card board power only, so a wall meter would read more",
   "record_published": null
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/price-the-evga-board-measured-as-listed-that-day-450-00-plus-12-00-shipping",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, the EVGA board measured, as listed that day",
   "value": "$450.00 plus $12.00 shipping",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    36699,
    36761
   ],
   "quote": "the EVGA FTW3 Ultra asked $450.00 plus $12.00 shipping on Jawa",
   "identity_basis": "row",
   "identity_token": "the EVGA FTW3 Ultra",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    36699,
    36761
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this as a note beside the model price and says it is not the price of the rig",
   "record_published": true
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/price-the-gigabyte-board-measured-as-listed-that-day-620-00",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, the Gigabyte board measured, as listed that day",
   "value": "$620.00",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read between 15:00 and 15:30 UTC on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    36763,
    36836
   ],
   "quote": "and the Gigabyte $620.00 from a marketplace seller shipping from overseas",
   "identity_basis": "section",
   "identity_token": "two used GeForce RTX 3080 10 GB, any make",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    35538,
    35579
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cheaper Gigabyte listings that day being out of stock — the page's own qualification",
   "record_published": true
  },
  {
   "id": "rtx-3080/two-used-3080s-priced/completed-sales-over-the-twelve-months-average-and-count-358-00-each-across-318-",
   "identity": "rtx-3080",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "completed sales over the twelve months, average and count",
   "value": "$358.00 each across 318 sales",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "over the twelve months to 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    38752,
    38790
   ],
   "quote": "averaged $358.00 each across 318 sales",
   "identity_basis": "section",
   "identity_token": "two used GeForce RTX 3080 10 GB, any make",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    35538,
    35579
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page calls this a supporting figure rather than a price",
   "record_published": true
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/price-one-card-renewed-1-879-99",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, one card, renewed",
   "value": "**$1,879.99**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read 17 September 2026 in the same half hour"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    38127,
    38261
   ],
   "quote": "**$1,879.99**, \"EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)\"",
   "identity_basis": "row",
   "identity_token": "EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    38127,
    38261
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "read by hand at the store page, because that retailer serves no price to this workshop's fetcher; it replaces an earlier figure this workshop published on 13 September that an operator ruled inaccurate and struck",
   "record_published": true
  },
  {
   "id": "rtx-3090/two-used-3080s-priced/the-exhibit-s-own-ratio-one-card-against-the-two-2-51-times",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, one card against the two",
   "value": "**2.51 times**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read 17 September 2026 in the same half hour"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-used-3080s-priced",
    "section": "what-it-costs-priced-17-september-2026",
    "md_sha": "3c182ef9be2365bbd8d7335c5cbad3ef863f83a2675fac1ee6d9ba6e628fe467"
   },
   "cite": "what-it-costs-priced-17-september-2026",
   "span": [
    38608,
    38636
   ],
   "quote": "it is **2.51 times** as much",
   "identity_basis": "section",
   "identity_token": "EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)",
   "identity_cite": "what-it-costs-priced-17-september-2026",
   "identity_span": [
    38143,
    38260
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": true
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-board-s-own-default-limit-and-the-driver-s-maximum-350-w-the-driver-s-max-li",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the board's own default limit, and the driver's maximum",
   "value": "**350 W** (the driver's `max_limit` is 366 W)",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-cards-named-once",
    "md_sha": "a6315fdebe9410b0dbed118c5350825ba2f2e16a51b8a76a322844fb1b5756fd"
   },
   "cite": "the-two-cards-named-once",
   "span": [
    3999,
    4175
   ],
   "quote": "| **a 3090** | One EVGA GeForce RTX 3090 XC3 Ultra 24 GB, 24,576 MiB | **350 W** (the driver's `max_limit` is 366 W) | 250 W = **71 %** · 300 W = **86 %** · 350 W = **100 %** |",
   "identity_basis": "row",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    3999,
    4175
   ],
   "cell": {
    "column": 2,
    "header_span": [
     3880,
     3980
    ],
    "header": "| Label | The board | Its own default limit, read from the driver | Each cap as a fraction of that |",
    "head_cell": "Its own default limit, read from the driver"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/each-cap-as-a-fraction-of-the-board-s-own-limit-250-w-71-300-w-86-350-w-100",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "each cap as a fraction of the board's own limit",
   "value": "250 W = **71 %** · 300 W = **86 %** · 350 W = **100 %**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-cards-named-once",
    "md_sha": "a6315fdebe9410b0dbed118c5350825ba2f2e16a51b8a76a322844fb1b5756fd"
   },
   "cite": "the-two-cards-named-once",
   "span": [
    3999,
    4175
   ],
   "quote": "| **a 3090** | One EVGA GeForce RTX 3090 XC3 Ultra 24 GB, 24,576 MiB | **350 W** (the driver's `max_limit` is 366 W) | 250 W = **71 %** · 300 W = **86 %** · 350 W = **100 %** |",
   "identity_basis": "row",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    3999,
    4175
   ],
   "cell": {
    "column": 3,
    "header_span": [
     3880,
     3980
    ],
    "header": "| Label | The board | Its own default limit, read from the driver | Each cap as a fraction of that |",
    "head_cell": "Each cap as a fraction of that"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-board-s-own-default-limit-and-the-driver-s-maximum-450-w-max-limit-480-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the board's own default limit, and the driver's maximum",
   "value": "**450 W** (`max_limit` 480 W)",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-cards-named-once",
    "md_sha": "a6315fdebe9410b0dbed118c5350825ba2f2e16a51b8a76a322844fb1b5756fd"
   },
   "cite": "the-two-cards-named-once",
   "span": [
    4176,
    4362
   ],
   "quote": "| **a 3090 Ti** | One EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, 24,564 MiB | **450 W** (`max_limit` 480 W) | 300 W = **67 %** · 350 W = **78 %** · 400 W = **89 %** · 450 W = **100 %** |",
   "identity_basis": "row",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4176,
    4362
   ],
   "cell": {
    "column": 2,
    "header_span": [
     3880,
     3980
    ],
    "header": "| Label | The board | Its own default limit, read from the driver | Each cap as a fraction of that |",
    "head_cell": "Its own default limit, read from the driver"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/each-cap-as-a-fraction-of-the-board-s-own-limit-300-w-67-350-w-78-400-w-89-450-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "each cap as a fraction of the board's own limit",
   "value": "300 W = **67 %** · 350 W = **78 %** · 400 W = **89 %** · 450 W = **100 %**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-cards-named-once",
    "md_sha": "a6315fdebe9410b0dbed118c5350825ba2f2e16a51b8a76a322844fb1b5756fd"
   },
   "cite": "the-two-cards-named-once",
   "span": [
    4176,
    4362
   ],
   "quote": "| **a 3090 Ti** | One EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB, 24,564 MiB | **450 W** (`max_limit` 480 W) | 300 W = **67 %** · 350 W = **78 %** · 400 W = **89 %** · 450 W = **100 %** |",
   "identity_basis": "row",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4176,
    4362
   ],
   "cell": {
    "column": 3,
    "header_span": [
     3880,
     3980
    ],
    "header": "| Label | The board | Its own default limit, read from the driver | Each cap as a fraction of that |",
    "head_cell": "Each cap as a fraction of that"
   },
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-136-139",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "136.139",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10515,
    10602
   ],
   "quote": "| Mixture, tok/s at 131,072 | 136.139 | **145.908** | 136.618 | **147.823** | 147.560 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-145-908",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**145.908**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10515,
    10602
   ],
   "quote": "| Mixture, tok/s at 131,072 | 136.139 | **145.908** | 136.618 | **147.823** | 147.560 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-136-618",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "136.618",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10515,
    10602
   ],
   "quote": "| Mixture, tok/s at 131,072 | 136.139 | **145.908** | 136.618 | **147.823** | 147.560 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-fits-and-what-the-cache-costs"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-147-823",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**147.823**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10515,
    10602
   ],
   "quote": "| Mixture, tok/s at 131,072 | 136.139 | **145.908** | 136.618 | **147.823** | 147.560 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-147-560",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "147.560",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10515,
    10602
   ],
   "quote": "| Mixture, tok/s at 131,072 | 136.139 | **145.908** | 136.618 | **147.823** | 147.560 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-47-838",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**47.838**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10603,
    10686
   ],
   "quote": "| Dense, tok/s at 65,536 | **47.838** | 42.969 | 46.880 | **53.354** | **57.396** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-42-969",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "42.969",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10603,
    10686
   ],
   "quote": "| Dense, tok/s at 65,536 | **47.838** | 42.969 | 46.880 | **53.354** | **57.396** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-46-880",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "46.880",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10603,
    10686
   ],
   "quote": "| Dense, tok/s at 65,536 | **47.838** | 42.969 | 46.880 | **53.354** | **57.396** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-53-354",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**53.354**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10603,
    10686
   ],
   "quote": "| Dense, tok/s at 65,536 | **47.838** | 42.969 | 46.880 | **53.354** | **57.396** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-57-396",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**57.396**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "65,536",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10603,
    10686
   ],
   "quote": "| Dense, tok/s at 65,536 | **47.838** | 42.969 | 46.880 | **53.354** | **57.396** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-241-53-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "241.53 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10687,
    10764
   ],
   "quote": "| Mean draw, mixture | 241.53 W | 240.58 W | 240.49 W | 282.64 W | 283.01 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-240-58-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "240.58 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10687,
    10764
   ],
   "quote": "| Mean draw, mixture | 241.53 W | 240.58 W | 240.49 W | 282.64 W | 283.01 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-240-49-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "240.49 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10687,
    10764
   ],
   "quote": "| Mean draw, mixture | 241.53 W | 240.58 W | 240.49 W | 282.64 W | 283.01 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-282-64-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "282.64 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10687,
    10764
   ],
   "quote": "| Mean draw, mixture | 241.53 W | 240.58 W | 240.49 W | 282.64 W | 283.01 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-283-01-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "283.01 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10687,
    10764
   ],
   "quote": "| Mean draw, mixture | 241.53 W | 240.58 W | 240.49 W | 282.64 W | 283.01 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-289-86-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "289.86 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10765,
    10840
   ],
   "quote": "| Mean draw, dense | 289.86 W | 296.64 W | 273.35 W | 338.05 W | 418.76 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-296-64-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "296.64 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10765,
    10840
   ],
   "quote": "| Mean draw, dense | 289.86 W | 296.64 W | 273.35 W | 338.05 W | 418.76 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-273-35-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "273.35 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10765,
    10840
   ],
   "quote": "| Mean draw, dense | 289.86 W | 296.64 W | 273.35 W | 338.05 W | 418.76 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-338-05-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "338.05 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10765,
    10840
   ],
   "quote": "| Mean draw, dense | 289.86 W | 296.64 W | 273.35 W | 338.05 W | 418.76 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-418-76-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "418.76 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10765,
    10840
   ],
   "quote": "| Mean draw, dense | 289.86 W | 296.64 W | 273.35 W | 338.05 W | 418.76 W |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-96-6",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "96.6 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10841,
    10927
   ],
   "quote": "| Draw as a share of the cap, dense | 96.6 % | 98.9 % | **78.1 %** | 96.6 % | 93.1 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-98-9",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "98.9 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10841,
    10927
   ],
   "quote": "| Draw as a share of the cap, dense | 96.6 % | 98.9 % | **78.1 %** | 96.6 % | 93.1 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-78-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "**78.1 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10841,
    10927
   ],
   "quote": "| Draw as a share of the cap, dense | 96.6 % | 98.9 % | **78.1 %** | 96.6 % | 93.1 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-96-6",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "96.6 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10841,
    10927
   ],
   "quote": "| Draw as a share of the cap, dense | 96.6 % | 98.9 % | **78.1 %** | 96.6 % | 93.1 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-93-1",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "93.1 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10841,
    10927
   ],
   "quote": "| Draw as a share of the cap, dense | 96.6 % | 98.9 % | **78.1 %** | 96.6 % | 93.1 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100-of-samples",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 % of samples",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10928,
    11052
   ],
   "quote": "| The cap flag, dense | Active in 100 % of samples | Active in 100 % | Active in 100 % | Active in 100 % | Active in 100 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10928,
    11052
   ],
   "quote": "| The cap flag, dense | Active in 100 % of samples | Active in 100 % | Active in 100 % | Active in 100 % | Active in 100 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10928,
    11052
   ],
   "quote": "| The cap flag, dense | Active in 100 % of samples | Active in 100 % | Active in 100 % | Active in 100 % | Active in 100 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10928,
    11052
   ],
   "quote": "| The cap flag, dense | Active in 100 % of samples | Active in 100 % | Active in 100 % | Active in 100 % | Active in 100 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    10928,
    11052
   ],
   "quote": "| The cap flag, dense | Active in 100 % of samples | Active in 100 % | Active in 100 % | Active in 100 % | Active in 100 % |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-100-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 100 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11053,
    11154
   ],
   "quote": "| The cap flag, mixture | Active in 100 % | Active in 83.3 % | Active in 83.3 % | **0 %** | **0 %** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-83-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 83.3 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11053,
    11154
   ],
   "quote": "| The cap flag, mixture | Active in 100 % | Active in 83.3 % | Active in 83.3 % | **0 %** | **0 %** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-in-83-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active in 83.3 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11053,
    11154
   ],
   "quote": "| The cap flag, mixture | Active in 100 % | Active in 83.3 % | Active in 83.3 % | **0 %** | **0 %** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-0",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the cap flag, the driver's own report",
   "value": "**0 %**",
   "non_figure": "0 %",
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11053,
    11154
   ],
   "quote": "| The cap flag, mixture | Active in 100 % | Active in 83.3 % | Active in 83.3 % | **0 %** | **0 %** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-0-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the cap flag, the driver's own report",
   "value": "**0 %**",
   "non_figure": "0 %",
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11053,
    11154
   ],
   "quote": "| The cap flag, mixture | Active in 100 % | Active in 83.3 % | Active in 83.3 % | **0 %** | **0 %** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-clock-median-1-275-mhz",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core clock, median",
   "value": "1,275 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11155,
    11245
   ],
   "quote": "| Core clock median, dense | 1,275 MHz | **960 MHz** | 1,215 MHz | 1,305 MHz | 1,965 MHz |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-clock-median-960-mhz",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core clock, median",
   "value": "**960 MHz**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11155,
    11245
   ],
   "quote": "| Core clock median, dense | 1,275 MHz | **960 MHz** | 1,215 MHz | 1,305 MHz | 1,965 MHz |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-clock-median-1-215-mhz",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core clock, median",
   "value": "1,215 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11155,
    11245
   ],
   "quote": "| Core clock median, dense | 1,275 MHz | **960 MHz** | 1,215 MHz | 1,305 MHz | 1,965 MHz |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-clock-median-1-305-mhz",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core clock, median",
   "value": "1,305 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11155,
    11245
   ],
   "quote": "| Core clock median, dense | 1,275 MHz | **960 MHz** | 1,215 MHz | 1,305 MHz | 1,965 MHz |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-clock-median-1-965-mhz",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core clock, median",
   "value": "1,965 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11155,
    11245
   ],
   "quote": "| Core clock median, dense | 1,275 MHz | **960 MHz** | 1,215 MHz | 1,305 MHz | 1,965 MHz |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-059-20",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,059.20",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11246,
    11340
   ],
   "quote": "| ↓ energy, dense, J/1,000 tokens | 6,059.20 | 6,902.38 | **5,842.92** | 6,343.81 | 7,295.49 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-902-38",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,902.38",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11246,
    11340
   ],
   "quote": "| ↓ energy, dense, J/1,000 tokens | 6,059.20 | 6,902.38 | **5,842.92** | 6,343.81 | 7,295.49 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/joules-per-1-000-tokens-5-842-92",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "**5,842.92**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11246,
    11340
   ],
   "quote": "| ↓ energy, dense, J/1,000 tokens | 6,059.20 | 6,902.38 | **5,842.92** | 6,343.81 | 7,295.49 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-343-81",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,343.81",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11246,
    11340
   ],
   "quote": "| ↓ energy, dense, J/1,000 tokens | 6,059.20 | 6,902.38 | **5,842.92** | 6,343.81 | 7,295.49 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-7-295-49",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "7,295.49",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11246,
    11340
   ],
   "quote": "| ↓ energy, dense, J/1,000 tokens | 6,059.20 | 6,902.38 | **5,842.92** | 6,343.81 | 7,295.49 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-771-71",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,771.71",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11341,
    11441
   ],
   "quote": "| ↓ energy, mixture, J/1,000 tokens | 1,771.71 | **1,648.85** | **1,758.13** | 1,909.96 | 1,915.13 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 1,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 300 W *(86 % of its limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 300 W *(86 % of its limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-648-85",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "**1,648.85**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11341,
    11441
   ],
   "quote": "| ↓ energy, mixture, J/1,000 tokens | 1,771.71 | **1,648.85** | **1,758.13** | 1,909.96 | 1,915.13 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 2,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 300 W *(67 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-758-13",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "**1,758.13**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-18, 00:22 to 00:54 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11341,
    11441
   ],
   "quote": "| ↓ energy, mixture, J/1,000 tokens | 1,771.71 | **1,648.85** | **1,758.13** | 1,909.96 | 1,915.13 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 3,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090, 350 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090, 350 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-909-96",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,909.96",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11341,
    11441
   ],
   "quote": "| ↓ energy, mixture, J/1,000 tokens | 1,771.71 | **1,648.85** | **1,758.13** | 1,909.96 | 1,915.13 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 4,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 350 W *(78 %)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-915-13",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,915.13",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "2026-09-17 (14:54 to 23:46 UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    11341,
    11441
   ],
   "quote": "| ↓ energy, mixture, J/1,000 tokens | 1,771.71 | **1,648.85** | **1,758.13** | 1,909.96 | 1,915.13 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    10315,
    10483
   ],
   "cell": {
    "column": 5,
    "header_span": [
     10315,
     10483
    ],
    "header": "| | A 3090, 300 W *(86 % of its limit)* | A 3090 Ti, 300 W *(67 %)* | A 3090, 350 W *(its own limit)* | A 3090 Ti, 350 W *(78 %)* | A 3090 Ti, 450 W *(its own limit)* |",
    "head_cell": "A 3090 Ti, 450 W *(its own limit)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads `a 3090 Ti, 450 W *(its own limit)*`",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-235-38-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "235.38 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-94-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "94.2 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/memory-clock-held-under-load-9-501-mhz",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "9,501 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-912-1-gb-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "912.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-504-9-gb-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "504.9 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-55-4",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**55.4 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-33-265",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "33.265",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    13985,
    14118
   ],
   "quote": "| A 3090, 250 W | 71 % | Ladder rung | 235.38 W | 94.2 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 504.9 GB/s | **55.4 %** | 33.265 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    13985,
    14118
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the cap table at the top of this section does not carry this rung: it prints a 3090 at 300 and 350 W and the Ti at 300, 350 and 450, and this reading is printed nowhere else on the page",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-289-86-w-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "289.86 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-96-6-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "96.6 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/memory-clock-held-under-load-9-501-mhz-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "9,501 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-912-1-gb-s-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "912.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-726-1-gb-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "726.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-79-6",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**79.6 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-47-838-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "47.838",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "single-window repeat",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14119,
    14265
   ],
   "quote": "| A 3090, 300 W | 86 % | **single-window repeat** | 289.86 W | 96.6 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 726.1 GB/s | **79.6 %** | 47.838 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14119,
    14265
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/mean-draw-273-35-w-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "273.35 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-78-1-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "78.1 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/memory-clock-held-under-load-9-501-mhz-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "9,501 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-912-1-gb-s-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "912.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-711-5-gb-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "711.5 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-78-0",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**78.0 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-tokens-a-second-46-880-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "46.880",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14266,
    14400
   ],
   "quote": "| A 3090, 350 W | 100 % | Ladder rung | 273.35 W | 78.1 % | Active 100 % | 9,501 MHz | 912.1 GB/s | 711.5 GB/s | **78.0 %** | 46.880 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14266,
    14400
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-296-64-w-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "296.64 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-98-9-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "98.9 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/memory-clock-held-under-load-10-251-mhz",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "10,251 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-984-1-gb-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "984.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-652-2-gb-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "652.2 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-66-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**66.3 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-42-969-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "42.969",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14401,
    14538
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | Ladder rung | 296.64 W | 98.9 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 652.2 GB/s | **66.3 %** | 42.969 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14401,
    14538
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-338-05-w-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "338.05 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-96-6-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "96.6 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/memory-clock-held-under-load-10-251-mhz-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "10,251 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-984-1-gb-s-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "984.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-809-8-gb-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "809.8 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-82-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**82.3 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-53-354-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "53.354",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14539,
    14676
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | Ladder rung | 338.05 W | 96.6 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 809.8 GB/s | **82.3 %** | 53.354 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14539,
    14676
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-382-16-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "382.16 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-95-5",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "95.5 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/memory-clock-held-under-load-10-251-mhz-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "10,251 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-984-1-gb-s-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "984.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-864-1-gb-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "864.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-87-8",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**87.8 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-56-936",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "56.936",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14677,
    14814
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | Ladder rung | 382.16 W | 95.5 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 864.1 GB/s | **87.8 %** | 56.936 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14677,
    14814
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "the-two-models-want-different-caps"
   },
   "later_reading_of": null,
   "note": "the cap table at the top of this section does not carry this rung: it prints a 3090 at 300 and 350 W and the Ti at 300, 350 and 450, and this reading is printed again in §The two models want different caps",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/mean-draw-418-76-w-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "mean draw",
   "value": "418.76 W",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 3,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Mean draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/draw-as-a-share-of-the-cap-93-1-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "draw as a share of the cap",
   "value": "93.1 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 4,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "% of the cap drawn"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-driver-s-own-report-active-100-4",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the driver's own report",
   "value": "Active 100 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 5,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/memory-clock-held-under-load-10-251-mhz-4",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "memory clock held under load",
   "value": "10,251 MHz",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 6,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Memory clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-bandwidth-ceiling-at-the-memory-clock-observed-984-1-gb-s-4",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the bandwidth ceiling at the memory clock observed",
   "value": "984.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 7,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Ceiling at that clock"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-bandwidth-while-decoding-871-1-gb-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed bandwidth while decoding",
   "value": "871.1 GB/s",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 8,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Streamed"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/streamed-as-a-share-of-the-ceiling-88-5",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "streamed as a share of the ceiling",
   "value": "**88.5 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 9,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "**% of the ceiling**"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-57-396-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "57.396",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ladder rung",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    14815,
    14953
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | Ladder rung | 418.76 W | 93.1 % | Active 100 % | 10,251 MHz | 984.1 GB/s | 871.1 GB/s | **88.5 %** | 57.396 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "identity_span": [
    14815,
    14953
   ],
   "cell": {
    "column": 10,
    "header_span": [
     13749,
     13929
    ],
    "header": "| Card and cap | % of its own limit | Arm shape | Mean draw | % of the cap drawn | Cap flag | Memory clock | Ceiling at that clock | Streamed | **% of the ceiling** | Dense tok/s |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-145-908-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "145.908",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "—",
   "non_figure": "—",
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-648-85-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "**1,648.85**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-42-969-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "42.969",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "—",
   "non_figure": "—",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-902-38-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,902.38",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18695,
    18764
   ],
   "quote": "| 300 W | 67 % | 145.908 | — | **1,648.85** | 42.969 | — | 6,902.38 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-147-823-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**147.823**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-1-31",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "+1.31 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-909-96-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,909.96 ⚑",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page marks this cell ⚑ and explains it below the table: between 350 and 400 W the mixture drew 28.4 W less at identical core clocks and identical decode, because the fan read 49 per cent at one rung and 0 at the other. The bench flags the step and corrects nothing",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-53-354-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "53.354",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-24-17",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "**+24.17 %**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-343-81-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "**6,343.81**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18765,
    18857
   ],
   "quote": "| 350 W | 78 % | **147.823** | +1.31 % | 1,909.96 ⚑ | 53.354 | **+24.17 %** | **6,343.81** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-147-599",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "147.599",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-0-15",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "−0.15 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-723-81",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,723.81 ⚑",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page marks this cell ⚑ and explains it below the table: between 350 and 400 W the mixture drew 28.4 W less at identical core clocks and identical decode, because the fan read 49 per cent at one rung and 0 at the other. The bench flags the step and corrects nothing",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-56-936-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "56.936",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-6-71",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "+6.71 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-6-713-05",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "6,713.05",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18858,
    18937
   ],
   "quote": "| 400 W | 89 % | 147.599 | −0.15 % | 1,723.81 ⚑ | 56.936 | +6.71 % | 6,713.05 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-147-560-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "147.560",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-0-03",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "−0.03 %",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-1-915-13-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "1,915.13",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Mixture J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-tokens-a-second-57-396-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second",
   "value": "**57.396**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense tok/s"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-step-against-the-rung-below-it-in-watts-0-81",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the exhibit's own step against the rung below it in watts",
   "value": "+0.81 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Step"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/joules-per-1-000-tokens-7-295-49-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "7,295.49",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-two-models-want-different-caps",
    "md_sha": "ef526254df7764d5c98d110062e1dc67b70bcc8657d45e9f798e03fd4977c0ac"
   },
   "cite": "the-two-models-want-different-caps",
   "span": [
    18938,
    19020
   ],
   "quote": "| 450 W | 100 % | 147.560 | −0.03 % | 1,915.13 | **57.396** | +0.81 % | 7,295.49 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     18547,
     18653
    ],
    "header": "| Cap | % of its own limit | Mixture tok/s | Step | Mixture J/1,000 | Dense tok/s | Step | Dense J/1,000 |",
    "head_cell": "Dense J/1,000"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": "the page's own words for this table's context: *both models at their headline windows*. It does not print the two window sizes here, so this row does not either",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-455-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.455 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-397-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.397 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-363-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "**1.363 s**",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-369-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.369 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-683-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.683 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 5,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-560-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.560 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 6,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-509-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.509 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20408,
    20500
   ],
   "quote": "| `klein-4b-8st` | 1.455 s | 1.397 s | **1.363 s** | 1.369 s | 1.683 s | 1.560 s | 1.509 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 7,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-436-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.436 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-379-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.379 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-351-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.351 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-350-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "**1.350 s**",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-721-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.721 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 5,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-579-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.579 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 6,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-1-528-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "1.528 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-8st-pixel4",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20501,
    20600
   ],
   "quote": "| `klein-4b-8st-pixel4` | 1.436 s | 1.379 s | 1.351 s | **1.350 s** | 1.721 s | 1.579 s | 1.528 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 7,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-332-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.332 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 1,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-130-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.130 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 2,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-015-s",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.015 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 3,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-015-s-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.015 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 4,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-6-297-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "6.297 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    20208,
    20366
   ],
   "cell": {
    "column": 5,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-827-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.827 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 6,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-a-finished-image-median-of-two-timed-submits-5-670-s",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a finished image, median of two timed submits",
   "value": "5.670 s",
   "non_figure": null,
   "conditions": {
    "model": "klein-4b-32st",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20601,
    20690
   ],
   "quote": "| `klein-4b-32st` | 5.332 s | 5.130 s | 5.015 s | 5.015 s | 6.297 s | 5.827 s | 5.670 s |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": {
    "column": 7,
    "header_span": [
     20208,
     20366
    ],
    "header": "| Graph, batch 3 | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* | A 3090, 250 W *(71 %)* | 300 W *(86 %)* | 350 W *(100 %)* |",
    "head_cell": "350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-320",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "320",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-1-811",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "1.811",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-plateau-56-c-at-0-7-min",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "56 °C at 0.7 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-peak-58-c",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "58 °C",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/fan-max-72",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "72 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/median-draw-292-0-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "292.0 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-92-9",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "92.9 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22070,
    22164
   ],
   "quote": "| A 3090 Ti, 300 W | 67 % | 320 | 1.811 | 56 °C at 0.7 min | 58 °C | 72 % | 292.0 W | 92.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22070,
    22164
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-331",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "331",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-1-715",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "1.715",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-plateau-55-c-at-0-4-min",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "55 °C at 0.4 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-peak-60-c",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "60 °C",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/fan-max-73",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "73 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/median-draw-316-2-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "316.2 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-87-9",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "87.9 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22165,
    22259
   ],
   "quote": "| A 3090 Ti, 350 W | 78 % | 331 | 1.715 | 55 °C at 0.4 min | 60 °C | 73 % | 316.2 W | 87.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22165,
    22259
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-347",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "347",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-1-676",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "1.676",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-plateau-59-c-at-0-3-min",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "59 °C at 0.3 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-peak-61-c",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "61 °C",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/fan-max-75",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "75 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/median-draw-355-4-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "355.4 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-16-0",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "16.0 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22260,
    22354
   ],
   "quote": "| A 3090 Ti, 400 W | 89 % | 347 | 1.676 | 59 °C at 0.3 min | 61 °C | 75 % | 355.4 W | 16.0 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22260,
    22354
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-348",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "**348**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-1-672",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "**1.672**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-plateau-57-c-at-0-3-min",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "57 °C at 0.3 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/core-temperature-peak-62-c",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "**62 °C**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/fan-max-75-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "75 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/median-draw-357-1-w",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "357.1 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-14-5",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "14.5 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22355,
    22462
   ],
   "quote": "| A 3090 Ti, 450 W | 100 % | **348** | **1.672** | 57 °C at 0.3 min | **62 °C** | 75 % | 357.1 W | 14.5 % |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "renders",
   "identity_span": [
    22355,
    22462
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-269",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "269",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-2-156",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "2.156",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-temperature-plateau-65-c-at-0-5-min",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "65 °C at 0.5 min",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-temperature-peak-67-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "67 °C",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/fan-max-58",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "58 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/median-draw-244-3-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "244.3 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-99-9",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "99.9 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22463,
    22554
   ],
   "quote": "| A 3090, 250 W | 71 % | 269 | 2.156 | 65 °C at 0.5 min | 67 °C | 58 % | 244.3 W | 99.9 % |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22463,
    22554
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/images-drawn-in-ten-minutes-311",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "images drawn in ten minutes",
   "value": "311",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 2,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Images"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/seconds-an-image-over-ten-minutes-of-continuous-drawing-1-886",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds an image, over ten minutes of continuous drawing",
   "value": "1.886",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 3,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "s/image"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-temperature-plateau-71-c-at-0-9-min",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "**71 °C at 0.9 min**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 4,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Plateau"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/core-temperature-peak-74-c",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature peak",
   "value": "**74 °C**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 5,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Peak"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/fan-max-67",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "fan max",
   "value": "67 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 6,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Fan max"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/median-draw-325-8-w",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "median draw",
   "value": "325.8 W",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 7,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Median draw"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-cap-flag-the-share-of-samples-it-was-set-in-100",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the cap flag, the share of samples it was set in",
   "value": "**100 %**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    22555,
    22658
   ],
   "quote": "| A 3090, 350 W | 100 % | 311 | 1.886 | **71 °C at 0.9 min** | **74 °C** | 67 % | 325.8 W | **100 %** |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "renders",
   "identity_span": [
    22555,
    22658
   ],
   "cell": {
    "column": 8,
    "header_span": [
     21916,
     22024
    ],
    "header": "| Card and cap | % of its own limit | Images | s/image | Plateau | Peak | Fan max | Median draw | Cap flag |",
    "head_cell": "Cap flag"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/largest-window-held-whole-placement-left-to-the-server-131-072-whole-20-129-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, placement left to the server",
   "value": "**131,072** whole, 20,129 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    24914,
    25019
   ],
   "quote": "| Mixture, placement left to the server | **131,072** whole, 20,129 MiB | **131,072** whole, 20,142 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/largest-window-held-whole-placement-left-to-the-server-131-072-whole-20-142-mib",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, placement left to the server",
   "value": "**131,072** whole, 20,142 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    24914,
    25019
   ],
   "quote": "| Mixture, placement left to the server | **131,072** whole, 20,129 MiB | **131,072** whole, 20,142 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/largest-window-held-whole-placement-left-to-the-server-65-536-whole-20-589-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, placement left to the server",
   "value": "65,536 whole, 20,589 MiB",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25020,
    25113
   ],
   "quote": "| Dense, placement left to the server | 65,536 whole, 20,589 MiB | 65,536 whole, 20,602 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/largest-window-held-whole-placement-left-to-the-server-65-536-whole-20-602-mib",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, placement left to the server",
   "value": "65,536 whole, 20,602 MiB",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25020,
    25113
   ],
   "quote": "| Dense, placement left to the server | 65,536 whole, 20,589 MiB | 65,536 whole, 20,602 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-98-304-wh",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "**98,304** whole, 23,469 MiB",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25114,
    25208
   ],
   "quote": "| Dense, the layer count named | **98,304** whole, 23,469 MiB | **98,304** whole, 23,482 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/largest-window-held-whole-and-the-memory-it-used-the-layer-count-named-98-304-wh",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, and the memory it used, the layer count named",
   "value": "**98,304** whole, 23,482 MiB",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25114,
    25208
   ],
   "quote": "| Dense, the layer count named | **98,304** whole, 23,469 MiB | **98,304** whole, 23,482 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/largest-window-held-whole-the-layer-count-named-refuses-to-load-the-allocator-s-",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "largest window held whole, the layer count named",
   "value": "Refuses to load — the allocator's own out-of-memory error, in the file",
   "non_figure": "refuses to load",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25209,
    25374
   ],
   "quote": "| Dense at 131,072, layer count named | Refuses to load — the allocator's own out-of-memory error, in the file | Refuses to load — the same error, at all four caps |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/largest-window-held-whole-the-layer-count-named-refuses-to-load-the-same-error-a",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "non-figure",
   "metric": "largest window held whole, the layer count named",
   "value": "Refuses to load — the same error, at all four caps",
   "non_figure": "refuses to load",
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "not stated",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25209,
    25374
   ],
   "quote": "| Dense at 131,072, layer count named | Refuses to load — the allocator's own out-of-memory error, in the file | Refuses to load — the same error, at all four caps |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-146-308-tok-s-20-995-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "146.308 tok/s, 20,995 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16` cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25375,
    25485
   ],
   "quote": "| Mixture at 131,072, `f16` cache, both at 350 W | 146.308 tok/s, 20,995 MiB | **159.410 tok/s**, 21,008 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-159-410-tok-s-21-008-mib",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "**159.410 tok/s**, 21,008 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16` cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25375,
    25485
   ],
   "quote": "| Mixture at 131,072, `f16` cache, both at 350 W | 146.308 tok/s, 20,995 MiB | **159.410 tok/s**, 21,008 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-136-618-20-129-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "136.618, 20,129 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q8_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25486,
    25589
   ],
   "quote": "| Mixture at 131,072, `q8_0` — what both were measured at | 136.618, 20,129 MiB | 147.823, 20,142 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-147-823-20-142-mib",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "147.823, 20,142 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q8_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25486,
    25589
   ],
   "quote": "| Mixture at 131,072, `q8_0` — what both were measured at | 136.618, 20,129 MiB | 147.823, 20,142 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": {
    "slug": "one-3090-against-one-3090-ti",
    "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own"
   },
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-135-249-19-415-mib",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "135.249, 19,415 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25590,
    25664
   ],
   "quote": "| Mixture at 131,072, `q4_0` | 135.249, 19,415 MiB | 146.780, 19,428 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 (24,576 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-at-the-full-window-and-the-memory-it-used-146-780-19-428-mib",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode at the full window, and the memory it used",
   "value": "146.780, 19,428 MiB",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    25590,
    25664
   ],
   "quote": "| Mixture at 131,072, `q4_0` | 135.249, 19,415 MiB | 146.780, 19,428 MiB |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    24849,
    24899
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24849,
     24899
    ],
    "header": "| | A 3090 (24,576 MiB) | A 3090 Ti (24,564 MiB) |",
    "head_cell": "A 3090 Ti (24,564 MiB)"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-126-650",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "126.650",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-134-754",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "134.754",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-136-274",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "136.274",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-145-817",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "145.817",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-147-201",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "147.201",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-146-996",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "146.996",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-147-364",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**147.364**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "1",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26562,
    26641
   ],
   "quote": "| 1 | 126.650 | 134.754 | 136.274 | 145.817 | 147.201 | 146.996 | **147.364** |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-195-522",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "195.522",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-210-419",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "210.419",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-214-039",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "214.039",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-228-601",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "228.601",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-231-908",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "231.908",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-232-647",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**232.647**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-231-240",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "231.240",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "2",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26642,
    26721
   ],
   "quote": "| 2 | 195.522 | 210.419 | 214.039 | 228.601 | 231.908 | **232.647** | 231.240 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-265-164",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "265.164",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "250 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 1,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 250 W *(71 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-288-438",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "288.438",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 2,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 300 W *(86 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-294-636",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "294.636",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 3,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090, 350 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-327-124",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "327.124",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "three-more-arms",
   "identity_span": [
    26339,
    26520
   ],
   "cell": {
    "column": 4,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "A 3090 Ti, 300 W *(67 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-337-044",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "337.044",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 5,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "350 W *(78 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-337-842",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "**337.842**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "400 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 6,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "400 W *(89 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-summed-across-conversations-tokens-a-second-337-620",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode summed across conversations, tokens a second",
   "value": "337.620",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "a 4,096-token window",
    "batch_or_steps": "256 tokens out",
    "concurrency": "4",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    26722,
    26801
   ],
   "quote": "| 4 | 265.164 | 288.438 | 294.636 | 327.124 | 337.044 | **337.842** | 337.620 |",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": {
    "column": 7,
    "header_span": [
     26339,
     26520
    ],
    "header": "| Conversations at once | A 3090, 250 W *(71 %)* | A 3090, 300 W *(86 %)* | A 3090, 350 W *(100 %)* | A 3090 Ti, 300 W *(67 %)* | 350 W *(78 %)* | 400 W *(89 %)* | 450 W *(100 %)* |",
    "head_cell": "450 W *(100 %)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-7-24",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "**7.24**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33709,
    33776
   ],
   "quote": "| Mixture model, tokens a second, both at 300 W | **7.24** | 6.88 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-6-88",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "6.88",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33709,
    33776
   ],
   "quote": "| Mixture model, tokens a second, both at 300 W | **7.24** | 6.88 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-2-54",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "**2.54**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33777,
    33842
   ],
   "quote": "| Dense model, tokens a second, both at 300 W | **2.54** | 2.03 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-2-03",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "2.03",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33777,
    33842
   ],
   "quote": "| Dense model, tokens a second, both at 300 W | **2.54** | 2.03 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-7-27",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "**7.27**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33843,
    33918
   ],
   "quote": "| Mixture model, each run as it comes *(350 W / 450 W)* | **7.27** | 6.96 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-6-96",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "6.96",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33843,
    33918
   ],
   "quote": "| Mixture model, each run as it comes *(350 W / 450 W)* | **7.27** | 6.96 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-2-49",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "2.49",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33919,
    33992
   ],
   "quote": "| Dense model, each run as it comes *(350 W / 450 W)* | 2.49 | **2.71** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/tokens-a-second-per-100-of-the-price-as-read-2-71",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens a second per $100 of the price as read",
   "value": "**2.71**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "256 tokens out",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33919,
    33992
   ],
   "quote": "| Dense model, each run as it comes *(350 W / 450 W)* | 2.49 | **2.71** |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/pictures-a-minute-per-100-of-the-price-as-read-2-05",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "pictures a minute per $100 of the price as read",
   "value": "**2.05**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "8 steps, batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33993,
    34065
   ],
   "quote": "| Pictures a minute, 8 steps, batch 3, both at 300 W | **2.05** | 1.95 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/pictures-a-minute-per-100-of-the-price-as-read-1-95",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "pictures a minute per $100 of the price as read",
   "value": "1.95",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "8 steps, batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    33993,
    34065
   ],
   "quote": "| Pictures a minute, 8 steps, batch 3, both at 300 W | **2.05** | 1.95 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/pictures-a-minute-per-100-of-the-price-as-read-2-11",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "pictures a minute per $100 of the price as read",
   "value": "**2.11**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "8 steps, batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    34066,
    34145
   ],
   "quote": "| Pictures a minute, each run as it comes *(350 W / 450 W)* | **2.11** | 2.07 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 1,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 *($1,879.99, a renewed listing)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/pictures-a-minute-per-100-of-the-price-as-read-2-07",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "pictures a minute per $100 of the price as read",
   "value": "2.07",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "450 W",
    "context": "not stated",
    "batch_or_steps": "8 steps, batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    34066,
    34145
   ],
   "quote": "| Pictures a minute, each run as it comes *(350 W / 450 W)* | **2.11** | 2.07 |",
   "identity_basis": "column-head",
   "identity_token": "a 3090 Ti",
   "identity_cite": "what-the-240-buys-on-those-two-readings",
   "identity_span": [
    33586,
    33692
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33586,
     33692
    ],
    "header": "| Per $100 of the price as read | A 3090 *($1,879.99, a renewed listing)* | A 3090 Ti *($2,119.99, new)* |",
    "head_cell": "A 3090 Ti *($2,119.99, new)*"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-exhibit-s-own-margin-over-the-ti-at-the-shared-300-w-cap-11-3",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the Ti at the shared 300 W cap",
   "value": "11.3 %",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "at-a-glance",
    "md_sha": "de6aa50dc9612e16c54cd81d983ffa76cf2d0263b6e697247b89df26ce7e0027"
   },
   "cite": "at-a-glance",
   "span": [
    5358,
    5536
   ],
   "quote": "| Dense model, tokens a second at 65,536 — both at 300 W | **47.838** | 42.969 | A 3090, by 11.3 % — **and the two rows above reverse it**; one cap step is the whole difference |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "at-a-glance",
   "identity_span": [
    5358,
    5536
   ],
   "cell": {
    "column": 3,
    "header_span": [
     4924,
     4957
    ],
    "header": "| | A 3090 | A 3090 Ti | Winner |",
    "head_cell": "Winner"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own reading of it, in the same cell: the two rows above reverse it, and one cap step is the whole difference",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-margin-over-a-3090-on-one-picture-run-as-they-come-9-3",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over a 3090 on one picture, run as they come",
   "value": "9.3 %",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "8 steps, 512 × 512, three at a time",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "at-a-glance",
    "md_sha": "de6aa50dc9612e16c54cd81d983ffa76cf2d0263b6e697247b89df26ce7e0027"
   },
   "cite": "at-a-glance",
   "span": [
    5537,
    5705
   ],
   "quote": "| ↓ one picture, 8 steps, 512 × 512, three at a time — each run as it comes | 1.509 s *(350 W)* | **1.369 s** *(450 W)* | **a 3090 Ti, by 9.3 %** — 0.907× the seconds |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "at-a-glance",
   "identity_span": [
    5537,
    5705
   ],
   "cell": {
    "column": 3,
    "header_span": [
     4924,
     4957
    ],
    "header": "| | A 3090 | A 3090 Ti | Winner |",
    "head_cell": "Winner"
   },
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W — which is why this row states no cap; the page prints the same margin as 0.907× the seconds",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-memory-clock-s-own-reported-maximum-9-751-mhz",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the memory clock's own reported maximum",
   "value": "9,751 MHz",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-question",
    "md_sha": "789bf463d676650e9608f0cce9a17437a09499b0746caf5e6f27c9f139f7e04a"
   },
   "cite": "the-question",
   "span": [
    7912,
    7949
   ],
   "quote": "9,501 MHz against a 9,751 MHz reading",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-memory-clock-s-own-reported-maximum-10-501",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the memory clock's own reported maximum",
   "value": "10,501",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-question",
    "md_sha": "789bf463d676650e9608f0cce9a17437a09499b0746caf5e6f27c9f139f7e04a"
   },
   "cite": "the-question",
   "span": [
    7951,
    7983
   ],
   "quote": "10,251 against a reported 10,501",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-exhibit-s-own-ratio-the-mixture-against-the-dense-model-2-8-times",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, the mixture against the dense model",
   "value": "2.8 times",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-question",
    "md_sha": "789bf463d676650e9608f0cce9a17437a09499b0746caf5e6f27c9f139f7e04a"
   },
   "cite": "the-question",
   "span": [
    8408,
    8456
   ],
   "quote": "the mixture writes 2.8 times a 3090's dense rate",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "the-question",
   "identity_span": [
    8408,
    8456
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-ratio-the-mixture-against-the-dense-model-3-4-times",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, the mixture against the dense model",
   "value": "3.4 times",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-question",
    "md_sha": "789bf463d676650e9608f0cce9a17437a09499b0746caf5e6f27c9f139f7e04a"
   },
   "cite": "the-question",
   "span": [
    8457,
    8480
   ],
   "quote": "and 3.4 times this Ti's",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-exhibit-s-own-change-from-its-300-w-rung-to-its-350-w-rung-2-0-per-cent-slow",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own change from its 300 W rung to its 350 W rung",
   "value": "**2.0 per cent slower**",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    15416,
    15514
   ],
   "quote": "Its dense rung at 350 W reads **2.0 per cent slower** than at 300 W on spreads that do not overlap",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page says in the same sentence that this is not a cap effect: a single-window repeat on a cold card against a ladder rung on a card eight degrees warmer, drawing 16.5 W less at 78.1 per cent of its cap",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/the-exhibit-s-own-change-from-its-300-w-rung-to-its-350-w-rung-0-35-per-cent",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own change from its 300 W rung to its 350 W rung",
   "value": "**0.35 per cent**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
    "md_sha": "149e76a72c9ee765eefd9d6c47f5347e9eb36eb95488b42e33c25e3a356b0234"
   },
   "cite": "what-a-power-cap-takes-and-what-each-card-does-at-its-own",
   "span": [
    15781,
    15816
   ],
   "quote": "Its mixture gains **0.35 per cent**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-ratio-its-seconds-against-a-3090-s-0-907-0-884-and-0-884",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own ratio, its seconds against a 3090's",
   "value": "0.907×, 0.884× and 0.884×",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "batch 3",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "renders",
    "md_sha": "c656c59b571d440a0c1480242b7e7a57e0aa8c2da697f0c5cbc3d17e1d0c1374"
   },
   "cite": "renders",
   "span": [
    20863,
    20945
   ],
   "quote": "**0.907×, 0.884× and 0.884× run as they come**: 9 to 12 per cent off every picture",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W — over the three certified graphs, in the order the render table prints them",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/what-an-unquantised-cache-buys-at-the-full-window-7-8-per-cent-on-the-ti",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what an unquantised cache buys at the full window",
   "value": "**+7.8 per cent on the Ti**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`, unquantised",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    26039,
    26066
   ],
   "quote": "**+7.8 per cent on the Ti**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/what-an-unquantised-cache-buys-at-the-full-window-7-1-on-a-3090",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "what an unquantised cache buys at the full window",
   "value": "**+7.1 on a 3090**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "Q4_K_M · `f16`, unquantised",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-fits-and-what-the-cache-costs",
    "md_sha": "18755e438eb86ba8088d858dabe51b17555258c362512efbbda74490e3a965bf"
   },
   "cite": "what-fits-and-what-the-cache-costs",
   "span": [
    26071,
    26089
   ],
   "quote": "**+7.1 on a 3090**",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "what-fits-and-what-the-cache-costs",
   "identity_span": [
    26071,
    26089
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/prompt-reading-tokens-a-second-2-549-966-tokens-a-second",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "prompt reading, tokens a second",
   "value": "**2,549.966 tokens a second**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27260,
    27314
   ],
   "quote": "is taken in at **2,549.966 tokens a second** by the Ti",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/prompt-reading-tokens-a-second-2-270-547",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "prompt reading, tokens a second",
   "value": "**2,270.547**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27315,
    27346
   ],
   "quote": "against **2,270.547** by a 3090",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    27315,
    27346
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-once-the-window-is-full-tokens-a-second-80-056",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode once the window is full, tokens a second",
   "value": "**80.056**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27348,
    27400
   ],
   "quote": "and written through at **80.056** against **73.327**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/decode-once-the-window-is-full-tokens-a-second-73-327",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode once the window is full, tokens a second",
   "value": "**73.327**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "350 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27348,
    27400
   ],
   "quote": "and written through at **80.056** against **73.327**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/prompt-reading-tokens-a-second-2-628-756",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "prompt reading, tokens a second",
   "value": "2,628.756",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27487,
    27533
   ],
   "quote": "the Ti's 450 W file reads 2,628.756 and 79.740",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/decode-once-the-window-is-full-tokens-a-second-79-740",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode once the window is full, tokens a second",
   "value": "79.740",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "Q4_K_M",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "450 W",
    "context": "98,000 tokens in a 131,072-token window",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27487,
    27533
   ],
   "quote": "the Ti's 450 W file reads 2,628.756 and 79.740",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/idle-draw-with-the-model-resident-median-and-mean-median-of-18-74-w-and-a-mean-o",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "idle draw with the model resident, median and mean",
   "value": "**median of 18.74 W and a mean of 43.04**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "57 samples over thirty seconds each",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27773,
    27827
   ],
   "quote": "a Ti reads a **median of 18.74 W and a mean of 43.04**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own reading of the pair: the median is the idle figure and the mean an upper bound containing a load",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/idle-draw-with-the-model-resident-median-and-mean-22-26-and-52-76",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "idle draw with the model resident, median and mean",
   "value": "**22.26 and 52.76**",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "57 samples over thirty seconds each",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27829,
    27856
   ],
   "quote": "a 3090, **22.26 and 52.76**",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "three-more-arms",
   "identity_span": [
    27829,
    27856
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own reading of the pair: the median is the idle figure and the mean an upper bound containing a load",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/idle-draw-the-empty-and-the-loaded-window-sharing-one-median-22-39-w",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "idle draw, the empty and the loaded window sharing one median",
   "value": "**22.39 W**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    27937,
    28006
   ],
   "quote": "its empty and loaded windows sharing the identical **22.39 W** median",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints this as its evidence that the gap between the median and the mean is a load spike inside the window, not a difference between an empty card and a loaded one",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-difference-in-what-each-card-costs-to-leave-on-6-a-year-on-the",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own difference in what each card costs to leave on",
   "value": "$6 a year on the medians and $16 on the means",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "three-more-arms",
    "md_sha": "54117837457a8660b649d7789ec9ba798329e5576167b264f96df02a23df0fbf"
   },
   "cite": "three-more-arms",
   "span": [
    28375,
    28475
   ],
   "quote": "**the faster card is the cheaper one to leave on, by $6 a year on the medians and $16 on the means**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "at 18.34 cents a kilowatt-hour, the United States residential average for June 2026 as the Energy Information Administration publishes it; board power, so a wall meter would read more for both",
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/weight-on-a-kitchen-scale-1-203-g",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "weight, on a kitchen scale",
   "value": "**1,203 g**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "Measured on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    29313,
    29367
   ],
   "quote": "the XC3 Ultra **1,203 g** and **about 2 inches** thick",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/thickness-by-tape-measure-about-2-inches",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "thickness, by tape measure",
   "value": "**about 2 inches**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "Measured on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    29313,
    29367
   ],
   "quote": "the XC3 Ultra **1,203 g** and **about 2 inches** thick",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/weight-on-a-kitchen-scale-2-157-g",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "weight, on a kitchen scale",
   "value": "**2,157 g**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "Measured on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    29369,
    29424
   ],
   "quote": "the FTW3 Ultra Ti **2,157 g** and **about 2.75 inches**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/thickness-by-tape-measure-about-2-75-inches",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "thickness, by tape measure",
   "value": "**about 2.75 inches**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "Measured on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    29369,
    29424
   ],
   "quote": "the FTW3 Ultra Ti **2,157 g** and **about 2.75 inches**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-difference-in-weight-and-thickness-954-g-heavier-1-79-times-th",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own difference in weight and thickness",
   "value": "954 g heavier, 1.79 times the weight, and about three quarters of an inch thicker",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "Measured on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    29427,
    29512
   ],
   "quote": "**954 g heavier, 1.79 times the weight, and about three quarters of an inch thicker**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/memory-chips-their-count-and-size-and-where-they-sit-twelve-2-gb-memory-chips-ev",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "memory chips, their count and size, and where they sit",
   "value": "twelve 2 GB memory chips, every one face-up beside the processor",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    28631,
    28713
   ],
   "quote": "A 3090 Ti carries twelve 2 GB memory chips, every one face-up beside the processor",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "the-card-itself-as-an-object",
   "identity_span": [
    28631,
    28713
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/memory-chips-their-count-and-size-and-where-they-sit-twenty-four-1-gb-chips-twel",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "memory chips, their count and size, and where they sit",
   "value": "twenty-four 1 GB chips, twelve on the front and twelve on the *back*",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-card-itself-as-an-object",
    "md_sha": "502b75dd38242d65e0ef5ae53269b9c2939d573a434d61e760b1d68d7ca6d2be"
   },
   "cite": "the-card-itself-as-an-object",
   "span": [
    28715,
    28798
   ],
   "quote": "a 3090 carries twenty-four 1 GB chips, twelve on the front and twelve on the *back*",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "the-card-itself-as-an-object",
   "identity_span": [
    28715,
    28798
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own consequence: this is the heat a person feels through a 3090's backplate, and the reason that backplate is a cooling part",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-gain-at-400-w-over-350-w-the-32-step-graph-400-w-wins-2-24-per",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own gain at 400 W over 350 W, the 32-step graph",
   "value": "**400 W wins 2.24 per cent**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "400 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "the-setting-this-page-would-choose-and-the-dissent",
    "md_sha": "1ca6d4bf206ca1a360d5a5513ba35e017160cb6373be7f3c60e9da8728ab7e1e"
   },
   "cite": "the-setting-this-page-would-choose-and-the-dissent",
   "span": [
    35241,
    35305
   ],
   "quote": "**400 W wins 2.24 per cent** there (5.1296 → 5.0150 s per image)",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [
    "the-instrument-and-the-control"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/the-exhibit-s-own-margin-in-board-power-over-a-3090-17-7-per-cent-more-board-pow",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin in board power over a 3090",
   "value": "**17.7 per cent** more board power on the mixture model and **53.2** more on the dense one",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-the-240-buys-on-those-two-readings",
    "md_sha": "b8cdf43cbd7e2b47f687a3a37234004fd0294d2ef6b3000a053087dfdecaa14f"
   },
   "cite": "what-the-240-buys-on-those-two-readings",
   "span": [
    34738,
    34837
   ],
   "quote": "it draws **17.7 per cent** more board power on the mixture model and **53.2** more on the dense one",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/one-3090-against-one-3090-ti/price-one-card-new-2-119-99",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, one card, new",
   "value": "**$2,119.99**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "ordered 15 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-it-cost",
    "md_sha": "ba1ce693c3c4c4f2c92d5d93e69a72e1022d27c9633520ce2878ea022dc7aa46"
   },
   "cite": "what-it-cost",
   "span": [
    31323,
    31386
   ],
   "quote": "bought **new** for **$2,119.99**, ordered **15 September 2026**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4194,
    4251
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page prints the listing title it was bought under: EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming, 24G-P5-4985-KR, 24GB GDDR6X, iCX3, ARGB LED, Backplate, Free eLeash",
   "record_published": true
  },
  {
   "id": "rtx-3090/one-3090-against-one-3090-ti/price-one-card-renewed-1-879-99",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, one card, renewed",
   "value": "**$1,879.99**",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "read by hand at the store page on 17 September 2026"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "one-3090-against-one-3090-ti",
    "section": "what-it-cost",
    "md_sha": "ba1ce693c3c4c4f2c92d5d93e69a72e1022d27c9633520ce2878ea022dc7aa46"
   },
   "cite": "what-it-cost",
   "span": [
    31962,
    32035
   ],
   "quote": "at **$1,879.99**, read by hand at the store page on **17 September 2026**",
   "identity_basis": "page",
   "identity_token": "one EVGA GeForce RTX 3090 XC3 Ultra 24 GB",
   "identity_cite": "the-two-cards-named-once",
   "identity_span": [
    4014,
    4067
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this 3090 has no price on record — it was already in the workshop — and this is a listing of the exact board, read by hand because on 18 September both retailers served no price to this workshop's fetcher",
   "record_published": true
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-median-15-47",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, median",
   "value": "15.47",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-18 20:41 → 22:57"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5113,
    5244
   ],
   "quote": "| A 3090 Ti, the first creations on the new machine | 350 W | Live journal | 7 | 15.47 | 15.42 | 15.52 | 2026-09-18 20:41 → 22:57 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5113,
    5244
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Median s per batch of three"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-minimum-15-42",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, minimum",
   "value": "15.42",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-18 20:41 → 22:57"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5113,
    5244
   ],
   "quote": "| A 3090 Ti, the first creations on the new machine | 350 W | Live journal | 7 | 15.47 | 15.42 | 15.52 | 2026-09-18 20:41 → 22:57 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5113,
    5244
   ],
   "cell": {
    "column": 5,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Min s"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-maximum-15-52",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, maximum",
   "value": "15.52",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-18 20:41 → 22:57"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5113,
    5244
   ],
   "quote": "| A 3090 Ti, the first creations on the new machine | 350 W | Live journal | 7 | 15.47 | 15.42 | 15.52 | 2026-09-18 20:41 → 22:57 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5113,
    5244
   ],
   "cell": {
    "column": 6,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Max s"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-median-15-39",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, median",
   "value": "15.39",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5245,
    5307
   ],
   "quote": "| A 3090 Ti | 350 W | Bench | 2 | 15.39 | — | — | 2026-09-17 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5245,
    5307
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Median s per batch of three"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-median-17-01",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, median",
   "value": "17.01",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5308,
    5367
   ],
   "quote": "| A 3090 | 350 W | Bench | 2 | 17.01 | — | — | 2026-09-17 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5308,
    5367
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Median s per batch of three"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-median-17-48",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, median",
   "value": "17.48",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5368,
    5427
   ],
   "quote": "| A 3090 | 300 W | Bench | 2 | 17.48 | — | — | 2026-09-17 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5368,
    5427
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Median s per batch of three"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/seconds-a-batch-of-three-median-18-89",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds a batch of three, median",
   "value": "18.89",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "250 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "2026-09-17"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5428,
    5487
   ],
   "quote": "| A 3090 | 250 W | Bench | 2 | 18.89 | — | — | 2026-09-17 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5428,
    5487
   ],
   "cell": {
    "column": 4,
    "header_span": [
     4985,
     5073
    ],
    "header": "| Card | Cap | Source | N | Median s per batch of three | Min s | Max s | Window (UTC) |",
    "head_cell": "Median s per batch of three"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/milliseconds-a-sleeve-minimum-2-060",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "milliseconds a sleeve, minimum",
   "value": "2,060",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-18 20:42 → 20:49"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6133,
    6222
   ],
   "quote": "| The new painter, a 3090 Ti | 350 W | 2 | — | 2,060 | 2,125 | 2026-09-18 20:42 → 20:49 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    6133,
    6222
   ],
   "cell": {
    "column": 4,
    "header_span": [
     6031,
     6097
    ],
    "header": "| Painter | Cap | N | Median ms | Min ms | Max ms | Window (UTC) |",
    "head_cell": "Min ms"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/milliseconds-a-sleeve-maximum-2-125",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "milliseconds a sleeve, maximum",
   "value": "2,125",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-18 20:42 → 20:49"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6133,
    6222
   ],
   "quote": "| The new painter, a 3090 Ti | 350 W | 2 | — | 2,060 | 2,125 | 2026-09-18 20:42 → 20:49 |",
   "identity_basis": "row",
   "identity_token": "a 3090 Ti",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    6133,
    6222
   ],
   "cell": {
    "column": 5,
    "header_span": [
     6031,
     6097
    ],
    "header": "| Painter | Cap | N | Median ms | Min ms | Max ms | Window (UTC) |",
    "head_cell": "Max ms"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/milliseconds-a-sleeve-median-2-551",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "milliseconds a sleeve, median",
   "value": "2,551",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-11 14:33 → 2026-09-18 01:38"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6223,
    6325
   ],
   "quote": "| The old painter, a 3090 | 280 W | 16 | 2,551 | 2,497 | 2,899 | 2026-09-11 14:33 → 2026-09-18 01:38 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    6223,
    6325
   ],
   "cell": {
    "column": 3,
    "header_span": [
     6031,
     6097
    ],
    "header": "| Painter | Cap | N | Median ms | Min ms | Max ms | Window (UTC) |",
    "head_cell": "Median ms"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/milliseconds-a-sleeve-minimum-2-497",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "milliseconds a sleeve, minimum",
   "value": "2,497",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-11 14:33 → 2026-09-18 01:38"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6223,
    6325
   ],
   "quote": "| The old painter, a 3090 | 280 W | 16 | 2,551 | 2,497 | 2,899 | 2026-09-11 14:33 → 2026-09-18 01:38 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    6223,
    6325
   ],
   "cell": {
    "column": 4,
    "header_span": [
     6031,
     6097
    ],
    "header": "| Painter | Cap | N | Median ms | Min ms | Max ms | Window (UTC) |",
    "head_cell": "Min ms"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/milliseconds-a-sleeve-maximum-2-899",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "milliseconds a sleeve, maximum",
   "value": "2,899",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "2026-09-11 14:33 → 2026-09-18 01:38"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6223,
    6325
   ],
   "quote": "| The old painter, a 3090 | 280 W | 16 | 2,551 | 2,497 | 2,899 | 2026-09-11 14:33 → 2026-09-18 01:38 |",
   "identity_basis": "row",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    6223,
    6325
   ],
   "cell": {
    "column": 5,
    "header_span": [
     6031,
     6097
    ],
    "header": "| Painter | Cap | N | Median ms | Min ms | Max ms | Window (UTC) |",
    "head_cell": "Max ms"
   },
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-over-a-3090-on-the-heavy-plate-at-a-shared-cap-8-5-per-",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over a 3090 on the heavy plate at a shared cap",
   "value": "8.5 per cent",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "why-this-card-and-why-350-w",
    "md_sha": "27965ba8ed79a05cacbe7577899ed51ca73c87595ee26924e9f4f3d078b3289b"
   },
   "cite": "why-this-card-and-why-350-w",
   "span": [
    3603,
    3673
   ],
   "quote": "this Ti draws the heavy plate 8.5 per cent faster than a 3090 at 300 W",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "both boards held at 300 W. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-over-a-3090-on-the-heavy-plate-at-a-shared-cap-9-5",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over a 3090 on the heavy plate at a shared cap",
   "value": "9.5",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "why-this-card-and-why-350-w",
    "md_sha": "27965ba8ed79a05cacbe7577899ed51ca73c87595ee26924e9f4f3d078b3289b"
   },
   "cite": "why-this-card-and-why-350-w",
   "span": [
    3646,
    3688
   ],
   "quote": "faster than a 3090 at 300 W and 9.5 at 350",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "both boards held at 350 W, which is a 3090's own limit and 78 per cent of this Ti's. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-over-a-3090-at-the-300-w-setting-nearest-the-old-machin",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over a 3090 at the 300 W setting nearest the old machine's cap",
   "value": "12 per cent",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "why-this-card-and-why-350-w",
    "md_sha": "27965ba8ed79a05cacbe7577899ed51ca73c87595ee26924e9f4f3d078b3289b"
   },
   "cite": "why-this-card-and-why-350-w",
   "span": [
    3747,
    3810
   ],
   "quote": "with this Ti at the 350 W the new machine now runs, 12 per cent",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "not a matched-cap comparison and the page says so in the same clause: this Ti at the 350 W the new machine runs, against a 3090 at the 300 W setting nearest the old machine's 280 W. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-in-core-temperature-over-a-3090-at-a-shared-cap-sixteen",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin in core temperature over a 3090 at a shared cap",
   "value": "sixteen degrees cooler at the plateau and fourteen at the peak",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "why-this-card-and-why-350-w",
    "md_sha": "27965ba8ed79a05cacbe7577899ed51ca73c87595ee26924e9f4f3d078b3289b"
   },
   "cite": "why-this-card-and-why-350-w",
   "span": [
    4120,
    4182
   ],
   "quote": "sixteen degrees cooler at the plateau and fourteen at the peak",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a 55 °C plateau and a 60 °C peak against a 3090's 71 and 74, both cards held at 350 W. The page prints the companion's own caveat beside it — those rows did not do equal work. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else",
   "record_published": null
  },
  {
   "id": "rtx-3090/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-bracket-for-the-old-machine-s-280-w-cap-seconds-a-batch-of-thr",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own bracket for the old machine's 280 W cap, seconds a batch of three",
   "value": "17.5 to 18.9 seconds per batch",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "280 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5555,
    5641
   ],
   "quote": "the honest before is a range: 17.5 to 18.9 seconds per batch on the bench's instrument",
   "identity_basis": "section",
   "identity_token": "a 3090",
   "identity_cite": "before-and-after-on-the-same-instruments",
   "identity_span": [
    5428,
    5454
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a bracket and not a reading: the bench ran 250 and 300 W rungs and no 280 W one, and the page rounds the two rows either side of the old machine's cap rather than interpolating between them",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-over-the-old-machine-at-its-own-cap-11-5-to-18-1-per-ce",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin over the old machine at its own cap",
   "value": "11.5 to 18.1 per cent faster",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5736,
    5812
   ],
   "quote": "the move is 11.5 to 18.1 per cent faster than the old machine at its own cap",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "a live median of seven creations on this Ti at 350 W read against the bracketed range for the old machine at 280 W, so the answer is a range: one instrument on each side, which is the join this page exists to make",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-spread-of-the-live-creations-above-the-bench-median-between-0-",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own spread of the live creations above the bench median",
   "value": "between 0.2 and 0.9 per cent above the bench's median",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "a batch of three, on the same 32-step graph the open set hangs",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    5891,
    5973
   ],
   "quote": "The seven live creations sit between 0.2 and 0.9 per cent above the bench's median",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own reason for quoting a bench figure at all: seven creations off a live service landed within a per cent of what two timed submits predicted",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/two-3090s-out-one-3090-ti-in/the-exhibit-s-own-margin-under-the-old-painter-s-median-17-and-19-per-cent-under",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "the exhibit's own margin under the old painter's median",
   "value": "17 and 19 per cent under the old median",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein 4B",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "350 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "two-3090s-out-one-3090-ti-in",
    "section": "before-and-after-on-the-same-instruments",
    "md_sha": "ead7fbce049160f65c850682568b83305ec0d216ee2380278bdc70e09789b6c1"
   },
   "cite": "before-and-after-on-the-same-instruments",
   "span": [
    6359,
    6432
   ],
   "quote": "the after stands as two readings: 17 and 19 per cent under the old median",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "what-moved-and-from-where",
   "identity_span": [
    2681,
    2725
   ],
   "cell": null,
   "conditions_cites": [
    "what-moved-and-from-where"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "two sleeves against a median of sixteen on the old card at 280 W, and the page says in the same sentence that two sleeves is not a median of its own",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-118-424",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "118.424",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    22964,
    23064
   ],
   "quote": "| **This board, fixed** | 95 W | 118.424 | 372.49 | 79.64 (76.95 · 79.64 · 94.22) | 672.50 | 148.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-372-49",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "372.49",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    22964,
    23064
   ],
   "quote": "| **This board, fixed** | 95 W | 118.424 | 372.49 | 79.64 (76.95 · 79.64 · 94.22) | 672.50 | 148.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-the-three-runs-means-79-64-76-95-79-64-94-22",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (the three runs' means)",
   "value": "79.64 (76.95 · 79.64 · 94.22)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    22964,
    23064
   ],
   "quote": "| **This board, fixed** | 95 W | 118.424 | 372.49 | 79.64 (76.95 · 79.64 · 94.22) | 672.50 | 148.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (the three runs' means)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-672-50",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "672.50",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    22964,
    23064
   ],
   "quote": "| **This board, fixed** | 95 W | 118.424 | 372.49 | 79.64 (76.95 · 79.64 · 94.22) | 672.50 | 148.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-148-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "148.7",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    22964,
    23064
   ],
   "quote": "| **This board, fixed** | 95 W | 118.424 | 372.49 | 79.64 (76.95 · 79.64 · 94.22) | 672.50 | 148.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-155-251",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "155.251",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23065,
    23203
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 155.251 | 372.92 | 122.57 (98.58 · 146.10 · 122.57) | 787.76 | 126.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-372-92",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "372.92",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23065,
    23203
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 155.251 | 372.92 | 122.57 (98.58 · 146.10 · 122.57) | 787.76 | 126.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-the-three-runs-means-122-57-98-58-146-10-122-57",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (the three runs' means)",
   "value": "122.57 (98.58 · 146.10 · 122.57)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23065,
    23203
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 155.251 | 372.92 | 122.57 (98.58 · 146.10 · 122.57) | 787.76 | 126.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (the three runs' means)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-787-76",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "787.76",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23065,
    23203
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 155.251 | 372.92 | 122.57 (98.58 · 146.10 · 122.57) | 787.76 | 126.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-126-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "126.7",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23065,
    23203
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 155.251 | 372.92 | 122.57 (98.58 · 146.10 · 122.57) | 787.76 | 126.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-155-288",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "155.288",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23204,
    23332
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 155.288 | 371.01 | 116.12 (116.12 · 115.76 · 146.42) | 747.97 | 133.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-371-01",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "371.01",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23204,
    23332
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 155.288 | 371.01 | 116.12 (116.12 · 115.76 · 146.42) | 747.97 | 133.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-the-three-runs-means-116-12-116-12-115-76-146-42",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (the three runs' means)",
   "value": "116.12 (116.12 · 115.76 · 146.42)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23204,
    23332
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 155.288 | 371.01 | 116.12 (116.12 · 115.76 · 146.42) | 747.97 | 133.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (the three runs' means)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-747-97",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "747.97",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23204,
    23332
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 155.288 | 371.01 | 116.12 (116.12 · 115.76 · 146.42) | 747.97 | 133.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-133-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "133.7",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    23204,
    23332
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 155.288 | 371.01 | 116.12 (116.12 · 115.76 · 146.42) | 747.97 | 133.7 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     22795,
     22933
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (the three runs' means) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. The page prints the three runs' means beside the median so the coarseness is visible: a mixture-model run here is about two seconds long, eight to ten samples at 2 Hz, and the three means spread widely for the same speed — the joule and per-100-W cells inherit it. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-28-839",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "28.839",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24528,
    24604
   ],
   "quote": "| **This board, fixed** | 95 W | 28.839 | 210.94 | 90.31 | 3,131.56 | 31.9 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-210-94",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "210.94",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24528,
    24604
   ],
   "quote": "| **This board, fixed** | 95 W | 28.839 | 210.94 | 90.31 | 3,131.56 | 31.9 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-90-31",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "90.31",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24528,
    24604
   ],
   "quote": "| **This board, fixed** | 95 W | 28.839 | 210.94 | 90.31 | 3,131.56 | 31.9 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-3-131-56",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "3,131.56",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24528,
    24604
   ],
   "quote": "| **This board, fixed** | 95 W | 28.839 | 210.94 | 90.31 | 3,131.56 | 31.9 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-31-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "31.9",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24528,
    24604
   ],
   "quote": "| **This board, fixed** | 95 W | 28.839 | 210.94 | 90.31 | 3,131.56 | 31.9 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-44-897",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "44.897",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24605,
    24717
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 44.897 | 203.72 | 135.76 | 3,019.40 | 33.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-203-72",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "203.72",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24605,
    24717
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 44.897 | 203.72 | 135.76 | 3,019.40 | 33.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-135-76",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "135.76",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24605,
    24717
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 44.897 | 203.72 | 135.76 | 3,019.40 | 33.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-3-019-40",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "3,019.40",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24605,
    24717
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 44.897 | 203.72 | 135.76 | 3,019.40 | 33.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-33-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "33.1",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140–150 W (n=374; 150 W in 335)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24605,
    24717
   ],
   "quote": "| **This board, dynamic boost** | 140–150 W (n=374; 150 W in 335) | 44.897 | 203.72 | 135.76 | 3,019.40 | 33.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-44-702",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s",
   "value": "44.702",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24718,
    24819
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 44.702 | 191.65 | 129.90 | 2,904.35 | 34.4 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-191-65",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "191.65",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24718,
    24819
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 44.702 | 191.65 | 129.90 | 2,904.35 | 34.4 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-129-90",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "129.90",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24718,
    24819
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 44.702 | 191.65 | 129.90 | 2,904.35 | 34.4 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-2-904-35",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "2,904.35",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24718,
    24819
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 44.702 | 191.65 | 129.90 | 2,904.35 | 34.4 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-per-100-w-34-4",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "tokens/s per 100 W",
   "value": "34.4",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat (n=385)",
    "context": "65,536",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    24718,
    24819
   ],
   "quote": "| **This board, its boot state** | 150 W, flat (n=385) | 44.702 | 191.65 | 129.90 | 2,904.35 | 34.4 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     24376,
     24497
    ],
    "header": "| Configuration | Power setting | Tokens/s | First token, ms | Board W (mean) | J per 1,000 tokens | Tokens/s per 100 W |",
    "head_cell": "Tokens/s per 100 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. On the dense model the runs are twenty seconds long and the three means agree within a few per cent, which is the page's own reason for printing one mean here and three above. The exhibit prints this ratio; the roster computes none.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/largest-window-held-whole-131-072-the-ladder-ran-out-of-rungs-first",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole",
   "value": "131,072 — the ladder ran out of rungs first",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "at every posture",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    25591,
    25698
   ],
   "quote": "| `gemma4:26b` | 131,072 — the ladder ran out of rungs first | 131,072 | 131,072 | 131,072, layers forced |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     25450,
     25568
    ],
    "header": "| Model | This board, at every posture | A 3090 at 350 W | A 3090 Ti at 350 W and at 450 W | Two 3080s at 300 W each |",
    "head_cell": "This board, at every posture"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/largest-window-held-whole-65-536-spilled-at-98-304",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole",
   "value": "65,536 — spilled at 98,304",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "at every posture",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "not stated"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    25699,
    25836
   ],
   "quote": "| `mistral-small3.2:24b` | 65,536 — spilled at 98,304 | 65,536 — spilled at 98,304 | 65,536 — spilled at 98,304 | 49,152, layers forced |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     25450,
     25568
    ],
    "header": "| Model | This board, at every posture | A 3090 at 350 W | A 3090 Ti at 350 W and at 450 W | Two 3080s at 300 W each |",
    "head_cell": "This board, at every posture"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/images-in-ten-minutes-294-600-9-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "images in ten minutes",
   "value": "294 (600.9 s)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Images in ten minutes"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/seconds-per-image-spread-1-958-1-918-2-019-n-294",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per image (spread)",
   "value": "1.958 (1.918–2.019, n=294)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Seconds per image (spread)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-ten-minute-window-0-0529",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the ten-minute window",
   "value": "0.0529",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-temperature-plateau-54-c-at-3-3-min",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "54 °C at 3.3 min",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Plateau (°C, and when reached)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-58",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core max °C",
   "value": "58",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-median-max-93-8-100-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W median/max",
   "value": "93.8 / 100.6",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26201,
    26363
   ],
   "quote": "| **This board, fixed** | 95 W | 294 (600.9 s) | 1.958 (1.918–2.019, n=294) | 0.0529 | 54 °C at 3.3 min | 58 | 93.8 / 100.6 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 7,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Board W median/max"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/images-in-ten-minutes-378-600-6-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "images in ten minutes",
   "value": "378 (600.6 s)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Images in ten minutes"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/seconds-per-image-spread-1-521-1-476-1-574-n-378",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per image (spread)",
   "value": "1.521 (1.476–1.574, n=378)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Seconds per image (spread)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-ten-minute-window-0-0633",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the ten-minute window",
   "value": "0.0633",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-temperature-plateau-68-c-at-5-5-min",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "68 °C at 5.5 min",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Plateau (°C, and when reached)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-71",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core max °C",
   "value": "71",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-median-max-145-9-156-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W median/max",
   "value": "145.9 / 156.9",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,141 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26364,
    26579
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — every one of 1,141 samples at 2 Hz | 378 (600.6 s) | 1.521 (1.476–1.574, n=378) | 0.0633 | 68 °C at 5.5 min | 71 | 145.9 / 156.9 | Unreadable; set to maximum by hand |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 7,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Board W median/max"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/images-in-ten-minutes-382-600-5-s",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "images in ten minutes",
   "value": "382 (600.5 s)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Images in ten minutes"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/seconds-per-image-spread-1-507-1-475-1-590-n-382",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "seconds per image (spread)",
   "value": "1.507 (1.475–1.590, n=382)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Seconds per image (spread)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-ten-minute-window-0-0629",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the ten-minute window",
   "value": "0.0629",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-temperature-plateau-68-c-at-4-6-min",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core temperature plateau",
   "value": "68 °C at 4.6 min",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Plateau (°C, and when reached)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-70",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "core max °C",
   "value": "70",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-median-max-146-4-155-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W median/max",
   "value": "146.4 / 155.2",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — every one of 1,140 samples at 2 Hz",
    "context": "not stated",
    "batch_or_steps": "Ten minutes of continuous drawing — the 8-step recipe, one image at a time",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    26580,
    26801
   ],
   "quote": "| **This board, its boot state** | 150 W, flat — every one of 1,140 samples at 2 Hz | 382 (600.5 s) | 1.507 (1.475–1.590, n=382) | 0.0629 | 68 °C at 4.6 min | 70 | 146.4 / 155.2 | Unreadable; not confirmed for this pass |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 7,
    "header_span": [
     25977,
     26162
    ],
    "header": "| Board and posture | Power setting | Images in ten minutes | Seconds per image (spread) | Wh per image | Plateau (°C, and when reached) | Core max °C | Board W median/max | Fan max % |",
    "head_cell": "Board W median/max"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-1-640-1-583-1-698",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "1.640 (1.583–1.698)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "8 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-0431",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.0431",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "8 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-1-627-1-579-1-675",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "1.627 (1.579–1.675)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "8 steps + pixel pass · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps + pixel pass · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-0429",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.0429",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "8 steps + pixel pass · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps + pixel pass · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-6-086-6-028-6-145",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "6.086 (6.028–6.145)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "32 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "32 steps · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-1607",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.1607",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "32 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28842,
    28986
   ],
   "quote": "| **This board, fixed** | 95 W | 1.640 (1.583–1.698) | 0.0431 | 1.627 (1.579–1.675) | 0.0429 | 6.086 (6.028–6.145) | 0.1607 | 94.3 / 96.0 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 7,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "32 steps · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-1-246-1-202-1-290",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "1.246 (1.202–1.290)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "8 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-0500",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.0500",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "8 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-1-245-1-194-1-296",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "1.245 (1.194–1.296)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "8 steps + pixel pass · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps + pixel pass · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-0513",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.0513",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "8 steps + pixel pass · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "8 steps + pixel pass · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/s-per-image-spread-n-2-4-649-4-584-4-715",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "s per image (spread, n=2)",
   "value": "4.649 (4.584–4.715)",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "32 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "32 steps · s per image (spread, n=2)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/wh-per-image-over-the-timed-batches-0-1934",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "Wh per image, over the timed batches",
   "value": "0.1934",
   "non_figure": null,
   "conditions": {
    "model": "FLUX.2 klein-4B",
    "quant": "not stated",
    "runtime": "ComfyUI",
    "runtime_version": "0.21.1",
    "cap": "150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches",
    "context": "not stated",
    "batch_or_steps": "32 steps · two timed batches of three images each after a warm-up submit that is excluded",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    28987,
    29209
   ],
   "quote": "| **This board, dynamic boost** | 150 W, flat — 82 samples at 2 Hz across the three recipes' timed batches | 1.246 (1.202–1.290) | 0.0500 | 1.245 (1.194–1.296) | 0.0513 | 4.649 (4.584–4.715) | 0.1934 | 148.1 / 151.3 | 50 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 7,
    "header_span": [
     28492,
     28799
    ],
    "header": "| Board and posture | Power setting | 8 steps · s per image (spread, n=2) | 8 steps · Wh per image | 8 steps + pixel pass · s per image (spread, n=2) | 8 steps + pixel pass · Wh per image | 32 steps · s per image (spread, n=2) | 32 steps · Wh per image | Board W median/max, 8 steps | Core max °C, 8 steps |",
    "head_cell": "32 steps · Wh per image"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. Watt-hours an image is energy over the timed window divided by the images finished in it, from each run's own power trace — the page's own definition, and it warns that the ten-minute and the batch tables are therefore not the same quantity. This board's render rows ran the same ComfyUI release as the desktop boards on a newer torch and driver, which a Blackwell board needs, so the page reads its render ratios as box against board rather than card against card.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-118-368",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "118.368",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33306,
    33373
   ],
   "quote": "| Fixed, 95 W | 1 | 118.368 | 102.226 | 118.368 | 345.52 | 857.75 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-102-226",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "102.226",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33306,
    33373
   ],
   "quote": "| Fixed, 95 W | 1 | 118.368 | 102.226 | 118.368 | 345.52 | 857.75 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-118-368",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "118.368",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33306,
    33373
   ],
   "quote": "| Fixed, 95 W | 1 | 118.368 | 102.226 | 118.368 | 345.52 | 857.75 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-345-52",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "345.52",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33306,
    33373
   ],
   "quote": "| Fixed, 95 W | 1 | 118.368 | 102.226 | 118.368 | 345.52 | 857.75 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-857-75",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "857.75",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33306,
    33373
   ],
   "quote": "| Fixed, 95 W | 1 | 118.368 | 102.226 | 118.368 | 345.52 | 857.75 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-199-413",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "199.413",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33374,
    33440
   ],
   "quote": "| Fixed, 95 W | 2 | 199.413 | 165.321 | 99.950 | 525.93 | 504.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-165-321",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "165.321",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33374,
    33440
   ],
   "quote": "| Fixed, 95 W | 2 | 199.413 | 165.321 | 99.950 | 525.93 | 504.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-99-950",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "99.950",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33374,
    33440
   ],
   "quote": "| Fixed, 95 W | 2 | 199.413 | 165.321 | 99.950 | 525.93 | 504.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-525-93",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "525.93",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33374,
    33440
   ],
   "quote": "| Fixed, 95 W | 2 | 199.413 | 165.321 | 99.950 | 525.93 | 504.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-504-27",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "504.27",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33374,
    33440
   ],
   "quote": "| Fixed, 95 W | 2 | 199.413 | 165.321 | 99.950 | 525.93 | 504.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-275-732",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "275.732",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33441,
    33507
   ],
   "quote": "| Fixed, 95 W | 4 | 275.732 | 231.239 | 69.255 | 702.42 | 396.86 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-231-239",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "231.239",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33441,
    33507
   ],
   "quote": "| Fixed, 95 W | 4 | 275.732 | 231.239 | 69.255 | 702.42 | 396.86 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-69-255",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "69.255",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33441,
    33507
   ],
   "quote": "| Fixed, 95 W | 4 | 275.732 | 231.239 | 69.255 | 702.42 | 396.86 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-702-42",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "702.42",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33441,
    33507
   ],
   "quote": "| Fixed, 95 W | 4 | 275.732 | 231.239 | 69.255 | 702.42 | 396.86 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-396-86",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "396.86",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33441,
    33507
   ],
   "quote": "| Fixed, 95 W | 4 | 275.732 | 231.239 | 69.255 | 702.42 | 396.86 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-155-231",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "155.231",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33508,
    33577
   ],
   "quote": "| Dynamic boost | 1 | 155.231 | 124.544 | 155.231 | 401.70 | 829.49 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-124-544",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "124.544",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33508,
    33577
   ],
   "quote": "| Dynamic boost | 1 | 155.231 | 124.544 | 155.231 | 401.70 | 829.49 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-155-231",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "155.231",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33508,
    33577
   ],
   "quote": "| Dynamic boost | 1 | 155.231 | 124.544 | 155.231 | 401.70 | 829.49 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-401-70",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "401.70",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33508,
    33577
   ],
   "quote": "| Dynamic boost | 1 | 155.231 | 124.544 | 155.231 | 401.70 | 829.49 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-829-49",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "829.49",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "1 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33508,
    33577
   ],
   "quote": "| Dynamic boost | 1 | 155.231 | 124.544 | 155.231 | 401.70 | 829.49 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-256-955",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "256.955",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33578,
    33647
   ],
   "quote": "| Dynamic boost | 2 | 256.955 | 202.821 | 128.550 | 528.06 | 626.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-202-821",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "202.821",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33578,
    33647
   ],
   "quote": "| Dynamic boost | 2 | 256.955 | 202.821 | 128.550 | 528.06 | 626.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-128-550",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "128.550",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33578,
    33647
   ],
   "quote": "| Dynamic boost | 2 | 256.955 | 202.821 | 128.550 | 528.06 | 626.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-528-06",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "528.06",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33578,
    33647
   ],
   "quote": "| Dynamic boost | 2 | 256.955 | 202.821 | 128.550 | 528.06 | 626.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-626-27",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "626.27",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "2 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33578,
    33647
   ],
   "quote": "| Dynamic boost | 2 | 256.955 | 202.821 | 128.550 | 528.06 | 626.27 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/summed-tokens-s-351-997",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "summed tokens/s",
   "value": "351.997",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33648,
    33716
   ],
   "quote": "| Dynamic boost | 4 | 351.997 | 283.934 | 87.994 | 694.10 | 405.97 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Summed tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/over-the-wall-clock-tokens-s-283-934",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "over the wall clock, tokens/s",
   "value": "283.934",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33648,
    33716
   ],
   "quote": "| Dynamic boost | 4 | 351.997 | 283.934 | 87.994 | 694.10 | 405.97 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Over the wall clock, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/per-stream-tokens-s-87-994",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "per stream, tokens/s",
   "value": "87.994",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33648,
    33716
   ],
   "quote": "| Dynamic boost | 4 | 351.997 | 283.934 | 87.994 | 694.10 | 405.97 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "Per stream"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-694-10",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "694.10",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33648,
    33716
   ],
   "quote": "| Dynamic boost | 4 | 351.997 | 283.934 | 87.994 | 694.10 | 405.97 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-405-97",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "405.97",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W over the arm (n=33 at five-second intervals; 150 W in 30 of them)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "4 streams",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    33648,
    33716
   ],
   "quote": "| Dynamic boost | 4 | 351.997 | 283.934 | 87.994 | 694.10 | 405.97 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 6,
    "header_span": [
     33152,
     33275
    ],
    "header": "| Posture | Streams | Summed tokens/s | Over the wall clock, tokens/s | Per stream | First token, ms | J per 1,000 tokens |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-ten-461-64-ms-p95-493-52-first-token-3",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of ten",
   "value": "461.64 ms (p95 493.52; first token 314.52 ms; 27.81 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one caller",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34396,
    34764
   ],
   "quote": "| A short gate-shaped call on `mistral-small3.2:24b`, one caller, with the model split 56.2 % on the card and 43.8 % in system memory at every posture — wall latency, median of ten | 461.64 ms (p95 493.52; first token 314.52 ms; 27.81 W mean) | 438.98 ms (p95 468.26; first token 298.34 ms; 32.11 W mean) | 430.85 ms (p95 448.20; first token 293.91 ms; 32.37 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Fixed 95 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-ten-438-98-ms-p95-468-26-first-token-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of ten",
   "value": "438.98 ms (p95 468.26; first token 298.34 ms; 32.11 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W over their short stages (n=11 and n=5 at five-second intervals)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one caller",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34396,
    34764
   ],
   "quote": "| A short gate-shaped call on `mistral-small3.2:24b`, one caller, with the model split 56.2 % on the card and 43.8 % in system memory at every posture — wall latency, median of ten | 461.64 ms (p95 493.52; first token 314.52 ms; 27.81 W mean) | 438.98 ms (p95 468.26; first token 298.34 ms; 32.11 W mean) | 430.85 ms (p95 448.20; first token 293.91 ms; 32.37 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Dynamic boost"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-ten-430-85-ms-p95-448-20-first-token-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of ten",
   "value": "430.85 ms (p95 448.20; first token 293.91 ms; 32.37 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "one caller",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34396,
    34764
   ],
   "quote": "| A short gate-shaped call on `mistral-small3.2:24b`, one caller, with the model split 56.2 % on the card and 43.8 % in system memory at every posture — wall latency, median of ten | 461.64 ms (p95 493.52; first token 314.52 ms; 27.81 W mean) | 438.98 ms (p95 468.26; first token 298.34 ms; 32.11 W mean) | 430.85 ms (p95 448.20; first token 293.91 ms; 32.37 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "This machine's boot state"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-eight-928-54-ms-p95-4-671-24-first-tok",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of eight",
   "value": "928.54 ms (p95 4,671.24; first token 465.75 ms; 66.59 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "four callers at once",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34765,
    35009
   ],
   "quote": "| The same, four callers at once — median of eight | 928.54 ms (p95 4,671.24; first token 465.75 ms; 66.59 W mean) | 948.56 ms (p95 4,567.46; first token 569.39 ms; 73.56 W mean) | 885.73 ms (p95 4,452.80; first token 470.94 ms; 70.93 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Fixed 95 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-eight-948-56-ms-p95-4-567-46-first-tok",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of eight",
   "value": "948.56 ms (p95 4,567.46; first token 569.39 ms; 73.56 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W over their short stages (n=11 and n=5 at five-second intervals)",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "four callers at once",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34765,
    35009
   ],
   "quote": "| The same, four callers at once — median of eight | 928.54 ms (p95 4,671.24; first token 465.75 ms; 66.59 W mean) | 948.56 ms (p95 4,567.46; first token 569.39 ms; 73.56 W mean) | 885.73 ms (p95 4,452.80; first token 470.94 ms; 70.93 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Dynamic boost"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/a-gate-shaped-call-wall-latency-median-of-eight-885-73-ms-p95-4-452-80-first-tok",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "a gate-shaped call, wall latency, median of eight",
   "value": "885.73 ms (p95 4,452.80; first token 470.94 ms; 70.93 W mean)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "four callers at once",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    34765,
    35009
   ],
   "quote": "| The same, four callers at once — median of eight | 928.54 ms (p95 4,671.24; first token 465.75 ms; 66.59 W mean) | 948.56 ms (p95 4,567.46; first token 569.39 ms; 73.56 W mean) | 885.73 ms (p95 4,452.80; first token 470.94 ms; 70.93 W mean) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "This machine's boot state"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-embedding-arm-texts-a-second-milliseconds-a-batch-and-j-per-1-000-texts-252-",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the embedding arm — texts a second, milliseconds a batch, and J per 1,000 texts",
   "value": "252.570 texts a second · 253.40 ms a batch · 65.67 J per 1,000 texts",
   "non_figure": null,
   "conditions": {
    "model": "nomic-embed-text",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "batches of 64 texts, 768 dimensions",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35010,
    35282
   ],
   "quote": "| `nomic-embed-text`, batches of 64 texts, 768 dimensions | 252.570 texts a second · 253.40 ms a batch · 65.67 J per 1,000 texts | 268.988 texts a second · 237.93 ms a batch · 83.79 J per 1,000 texts | 250.655 texts a second · 255.33 ms a batch · 90.19 J per 1,000 texts |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Fixed 95 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-embedding-arm-texts-a-second-milliseconds-a-batch-and-j-per-1-000-texts-268-",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the embedding arm — texts a second, milliseconds a batch, and J per 1,000 texts",
   "value": "268.988 texts a second · 237.93 ms a batch · 83.79 J per 1,000 texts",
   "non_figure": null,
   "conditions": {
    "model": "nomic-embed-text",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W over their short stages (n=11 and n=5 at five-second intervals)",
    "context": "not stated",
    "batch_or_steps": "batches of 64 texts, 768 dimensions",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35010,
    35282
   ],
   "quote": "| `nomic-embed-text`, batches of 64 texts, 768 dimensions | 252.570 texts a second · 253.40 ms a batch · 65.67 J per 1,000 texts | 268.988 texts a second · 237.93 ms a batch · 83.79 J per 1,000 texts | 250.655 texts a second · 255.33 ms a batch · 90.19 J per 1,000 texts |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Dynamic boost"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-embedding-arm-texts-a-second-milliseconds-a-batch-and-j-per-1-000-texts-250-",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the embedding arm — texts a second, milliseconds a batch, and J per 1,000 texts",
   "value": "250.655 texts a second · 255.33 ms a batch · 90.19 J per 1,000 texts",
   "non_figure": null,
   "conditions": {
    "model": "nomic-embed-text",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat",
    "context": "not stated",
    "batch_or_steps": "batches of 64 texts, 768 dimensions",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35010,
    35282
   ],
   "quote": "| `nomic-embed-text`, batches of 64 texts, 768 dimensions | 252.570 texts a second · 253.40 ms a batch · 65.67 J per 1,000 texts | 268.988 texts a second · 237.93 ms a batch · 83.79 J per 1,000 texts | 250.655 texts a second · 255.33 ms a batch · 90.19 J per 1,000 texts |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "This machine's boot state"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-card-empty-idle-w-7-18-w-57-samples",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the card empty, idle W",
   "value": "7.18 W (57 samples)",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35283,
    35391
   ],
   "quote": "| The card empty, idle — mean over 30 s | 7.18 W (57 samples) | 7.14 W (57 samples) | 11.57 W (58 samples) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Fixed 95 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-card-empty-idle-w-7-14-w-57-samples",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the card empty, idle W",
   "value": "7.14 W (57 samples)",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "140 to 150 W (n=19)",
    "context": "not stated",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35283,
    35391
   ],
   "quote": "| The card empty, idle — mean over 30 s | 7.18 W (57 samples) | 7.14 W (57 samples) | 11.57 W (58 samples) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Dynamic boost"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/the-card-empty-idle-w-11-57-w-58-samples",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "the card empty, idle W",
   "value": "11.57 W (58 samples)",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "150 W, flat",
    "context": "not stated",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35283,
    35391
   ],
   "quote": "| The card empty, idle — mean over 30 s | 7.18 W (57 samples) | 7.14 W (57 samples) | 11.57 W (58 samples) |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "This machine's boot state"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/held-resident-idle-w-and-what-residency-costs-9-03-w-so-residency-costs-1-85-w",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "held resident, idle W, and what residency costs",
   "value": "9.03 W, so residency costs 1.85 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35392,
    35566
   ],
   "quote": "| `gemma4:26b` held resident at 131,072, idle — mean over 30 s | 9.03 W, so residency costs 1.85 W | 11.34 W, so residency costs 4.20 W | 14.42 W, so residency costs 2.85 W |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Fixed 95 W"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/held-resident-idle-w-and-what-residency-costs-11-34-w-so-residency-costs-4-20-w",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "held resident, idle W, and what residency costs",
   "value": "11.34 W, so residency costs 4.20 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W (n=19)",
    "context": "131,072",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35392,
    35566
   ],
   "quote": "| `gemma4:26b` held resident at 131,072, idle — mean over 30 s | 9.03 W, so residency costs 1.85 W | 11.34 W, so residency costs 4.20 W | 14.42 W, so residency costs 2.85 W |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "Dynamic boost"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/held-resident-idle-w-and-what-residency-costs-14-42-w-so-residency-costs-2-85-w",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "held resident, idle W, and what residency costs",
   "value": "14.42 W, so residency costs 2.85 W",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "not stated",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W, flat",
    "context": "131,072",
    "batch_or_steps": "mean over 30 s",
    "concurrency": "not stated",
    "window": "00:10 to 00:56 UTC on 2026-09-22"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    35392,
    35566
   ],
   "quote": "| `gemma4:26b` held resident at 131,072, idle — mean over 30 s | 9.03 W, so residency costs 1.85 W | 11.34 W, so residency costs 4.20 W | 14.42 W, so residency costs 2.85 W |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     34313,
     34377
    ],
    "header": "| Arm | Fixed 95 W | Dynamic boost | This machine's boot state |",
    "head_cell": "This machine's boot state"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "this machine's boot state: Dynamic Boost running and the boot-time clock lock an operator added to this laptop, 1,200 to 2,550 MHz, read back from the card as in force at every gate — not a state a buyer's machine is in unless they set it. The driver answers `[N/A]` for fan speed on this board, and the page records that an operator, asked afterwards whether the fan was still at maximum for this third pass, was not sure.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-118-424-117-802-119-239",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "118.424 (117.802–119.239)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30100,
    30254
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 118.424 (117.802–119.239) | 372.49 | 79.64 | 672.50 | 42 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-372-49-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "372.49",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30100,
    30254
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 118.424 (117.802–119.239) | 372.49 | 79.64 | 672.50 | 42 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-79-64",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "79.64",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30100,
    30254
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 118.424 (117.802–119.239) | 372.49 | 79.64 | 672.50 | 42 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-672-50-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "672.50",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30100,
    30254
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 118.424 (117.802–119.239) | 372.49 | 79.64 | 672.50 | 42 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-42",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "42",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30100,
    30254
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 118.424 (117.802–119.239) | 372.49 | 79.64 | 672.50 | 42 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-121-379-120-822-123-017",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "121.379 (120.822–123.017)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30255,
    30355
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 121.379 (120.822–123.017) | 435.96 | 76.53 | 622.11 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-435-96",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "435.96",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30255,
    30355
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 121.379 (120.822–123.017) | 435.96 | 76.53 | 622.11 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-76-53",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "76.53",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30255,
    30355
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 121.379 (120.822–123.017) | 435.96 | 76.53 | 622.11 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-622-11",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "622.11",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30255,
    30355
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 121.379 (120.822–123.017) | 435.96 | 76.53 | 622.11 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-48",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "48",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30255,
    30355
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 121.379 (120.822–123.017) | 435.96 | 76.53 | 622.11 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-116-125-115-846-116-716",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "116.125 (115.846–116.716)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30356,
    30457
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 116.125 (115.846–116.716) | 326.27 | 86.87 | 744.29 | 56 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-326-27",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "326.27",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30356,
    30457
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 116.125 (115.846–116.716) | 326.27 | 86.87 | 744.29 | 56 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-86-87",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "86.87",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30356,
    30457
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 116.125 (115.846–116.716) | 326.27 | 86.87 | 744.29 | 56 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-744-29",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "744.29",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30356,
    30457
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 116.125 (115.846–116.716) | 326.27 | 86.87 | 744.29 | 56 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-56",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "56",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30356,
    30457
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 116.125 (115.846–116.716) | 326.27 | 86.87 | 744.29 | 56 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-56-174-56-039-56-374",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "56.174 (56.039–56.374)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30458,
    30562
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 56.174 (56.039–56.374) | 744.92 | 88.38 | 1,573.32 | 64 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-744-92",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "744.92",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30458,
    30562
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 56.174 (56.039–56.374) | 744.92 | 88.38 | 1,573.32 | 64 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-88-38",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "88.38",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30458,
    30562
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 56.174 (56.039–56.374) | 744.92 | 88.38 | 1,573.32 | 64 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-1-573-32",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "1,573.32",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30458,
    30562
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 56.174 (56.039–56.374) | 744.92 | 88.38 | 1,573.32 | 64 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-64",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "64",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30458,
    30562
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 56.174 (56.039–56.374) | 744.92 | 88.38 | 1,573.32 | 64 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-28-851-28-721-28-993",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "28.851 (28.721–28.993)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30717,
    30843
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 28.851 (28.721–28.993) | 214.37 | 93.13 | 3,227.99 | 45 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-214-37",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "214.37",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30717,
    30843
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 28.851 (28.721–28.993) | 214.37 | 93.13 | 3,227.99 | 45 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-93-13",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "93.13",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30717,
    30843
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 28.851 (28.721–28.993) | 214.37 | 93.13 | 3,227.99 | 45 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-3-227-99",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "3,227.99",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30717,
    30843
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 28.851 (28.721–28.993) | 214.37 | 93.13 | 3,227.99 | 45 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-45",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "45",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30717,
    30843
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 28.851 (28.721–28.993) | 214.37 | 93.13 | 3,227.99 | 45 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-86-808-85-795-87-364",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "86.808 (85.795–87.364)",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "measured here on its text path only",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30977,
    31151
   ],
   "quote": "| `minicpm-v4.5`, the vision model the workshop's products use to read pictures, measured here on its text path only | 86.808 (85.795–87.364) | 149.86 | 79.27 | 913.16 | 41 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-149-86",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "149.86",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "measured here on its text path only",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30977,
    31151
   ],
   "quote": "| `minicpm-v4.5`, the vision model the workshop's products use to read pictures, measured here on its text path only | 86.808 (85.795–87.364) | 149.86 | 79.27 | 913.16 | 41 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-79-27",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "79.27",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "measured here on its text path only",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30977,
    31151
   ],
   "quote": "| `minicpm-v4.5`, the vision model the workshop's products use to read pictures, measured here on its text path only | 86.808 (85.795–87.364) | 149.86 | 79.27 | 913.16 | 41 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-913-16",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "913.16",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "measured here on its text path only",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30977,
    31151
   ],
   "quote": "| `minicpm-v4.5`, the vision model the workshop's products use to read pictures, measured here on its text path only | 86.808 (85.795–87.364) | 149.86 | 79.27 | 913.16 | 41 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-41",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "41",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "95 W",
    "context": "not stated",
    "batch_or_steps": "measured here on its text path only",
    "concurrency": "one parallel slot",
    "window": "18:43 to 19:34 UTC on 2026-09-21"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    30977,
    31151
   ],
   "quote": "| `minicpm-v4.5`, the vision model the workshop's products use to read pictures, measured here on its text path only | 86.808 (85.795–87.364) | 149.86 | 79.27 | 913.16 | 41 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     29967,
     30073
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the fixed posture: Dynamic Boost stopped and the board enforcing its own default, 95 W of a possible 175, with nothing on this board able to set a limit at all — the limit was read back from the card at the open and the close of every stage and at 2 Hz inside every scored run, and a change would have voided the stage. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-155-251-155-128-155-593",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "155.251 (155.128–155.593)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31484,
    31639
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 155.251 (155.128–155.593) | 372.92 | 122.57 | 787.76 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-372-92-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "372.92",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31484,
    31639
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 155.251 (155.128–155.593) | 372.92 | 122.57 | 787.76 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-122-57",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "122.57",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31484,
    31639
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 155.251 (155.128–155.593) | 372.92 | 122.57 | 787.76 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-787-76-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "787.76",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31484,
    31639
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 155.251 (155.128–155.593) | 372.92 | 122.57 | 787.76 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": {
    "slug": "a-laptop-asked-the-desktops-questions",
    "cite": "the-tables"
   },
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-44",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "44",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q8_0` (the default)",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31484,
    31639
   ],
   "quote": "| `gemma4:26b`, KV cache `q8_0` (the default), 131,072 — the ladder's own rung, for reference | 155.251 (155.128–155.593) | 372.92 | 122.57 | 787.76 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page labels this row *the ladder's own rung, for reference*: the comparison table at the top of this same section prints the same reading with the posture's own cap in its cell. Two printings of one measurement are two rows on this roster, and each names the other's section — here its own, since both printings sit inside this page's one cited section. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-165-656-165-545-166-175",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "165.656 (165.545–166.175)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W in every one of the 24 samples of its own 2 Hz trace",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31640,
    31741
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 165.656 (165.545–166.175) | 439.78 | 116.01 | 700.31 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-439-78",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "439.78",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W in every one of the 24 samples of its own 2 Hz trace",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31640,
    31741
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 165.656 (165.545–166.175) | 439.78 | 116.01 | 700.31 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-116-01",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "116.01",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W in every one of the 24 samples of its own 2 Hz trace",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31640,
    31741
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 165.656 (165.545–166.175) | 439.78 | 116.01 | 700.31 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-700-31",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "700.31",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W in every one of the 24 samples of its own 2 Hz trace",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31640,
    31741
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 165.656 (165.545–166.175) | 439.78 | 116.01 | 700.31 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-44-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "44",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `f16`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "150 W in every one of the 24 samples of its own 2 Hz trace",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31640,
    31741
   ],
   "quote": "| `gemma4:26b`, KV cache `f16`, 131,072 | 165.656 (165.545–166.175) | 439.78 | 116.01 | 700.31 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance.",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-155-712-155-641-155-908",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "155.712 (155.641–155.908)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31742,
    31844
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 155.712 (155.641–155.908) | 336.23 | 148.71 | 955.03 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-336-23",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "336.23",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31742,
    31844
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 155.712 (155.641–155.908) | 336.23 | 148.71 | 955.03 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-148-71",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "148.71",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31742,
    31844
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 155.712 (155.641–155.908) | 336.23 | 148.71 | 955.03 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-955-03",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "955.03",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31742,
    31844
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 155.712 (155.641–155.908) | 336.23 | 148.71 | 955.03 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-44-3",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "44",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "KV cache `q4_0`",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31742,
    31844
   ],
   "quote": "| `gemma4:26b`, KV cache `q4_0`, 131,072 | 155.712 (155.641–155.908) | 336.23 | 148.71 | 955.03 | 44 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-72-159-72-024-72-171",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "72.159 (72.024–72.171)",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31845,
    31950
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 72.159 (72.024–72.171) | 721.95 | 123.99 | 1,718.01 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-721-95",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "721.95",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31845,
    31950
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 72.159 (72.024–72.171) | 721.95 | 123.99 | 1,718.01 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-123-99",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "123.99",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31845,
    31950
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 72.159 (72.024–72.171) | 721.95 | 123.99 | 1,718.01 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-1-718-01",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "1,718.01",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31845,
    31950
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 72.159 (72.024–72.171) | 721.95 | 123.99 | 1,718.01 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-48-2",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "48",
   "non_figure": null,
   "conditions": {
    "model": "gemma4:26b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "the window filled to 131,072",
    "batch_or_steps": "not stated",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31845,
    31950
   ],
   "quote": "| `gemma4:26b`, the window filled to 131,072 | 72.159 (72.024–72.171) | 721.95 | 123.99 | 1,718.01 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-44-946-44-917-44-963",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "44.946 (44.917–44.963)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31951,
    32078
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 44.946 (44.917–44.963) | 208.97 | 140.48 | 3,124.35 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-208-97",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "208.97",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31951,
    32078
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 44.946 (44.917–44.963) | 208.97 | 140.48 | 3,124.35 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-140-48",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "140.48",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31951,
    32078
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 44.946 (44.917–44.963) | 208.97 | 140.48 | 3,124.35 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-3-124-35",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "3,124.35",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31951,
    32078
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 44.946 (44.917–44.963) | 208.97 | 140.48 | 3,124.35 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-48-3",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "48",
   "non_figure": null,
   "conditions": {
    "model": "mistral-small3.2:24b",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "98,304",
    "batch_or_steps": "every layer forced onto the card",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    31951,
    32078
   ],
   "quote": "| `mistral-small3.2:24b`, every layer forced onto the card, 98,304 | 44.946 (44.917–44.963) | 208.97 | 140.48 | 3,124.35 | 48 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/tokens-s-spread-of-three-121-733-121-722-121-762",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "tokens/s (spread of three)",
   "value": "121.733 (121.722–121.762)",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "not stated",
    "batch_or_steps": "text path only",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    32212,
    32326
   ],
   "quote": "| `minicpm-v4.5`, the vision model, text path only | 121.733 (121.722–121.762) | 132.94 | 144.74 | 1,188.72 | 43 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 1,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Tokens/s (spread of three)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/first-token-ms-132-94",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "132.94",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "not stated",
    "batch_or_steps": "text path only",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    32212,
    32326
   ],
   "quote": "| `minicpm-v4.5`, the vision model, text path only | 121.733 (121.722–121.762) | 132.94 | 144.74 | 1,188.72 | 43 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 2,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "First token, ms"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/board-w-mean-144-74",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "board W (mean)",
   "value": "144.74",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "not stated",
    "batch_or_steps": "text path only",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    32212,
    32326
   ],
   "quote": "| `minicpm-v4.5`, the vision model, text path only | 121.733 (121.722–121.762) | 132.94 | 144.74 | 1,188.72 | 43 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 3,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Board W (mean)"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/j-per-1-000-tokens-1-188-72",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "J per 1,000 tokens",
   "value": "1,188.72",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "not stated",
    "batch_or_steps": "text path only",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    32212,
    32326
   ],
   "quote": "| `minicpm-v4.5`, the vision model, text path only | 121.733 (121.722–121.762) | 132.94 | 144.74 | 1,188.72 | 43 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 4,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "J per 1,000 tokens"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/core-max-c-43",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "core max °C",
   "value": "43",
   "non_figure": null,
   "conditions": {
    "model": "minicpm-v4.5",
    "quant": "a `q8_0` KV cache",
    "runtime": "ollama",
    "runtime_version": "0.32.13",
    "cap": "140 to 150 W",
    "context": "not stated",
    "batch_or_steps": "text path only",
    "concurrency": "one parallel slot",
    "window": "21:01 to 21:50 UTC the same day"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "the-tables",
    "md_sha": "349f66e7b939391611c37e45301b248b54c2967c6fe0d8fc9f44a2d5fa9e94d7"
   },
   "cite": "the-tables",
   "span": [
    32212,
    32326
   ],
   "quote": "| `minicpm-v4.5`, the vision model, text path only | 121.733 (121.722–121.762) | 132.94 | 144.74 | 1,188.72 | 43 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": {
    "column": 5,
    "header_span": [
     31351,
     31457
    ],
    "header": "| Arm | Tokens/s (spread of three) | First token, ms | Board W (mean) | J per 1,000 tokens | Core max °C |",
    "head_cell": "Core max °C"
   },
   "conditions_cites": [
    "how-it-was-measured"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the dynamic-boost posture, the state a laptop like this arrives in: Dynamic Boost running, no limit set because none can be, and the board floating its budget between its 95 W default and its 175 W maximum against the processor's draw — so the cap cell is the budget this pass sampled and not a setting anybody made. The driver answers `[N/A]` for fan speed on this board, so no result file carries a fan column; an operator set the laptop's fan to maximum by hand before the first pass, and the page asks that every temperature be read as a temperature under that condition and every figure that might have throttled on an ordinary fan curve as one that did not get the chance. the page's own witness for this arm: *the other arms touched 140 W*, against the `f16` arm's flat 150 W",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/a-laptop-asked-the-desktops-questions/price-the-whole-machine-this-part-is-soldered-into-new-3-099-99",
   "identity": "rtx-5090-laptop",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "price, the whole machine this part is soldered into, new",
   "value": "$3,099.99",
   "non_figure": null,
   "conditions": {
    "model": "not stated",
    "quant": "not stated",
    "runtime": "not stated",
    "runtime_version": "not stated",
    "cap": "not stated",
    "context": "not stated",
    "batch_or_steps": "not stated",
    "concurrency": "not stated",
    "window": "ordered 2026-07-03"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "a-laptop-asked-the-desktops-questions",
    "section": "what-this-page-cannot-say",
    "md_sha": "270fee03a59f8176eb8013867744d9282b28f80e0423eba21ef6a092a00e7c55"
   },
   "cite": "what-this-page-cannot-say",
   "span": [
    38820,
    38915
   ],
   "quote": "The whole machine was bought new from Micro Center for $3,099.99 before tax, ordered 2026-07-03",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop",
   "identity_cite": "the-board-and-the-two-things-it-will-not-do",
   "identity_span": [
    6897,
    6967
   ],
   "cell": null,
   "conditions_cites": [],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "before tax, from Micro Center, and it is the price of the whole laptop rather than of this board: the page's own sentence is that this part *is soldered into a laptop and is not sold as a card*, and that *a buyer is weighing that whole machine against a card plus the rest of a computer*. The same section is where the page rules that no reading from this bench may fill a cell on this shelf's comparison page, its price tables included",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-tokens-a-second-scored-78-126",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second (scored)",
   "value": "78.126",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9848,
    9902
   ],
   "quote": "| Decode, tokens a second (scored) | 78.126 | 85.216 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-tokens-a-second-scored-85-216",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second (scored)",
   "value": "85.216",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9848,
    9902
   ],
   "quote": "| Decode, tokens a second (scored) | 78.126 | 85.216 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/spread-of-the-ten-runs-0-50",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "spread of the ten runs",
   "value": "0.50 %",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9903,
    9947
   ],
   "quote": "| Spread of the ten runs | 0.50 % | 0.41 % |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/spread-of-the-ten-runs-0-41",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "spread of the ten runs",
   "value": "0.41 %",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9903,
    9947
   ],
   "quote": "| Spread of the ten runs | 0.50 % | 0.41 % |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-tokens-a-second-the-server-s-own-clock-78-118",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second (the server's own clock)",
   "value": "78.118",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9948,
    10018
   ],
   "quote": "| Decode, tokens a second (the server's own clock) | 78.118 | 85.553 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-tokens-a-second-the-server-s-own-clock-85-553",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "decode, tokens a second (the server's own clock)",
   "value": "85.553",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    9948,
    10018
   ],
   "quote": "| Decode, tokens a second (the server's own clock) | 78.118 | 85.553 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/first-token-ms-recorded-not-scored-226-0",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms (recorded, not scored)",
   "value": "226.0",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10019,
    10079
   ],
   "quote": "| First token, ms (recorded, not scored) | 226.0 | 1,018.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words: *what this reading is not: a first-token comparison.* Ollama's 1,018 ms is about 761 ms of its own load step, on a model already loaded; the page's first-token comparison is arm C's at one request. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/first-token-ms-recorded-not-scored-1-018-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms (recorded, not scored)",
   "value": "1,018.1",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10019,
    10079
   ],
   "quote": "| First token, ms (recorded, not scored) | 226.0 | 1,018.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own words: *what this reading is not: a first-token comparison.* Ollama's 1,018 ms is about 761 ms of its own load step, on a model already loaded; the page's first-token comparison is arm C's at one request. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/of-which-ollama-s-load-step-ms-761-1",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "of which Ollama's load step, ms",
   "value": "761.1",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10080,
    10127
   ],
   "quote": "| Of which Ollama's load step, ms | — | 761.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/total-speed-of-one-answer-tokens-a-second-send-to-last-token-73-364",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "total speed of one answer, tokens a second (send to last token)",
   "value": "73.364",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10128,
    10213
   ],
   "quote": "| Total speed of one answer, tokens a second (send to last token) | 73.364 | 63.865 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/total-speed-of-one-answer-tokens-a-second-send-to-last-token-63-865",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "total speed of one answer, tokens a second (send to last token)",
   "value": "63.865",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10128,
    10213
   ],
   "quote": "| Total speed of one answer, tokens a second (send to last token) | 73.364 | 63.865 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/board-power-while-busy-w-299-14",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board power while busy, W",
   "value": "299.14",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10214,
    10261
   ],
   "quote": "| Board power while busy, W | 299.14 | 289.88 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/board-power-while-busy-w-289-88",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board power while busy, W",
   "value": "289.88",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10214,
    10261
   ],
   "quote": "| Board power while busy, W | 299.14 | 289.88 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/joules-per-1-000-tokens-busy-watts-decode-speed-arithmetic-not-registered-3-829",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens, busy watts ÷ decode speed (arithmetic, not registered)",
   "value": "3,829",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.27.1",
    "cap": "300 W",
    "context": "a 4,352-token window",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · our kit's own serve line",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10262,
    10361
   ],
   "quote": "| Joules per 1,000 tokens, busy watts ÷ decode speed (arithmetic, not registered) | 3,829 | 3,402 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 1,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "vLLM 0.27.1"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/joules-per-1-000-tokens-busy-watts-decode-speed-arithmetic-not-registered-3-402",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens, busy watts ÷ decode speed (arithmetic, not registered)",
   "value": "3,402",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "One 536-token prompt (`series-p512`), 256 tokens out · `num_batch: 512`",
    "concurrency": "one request at a time",
    "window": "2026-09-28 from 11:29:59 to 11:33:35 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "one-request-at-a-time",
    "md_sha": "39c7c6317e2392cbee3b763d420fc3b6a209c6fbcac283acc934259f2a90b29b"
   },
   "cite": "one-request-at-a-time",
   "span": [
    10262,
    10361
   ],
   "quote": "| Joules per 1,000 tokens, busy watts ÷ decode speed (arithmetic, not registered) | 3,829 | 3,402 |",
   "identity_basis": "page",
   "identity_token": "RTX 3090",
   "identity_cite": "what-was-written-down-before-the-first-run",
   "identity_span": [
    43016,
    43083
   ],
   "cell": {
    "column": 2,
    "header_span": [
     9768,
     9831
    ],
    "header": "| Median of ten runs | vLLM 0.27.1 | Ollama 0.32.13, 16 slots |",
    "head_cell": "Ollama 0.32.13, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-212-999",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "212.999",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24345,
    24370
   ],
   "quote": "| 1 | 212.999 | 441.866 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "at one request a run holds one request, so its 95th percentile is that request's wait (the page's caption). answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-441-866",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "441.866",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "not stated",
    "concurrency": "1",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24345,
    24370
   ],
   "quote": "| 1 | 212.999 | 441.866 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "at one request a run holds one request, so its 95th percentile is that request's wait (the page's caption). answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-417-849",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "417.849",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "not stated",
    "concurrency": "2",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24371,
    24398
   ],
   "quote": "| 2 | 417.849 | 1,068.577 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-1-068-577",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "1,068.577",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "not stated",
    "concurrency": "2",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24371,
    24398
   ],
   "quote": "| 2 | 417.849 | 1,068.577 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-822-939",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "822.939",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "not stated",
    "concurrency": "4",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24399,
    24426
   ],
   "quote": "| 4 | 822.939 | 2,244.823 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-2-244-823",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "2,244.823",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "not stated",
    "concurrency": "4",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24399,
    24426
   ],
   "quote": "| 4 | 822.939 | 2,244.823 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-1-641-068",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "1,641.068",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "not stated",
    "concurrency": "8",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24427,
    24456
   ],
   "quote": "| 8 | 1,641.068 | 5,365.722 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-5-365-722",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "5,365.722",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "not stated",
    "concurrency": "8",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24427,
    24456
   ],
   "quote": "| 8 | 1,641.068 | 5,365.722 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-3-333-177",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "3,333.177",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24457,
    24487
   ],
   "quote": "| 16 | 3,333.177 | 20,935.24 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/the-wait-for-the-first-token-95th-percentile-ms-20-935-24",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "the wait for the first token, 95th percentile, ms",
   "value": "20,935.24",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "not stated",
    "concurrency": "16",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    24457,
    24487
   ],
   "quote": "| 16 | 3,333.177 | 20,935.24 |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24276,
     24327
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-82-693",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "82.693",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step",
    "concurrency": "1",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23515,
    23547
   ],
   "quote": "| 1 | 82.693 | 85.007 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-85-007",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "85.007",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · `num_batch: 512`",
    "concurrency": "1",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23515,
    23547
   ],
   "quote": "| 1 | 82.693 | 85.007 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-79-677",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "79.677",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step",
    "concurrency": "2",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23548,
    23579
   ],
   "quote": "| 2 | 79.677 | 76.43 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-76-43",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "76.43",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · `num_batch: 512`",
    "concurrency": "2",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23548,
    23579
   ],
   "quote": "| 2 | 79.677 | 76.43 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-72-183",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "72.183",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step",
    "concurrency": "4",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23580,
    23612
   ],
   "quote": "| 4 | 72.183 | 51.713 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-51-713",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "51.713",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · `num_batch: 512`",
    "concurrency": "4",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23580,
    23612
   ],
   "quote": "| 4 | 72.183 | 51.713 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's caption: at four and at eight the sixteen-slot runner's three scored runs *decoded at two speeds (52.026, 51.713 and 32.831 tokens a second at four; 23.816, 32.485 and 23.814 at eight), and each median above is one of them*. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-58-778",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "58.778",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step",
    "concurrency": "8",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23613,
    23645
   ],
   "quote": "| 8 | 58.778 | 23.816 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-23-816",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "23.816",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · `num_batch: 512`",
    "concurrency": "8",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23613,
    23645
   ],
   "quote": "| 8 | 58.778 | 23.816 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's caption: at four and at eight the sixteen-slot runner's three scored runs *decoded at two speeds (52.026, 51.713 and 32.831 tokens a second at four; 23.816, 32.485 and 23.814 at eight), and each median above is one of them*. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-41-428",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "41.428",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`google/gemma-4-12B-it-qat-w4a16-ct` at revision `1d2c2d7f`",
    "runtime": "vLLM",
    "runtime_version": "0.30.0",
    "cap": "300 W",
    "context": "a 5,120-token window",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step",
    "concurrency": "16",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23646,
    23679
   ],
   "quote": "| 16 | 41.428 | 24.449 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 1,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "vLLM 0.30.0"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-3090/ollama-and-vllm-on-one-rtx-3090/decode-per-request-tokens-a-second-24-449",
   "identity": "rtx-3090",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "decode per request, tokens a second",
   "value": "24.449",
   "non_figure": null,
   "conditions": {
    "model": "Gemma 4 12B",
    "quant": "`gemma4:12b-it-qat`, a 4-bit GGUF",
    "runtime": "Ollama",
    "runtime_version": "0.34.4",
    "cap": "300 W",
    "context": "16 slots of 5,120 tokens",
    "batch_or_steps": "the 536-token prompt with each request's seeded tag, 256 tokens out per request · `num_batch: 512`",
    "concurrency": "16",
    "window": "2026-10-01 (UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "ollama-and-vllm-on-one-rtx-3090",
    "section": "requests-at-once",
    "md_sha": "e4eea73ba93f75e4e82f383e5475832feb82d550094bb5a31793ed71059844ee"
   },
   "cite": "requests-at-once",
   "span": [
    23646,
    23679
   ],
   "quote": "| 16 | 41.428 | 24.449 | Differ |",
   "identity_basis": "section",
   "identity_token": "GeForce RTX 3090 XC3 Ultra",
   "identity_cite": "requests-at-once",
   "identity_span": [
    13973,
    14097
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23432,
     23493
    ],
    "header": "| At once | vLLM 0.30.0 | Ollama 0.34.4, 16 slots | Answers |",
    "head_cell": "Ollama 0.34.4, 16 slots"
   },
   "conditions_cites": [
    "what-was-written-down-before-the-first-run",
    "two-names-one-card",
    "what-each-holds"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6)",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-19-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "19.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39217,
    39279
   ],
   "quote": "| 4K (4,038 tokens) | 19.1 | 7.6 | 1.9 | 2.5× | 10.0× | 4.0× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 512 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-7-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "7.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39217,
    39279
   ],
   "quote": "| 4K (4,038 tokens) | 19.1 | 7.6 | 1.9 | 2.5× | 10.0× | 4.0× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-1-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "1.9",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39217,
    39279
   ],
   "quote": "| 4K (4,038 tokens) | 19.1 | 7.6 | 1.9 | 2.5× | 10.0× | 4.0× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "The Strata engine (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-125-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "125.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39280,
    39347
   ],
   "quote": "| 32K (32,038 tokens) | 125.6 | 50.0 | 12.1 | 2.5× | 10.4× | 4.1× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 512 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-50-0",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "50.0",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39280,
    39347
   ],
   "quote": "| 32K (32,038 tokens) | 125.6 | 50.0 | 12.1 | 2.5× | 10.4× | 4.1× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-12-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "12.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39280,
    39347
   ],
   "quote": "| 32K (32,038 tokens) | 125.6 | 50.0 | 12.1 | 2.5× | 10.4× | 4.1× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "The Strata engine (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-545-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "545.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39348,
    39418
   ],
   "quote": "| 128K (128,037 tokens) | 545.1 | 204.4 | 47.6 | 2.7× | 11.5× | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 512 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-204-4",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "204.4",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39348,
    39418
   ],
   "quote": "| 128K (128,037 tokens) | 545.1 | 204.4 | 47.6 | 2.7× | 11.5× | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-47-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "47.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    39348,
    39418
   ],
   "quote": "| 128K (128,037 tokens) | 545.1 | 204.4 | 47.6 | 2.7× | 11.5× | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     39027,
     39180
    ],
    "header": "| Prompt length | llama.cpp, 512 (s) | llama.cpp, 4,096 (s) | The Strata engine (s) | 512 ÷ 4,096 | The Strata engine's lead at 512 | Its lead at 4,096 |",
    "head_cell": "The Strata engine (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-211-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "211.7",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40323,
    40355
   ],
   "quote": "| 4K | 211.7 | 533.7 | 2,147.5 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 512 (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-533-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "533.7",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40323,
    40355
   ],
   "quote": "| 4K | 211.7 | 533.7 | 2,147.5 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-147-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,147.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40323,
    40355
   ],
   "quote": "| 4K | 211.7 | 533.7 | 2,147.5 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-255-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "255.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40356,
    40389
   ],
   "quote": "| 32K | 255.1 | 640.9 | 2,655.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 512 (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-640-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "640.9",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40356,
    40389
   ],
   "quote": "| 32K | 255.1 | 640.9 | 2,655.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-655-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,655.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40356,
    40389
   ],
   "quote": "| 32K | 255.1 | 640.9 | 2,655.1 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-234-9",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "234.9",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40390,
    40424
   ],
   "quote": "| 128K | 234.9 | 626.7 | 2,705.0 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 512 (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-626-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "626.7",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40390,
    40424
   ],
   "quote": "| 128K | 234.9 | 626.7 | 2,705.0 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-705-0",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,705.0",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-first-word-the-laptop-the-iq3-s-file",
    "md_sha": "147665344665639d735df0a1bc09ce72d5b76828931d9d2f64949ff2afa0d303"
   },
   "cite": "the-first-word-the-laptop-the-iq3-s-file",
   "span": [
    40390,
    40424
   ],
   "quote": "| 128K | 234.9 | 626.7 | 2,705.0 |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     40149,
     40301
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp, 512 (tok/s, engine's clock) | Prompt rate, llama.cpp, 4,096 (tok/s) | Prompt rate, the Strata engine (tok/s) |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-24-4",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "24.4",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45268,
    45309
   ],
   "quote": "| 4K | 24.4 | 22.5 | 85.6 | 3.5× | 3.8× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 512 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-22-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "22.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45268,
    45309
   ],
   "quote": "| 4K | 24.4 | 22.5 | 85.6 | 3.5× | 3.8× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-85-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "85.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45268,
    45309
   ],
   "quote": "| 4K | 24.4 | 22.5 | 85.6 | 3.5× | 3.8× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "The Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-20-7",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "20.7",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45310,
    45352
   ],
   "quote": "| 32K | 20.7 | 19.4 | 87.1 | 4.2× | 4.5× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 512 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-19-4",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "19.4",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45310,
    45352
   ],
   "quote": "| 32K | 20.7 | 19.4 | 87.1 | 4.2× | 4.5× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-87-1",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "87.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45310,
    45352
   ],
   "quote": "| 32K | 20.7 | 19.4 | 87.1 | 4.2× | 4.5× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "The Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-14-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "14.2",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 512-token micro-batch (its default) · 8 threads · 34 of 48 layers' experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-02T23:03:15Z → 23:40:55Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45353,
    45396
   ],
   "quote": "| 128K | 14.2 | 13.0 | 69.8 | 4.9× | 5.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 512 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-13-0",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "13.0",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · a 4,096-token micro-batch · 8 threads · 42 layers in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-05T14:43:47Z → 14:59:47Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45353,
    45396
   ],
   "quote": "| 128K | 14.2 | 13.0 | 69.8 | 4.9× | 5.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "llama.cpp, 4,096 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's control, registered after the first readings: a control, no band; measured while two of our audit jobs ran on the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-69-8",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "69.8",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · reading 8,192 tokens at a time · about a third of the experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-03T00:46:41Z → 00:50:45Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "writing-the-answer-the-laptop-the-iq3-s-file",
    "md_sha": "1a9042845db5b8f3380b519e87f722a6fa907f02df375415474fd5b577919632"
   },
   "cite": "writing-the-answer-the-laptop-the-iq3-s-file",
   "span": [
    45353,
    45396
   ],
   "quote": "| 128K | 14.2 | 13.0 | 69.8 | 4.9× | 5.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 3,
    "header_span": [
     45077,
     45236
    ],
    "header": "| Prompt length | llama.cpp, 512 (tok/s) | llama.cpp, 4,096 (tok/s) | The Strata engine (tok/s) | The Strata engine ÷ llama.cpp at 512 | ÷ llama.cpp at 4,096 |",
    "head_cell": "The Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "measured while another job's dry run shared the laptop",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-9-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "9.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51571,
    51652
   ],
   "quote": "| 4K | 9.5 | OUT-OF-BAND (opened: no) | — | 32.5 | OUT-OF-BAND (opened: no) | — |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "First word, llama.cpp (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-32-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "32.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51571,
    51652
   ],
   "quote": "| 4K | 9.5 | OUT-OF-BAND (opened: no) | — | 32.5 | OUT-OF-BAND (opened: no) | — |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 4,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "Writing, llama.cpp (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-73-3",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "73.3",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51653,
    51703
   ],
   "quote": "| 32K | 73.3 | 11.5 | 6.4× | 26.6 | 115.2 | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "First word, llama.cpp (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-11-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "11.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · reading 8,192 tokens at a time",
    "concurrency": "one request at a time",
    "window": "2026-10-03T05:09:47Z → 05:13:25Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51653,
    51703
   ],
   "quote": "| 32K | 73.3 | 11.5 | 6.4× | 26.6 | 115.2 | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "First word, the Strata engine (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-26-6",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "26.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51653,
    51703
   ],
   "quote": "| 32K | 73.3 | 11.5 | 6.4× | 26.6 | 115.2 | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 4,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "Writing, llama.cpp (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-115-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "115.2",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · reading 8,192 tokens at a time",
    "concurrency": "one request at a time",
    "window": "2026-10-03T05:09:47Z → 05:13:25Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51653,
    51703
   ],
   "quote": "| 32K | 73.3 | 11.5 | 6.4× | 26.6 | 115.2 | 4.3× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 5,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "Writing, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-321-8",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "321.8",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51704,
    51756
   ],
   "quote": "| 128K | 321.8 | 45.2 | 7.1× | 17.5 | 102.5 | 5.9× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 1,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "First word, llama.cpp (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/first-word-seconds-45-2",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "45.2",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · reading 8,192 tokens at a time",
    "concurrency": "one request at a time",
    "window": "2026-10-03T05:09:47Z → 05:13:25Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51704,
    51756
   ],
   "quote": "| 128K | 321.8 | 45.2 | 7.1× | 17.5 | 102.5 | 5.9× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 2,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "First word, the Strata engine (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-17-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "17.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 8 threads · 24 of 48 layers of experts in system memory",
    "concurrency": "one request at a time",
    "window": "2026-10-03T06:07:35Z → 06:31:09Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51704,
    51756
   ],
   "quote": "| 128K | 321.8 | 45.2 | 7.1× | 17.5 | 102.5 | 5.9× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 4,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "Writing, llama.cpp (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the clean re-read, the figure of record; llama.cpp's memory limit reached, as the page's caption says, and the two llama.cpp attempts agree within about 6 % on every figure",
   "record_published": null
  },
  {
   "id": "rtx-5090-laptop/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-102-5",
   "identity": "rtx-5090-laptop",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "102.5",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "Q2_0",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.25",
    "cap": "a limit of 135-150 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · reading 8,192 tokens at a time",
    "concurrency": "one request at a time",
    "window": "2026-10-03T05:09:47Z → 05:13:25Z"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-q2-0-pair-the-laptop",
    "md_sha": "3d9b8d4eb4937a6db0b05d53f93353a39dc0131121dc56115dfdefec73a1d4ce"
   },
   "cite": "the-q2-0-pair-the-laptop",
   "span": [
    51704,
    51756
   ],
   "quote": "| 128K | 321.8 | 45.2 | 7.1× | 17.5 | 102.5 | 5.9× |",
   "identity_basis": "page",
   "identity_token": "RTX 5090 Laptop GPU (24 GB)",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    29890,
    29936
   ],
   "cell": {
    "column": 5,
    "header_span": [
     51323,
     51534
    ],
    "header": "| Prompt length | First word, llama.cpp (s) | First word, the Strata engine (s) | llama.cpp ÷ the Strata engine | Writing, llama.cpp (tok/s) | Writing, the Strata engine (tok/s) | The Strata engine ÷ llama.cpp |",
    "head_cell": "Writing, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-102-889",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "102.889",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59065,
    59163
   ],
   "quote": "| 4K | 102.889 | 205.263 | +102.374 | 32.475 | Beyond unmeasured night-to-night variation | 2.0× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, 2026-10-06 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-205-263",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "205.263",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59065,
    59163
   ],
   "quote": "| 4K | 102.889 | 205.263 | +102.374 | 32.475 | Beyond unmeasured night-to-night variation | 2.0× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 2026-10-07 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-76-337",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "76.337",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59164,
    59276
   ],
   "quote": "| 32K (exploratory) | 76.337 | 189.155 | +112.818 | 45.300 | Beyond unmeasured night-to-night variation | 2.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, 2026-10-06 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-189-155",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "189.155",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59164,
    59276
   ],
   "quote": "| 32K (exploratory) | 76.337 | 189.155 | +112.818 | 45.300 | Beyond unmeasured night-to-night variation | 2.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 2026-10-07 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-38-576",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "38.576",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59277,
    59390
   ],
   "quote": "| 128K (exploratory) | 38.576 | 172.325 | +133.749 | 44.019 | Beyond unmeasured night-to-night variation | 4.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, 2026-10-06 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-172-325",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "172.325",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    59277,
    59390
   ],
   "quote": "| 128K (exploratory) | 38.576 | 172.325 | +133.749 | 44.019 | Beyond unmeasured night-to-night variation | 4.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     58858,
     59029
    ],
    "header": "| Prompt length | llama.cpp, 2026-10-06 (tok/s) | The Strata engine 0.1.39, 2026-10-07 (tok/s) | Difference (tok/s) | The cross-night threshold (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 2026-10-07 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-2-52",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "2.52",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61382,
    61473
   ],
   "quote": "| 4K | 2.52 | 1.06 | 1.456 | 3.033 | Within unmeasured night-to-night variation; no ratio |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, llama.cpp, 2026-10-06 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-1-06",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "1.06",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61382,
    61473
   ],
   "quote": "| 4K | 2.52 | 1.06 | 1.456 | 3.033 | Within unmeasured night-to-night variation; no ratio |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, the Strata engine, 2026-10-07 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-20-91",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "20.91",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61474,
    61542
   ],
   "quote": "| 32K (exploratory) | 20.91 | 6.37 | 14.535 | 5.703 | Beyond; 3.3× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, llama.cpp, 2026-10-06 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-6-37",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "6.37",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61474,
    61542
   ],
   "quote": "| 32K (exploratory) | 20.91 | 6.37 | 14.535 | 5.703 | Beyond; 3.3× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, the Strata engine, 2026-10-07 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-94-22",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "94.22",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61543,
    61614
   ],
   "quote": "| 128K (exploratory) | 94.22 | 26.63 | 67.593 | 16.772 | Beyond; 3.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, llama.cpp, 2026-10-06 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/first-word-seconds-26-63",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "26.63",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    61543,
    61614
   ],
   "quote": "| 128K (exploratory) | 94.22 | 26.63 | 67.593 | 16.772 | Beyond; 3.5× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     61203,
     61351
    ],
    "header": "| Prompt length | First word, llama.cpp, 2026-10-06 (s) | First word, the Strata engine, 2026-10-07 (s) | Difference (s) | Threshold (s) | Verdict |",
    "head_cell": "First word, the Strata engine, 2026-10-07 (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-1-619-4",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "1,619.4",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62349,
    62480
   ],
   "quote": "| 4K | 1,619.4 | 3,920.0 | Within unmeasured night-to-night variation (difference 2,300.6 under a threshold of 3,097.5); no ratio |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, llama.cpp (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-3-920-0",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "3,920.0",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62349,
    62480
   ],
   "quote": "| 4K | 1,619.4 | 3,920.0 | Within unmeasured night-to-night variation (difference 2,300.6 under a threshold of 3,097.5); no ratio |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-1-539-1",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "1,539.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62481,
    62537
   ],
   "quote": "| 32K (exploratory) | 1,539.1 | 5,143.9 | Beyond; 3.3× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, llama.cpp (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-5-143-9",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "5,143.9",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62481,
    62537
   ],
   "quote": "| 32K (exploratory) | 1,539.1 | 5,143.9 | Beyond; 3.3× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-1-362-8",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "1,362.8",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · every layer on the card · 8 threads · a 512-token micro-batch",
    "concurrency": "one request at a time",
    "window": "2026-10-06"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62538,
    62595
   ],
   "quote": "| 128K (exploratory) | 1,362.8 | 4,908.1 | Beyond; 3.6× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 1,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, llama.cpp (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-pro-6000-blackwell/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-4-908-1",
   "identity": "rtx-pro-6000-blackwell",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "4,908.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · all 24,576 experts on the card · the prompt path reading 8,192 tokens at a time · the draft head on",
    "concurrency": "one request at a time",
    "window": "2026-10-07"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-desktop-card-one-engine-a-night",
    "md_sha": "49b998359cff94c8bc7d125a228d058034b0c66b21b2dcc9402f3d8bb9654e55"
   },
   "cite": "the-desktop-card-one-engine-a-night",
   "span": [
    62538,
    62595
   ],
   "quote": "| 128K (exploratory) | 1,362.8 | 4,908.1 | Beyond; 3.6× |",
   "identity_basis": "page",
   "identity_token": "RTX PRO 6000 Blackwell Workstation Edition",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    30916,
    30977
   ],
   "cell": {
    "column": 2,
    "header_span": [
     62211,
     62328
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp (tok/s, engine's clock) | Prompt rate, the Strata engine (tok/s) | Verdict |",
    "head_cell": "Prompt rate, the Strata engine (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-22-490",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "22.490",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71005,
    71086
   ],
   "quote": "| 4K | 22.490 | 98.291 | +75.801 | 31.989 | Beyond the registered margin | 4.4× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-98-291",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "98.291",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71005,
    71086
   ],
   "quote": "| 4K | 22.490 | 98.291 | +75.801 | 31.989 | Beyond the registered margin | 4.4× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-19-620",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "19.620",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71087,
    71183
   ],
   "quote": "| 32K (exploratory) | 19.620 | 92.571 | +72.951 | 12.375 | Beyond the registered margin | 4.7× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-92-571",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "92.571",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71087,
    71183
   ],
   "quote": "| 32K (exploratory) | 19.620 | 92.571 | +72.951 | 12.375 | Beyond the registered margin | 4.7× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-13-993",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "13.993",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71184,
    71281
   ],
   "quote": "| 128K (exploratory) | 13.993 | 81.452 | +67.459 | 28.434 | Beyond the registered margin | 5.8× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-81-452",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "81.452",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    71184,
    71281
   ],
   "quote": "| 128K (exploratory) | 13.993 | 81.452 | +67.459 | 28.434 | Beyond the registered margin | 5.8× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     70717,
     70969
    ],
    "header": "| Prompt length | llama.cpp, `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (tok/s) | The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s) | Difference (tok/s) | The registered margin (tok/s) | Verdict | Ratio |",
    "head_cell": "The Strata engine 0.1.39, 8,144 of 24,576 experts on the card (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-16-87",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "16.87",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76607,
    76674
   ],
   "quote": "| 4K | 16.87 | 1.92 | 15.84 | Beyond the registered margin | 8.8× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-1-92",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "1.92",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76607,
    76674
   ],
   "quote": "| 4K | 16.87 | 1.92 | 15.84 | Beyond the registered margin | 8.8× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-15-84",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "15.84",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76607,
    76674
   ],
   "quote": "| 4K | 16.87 | 1.92 | 15.84 | Beyond the registered margin | 8.8× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp b11514, no draft head, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-138-02",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "138.02",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76675,
    76761
   ],
   "quote": "| 32K (exploratory) | 138.02 | 12.14 | 130.90 | Beyond the registered margin | 11.4× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-12-14",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "12.14",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76675,
    76761
   ],
   "quote": "| 32K (exploratory) | 138.02 | 12.14 | 130.90 | Beyond the registered margin | 11.4× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-130-90",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "130.90",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76675,
    76761
   ],
   "quote": "| 32K (exploratory) | 138.02 | 12.14 | 130.90 | Beyond the registered margin | 11.4× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp b11514, no draft head, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-566-57",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "566.57",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76762,
    76849
   ],
   "quote": "| 128K (exploratory) | 566.57 | 51.05 | 528.72 | Beyond the registered margin | 11.1× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-51-05",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "51.05",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76762,
    76849
   ],
   "quote": "| 128K (exploratory) | 566.57 | 51.05 | 528.72 | Beyond the registered margin | 11.1× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/first-word-seconds-528-72",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "first word, seconds",
   "value": "528.72",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    76762,
    76849
   ],
   "quote": "| 128K (exploratory) | 566.57 | 51.05 | 528.72 | Beyond the registered margin | 11.1× |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     76284,
     76576
    ],
    "header": "| Prompt length | First word, llama.cpp `3cf03257`, 34 of 48 layers' experts in system memory, 8 of 64 threads (s) | First word, the Strata engine 0.1.39, 8,144 of 24,576 experts on the card (s) | First word, llama.cpp b11514, no draft head, 8 of 64 threads (s) | The pair's verdict | Ratio |",
    "head_cell": "First word, llama.cpp b11514, no draft head, 8 of 64 threads (s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-239-7",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "239.7",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78007,
    78070
   ],
   "quote": "| 4K | 239.7 | 2,143.1 | 255.2 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-143-1",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,143.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78007,
    78070
   ],
   "quote": "| 4K | 239.7 | 2,143.1 | 255.2 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, the Strata engine 0.1.39 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-255-2",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "255.2",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78007,
    78070
   ],
   "quote": "| 4K | 239.7 | 2,143.1 | 255.2 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp b11514, no draft head (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-232-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "232.3",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78071,
    78149
   ],
   "quote": "| 32K (exploratory) | 232.3 | 2,671.6 | 244.9 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-671-6",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,671.6",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78071,
    78149
   ],
   "quote": "| 32K (exploratory) | 232.3 | 2,671.6 | 244.9 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, the Strata engine 0.1.39 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-244-9",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "244.9",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78071,
    78149
   ],
   "quote": "| 32K (exploratory) | 232.3 | 2,671.6 | 244.9 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp b11514, no draft head (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-226-1",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "226.1",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "3cf03257",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78150,
    78229
   ],
   "quote": "| 128K (exploratory) | 226.1 | 2,535.0 | 242.3 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 1,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-2-535-0",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "2,535.0",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "the Strata engine",
    "runtime_version": "0.1.39",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · 8,144 of 24,576 experts on the card",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78150,
    78229
   ],
   "quote": "| 128K (exploratory) | 226.1 | 2,535.0 | 242.3 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 2,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, the Strata engine 0.1.39 (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/prompt-rate-tokens-a-second-the-engine-s-clock-242-3",
   "identity": "rtx-3090-ti",
   "road": "page-table",
   "kind": "figure",
   "metric": "prompt rate, tokens a second, the engine's clock",
   "value": "242.3",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    78150,
    78229
   ],
   "quote": "| 128K (exploratory) | 226.1 | 2,535.0 | 242.3 | Beyond the registered margin |",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": {
    "column": 3,
    "header_span": [
     77780,
     77981
    ],
    "header": "| Prompt length | Prompt rate, llama.cpp `3cf03257` (tok/s, engine's clock) | Prompt rate, the Strata engine 0.1.39 (tok/s) | Prompt rate, llama.cpp b11514, no draft head (tok/s) | The pair's verdict |",
    "head_cell": "Prompt rate, llama.cpp b11514, no draft head (tok/s)"
   },
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-22-657",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "22.657",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "4K (4,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    74601,
    74683
   ],
   "quote": "llama.cpp b11514 wrote 22.657 / 21.257 / 17.681 tokens a second at 4K / 32K / 128K",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the first figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-21-257",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "21.257",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "32K (32,038 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    74601,
    74683
   ],
   "quote": "llama.cpp b11514 wrote 22.657 / 21.257 / 17.681 tokens a second at 4K / 32K / 128K",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the second figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090-ti/flash-next-why-the-first-word-comes-sooner/writing-speed-tokens-a-second-17-681",
   "identity": "rtx-3090-ti",
   "road": "hand-carried",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "17.681",
   "non_figure": null,
   "conditions": {
    "model": "Qwen3.8-Flash-Next",
    "quant": "IQ3_S",
    "runtime": "llama.cpp",
    "runtime_version": "b11514",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "128K (128,037 tokens) · no draft head · 8 of 64 threads",
    "concurrency": "one request at a time",
    "window": "2026-10-09 (09:32Z to 10:51Z, UTC)"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "flash-next-why-the-first-word-comes-sooner",
    "section": "the-24-gb-desktop-card-both-engines-in-one-window",
    "md_sha": "63fbc31c9ed596addc03d765d8942ce2521286e5232214540009abfab3009b64"
   },
   "cite": "the-24-gb-desktop-card-both-engines-in-one-window",
   "span": [
    74601,
    74683
   ],
   "quote": "llama.cpp b11514 wrote 22.657 / 21.257 / 17.681 tokens a second at 4K / 32K / 128K",
   "identity_basis": "page",
   "identity_token": "GeForce RTX 3090 Ti",
   "identity_cite": "the-question-and-the-instruments",
   "identity_span": [
    31932,
    31987
   ],
   "cell": null,
   "conditions_cites": [
    "the-question-and-the-instruments",
    "what-it-does-not-say",
    "the-first-word-the-laptop-the-iq3-s-file"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the third figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-60-545-0-033",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "60.545 (0.033)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 4,096, tokens/s (run range, tokens/s)"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/first-token-ms-548-22",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "548.22",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "First token at 4,096, ms"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/board-w-259-9",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W",
   "value": "259.9",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 4,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Board W at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/joules-per-1-000-tokens-4-293-87",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "4,293.87",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 5,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Joules per 1,000 tokens at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-60-41",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "60.41",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "32,768",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 6,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 32,768, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/largest-window-held-whole-tokens-98-304",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, tokens",
   "value": "98,304",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24008,
    24086
   ],
   "quote": "| Q8_0 | 12.13 | 60.545 (0.033) | 548.22 | 259.9 | 4,293.87 | 60.41 | 98,304 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 7,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Largest window held whole, tokens"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the ladder stops at the first window that does not fit whole; *topped out* is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-65-724-0-441",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "65.724 (0.441)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 4,096, tokens/s (run range, tokens/s)"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/first-token-ms-518-25",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "518.25",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "First token at 4,096, ms"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/board-w-270-81",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W",
   "value": "270.81",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 4,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Board W at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/joules-per-1-000-tokens-4-102-0",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "4,102.0",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 5,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Joules per 1,000 tokens at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-64-83",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "64.83",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "32,768",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 6,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 32,768, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/largest-window-held-whole-tokens-131-072-topped-out",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, tokens",
   "value": "131,072, topped out",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24087,
    24177
   ],
   "quote": "| Q6_K | 9.37 | 65.724 (0.441) | 518.25 | 270.81 | 4,102.0 | 64.83 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 7,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Largest window held whole, tokens"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the ladder stops at the first window that does not fit whole; *topped out* is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-78-942-0-068",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "78.942 (0.068)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 4,096, tokens/s (run range, tokens/s)"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/first-token-ms-516-45",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "516.45",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "First token at 4,096, ms"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/board-w-264-11",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W",
   "value": "264.11",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 4,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Board W at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/joules-per-1-000-tokens-3-345-62",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "3,345.62",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 5,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Joules per 1,000 tokens at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-78-158",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "78.158",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "32,768",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 6,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 32,768, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/largest-window-held-whole-tokens-131-072-topped-out-2",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, tokens",
   "value": "131,072, topped out",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24178,
    24272
   ],
   "quote": "| Q5_K_M | 8.13 | 78.942 (0.068) | 516.45 | 264.11 | 3,345.62 | 78.158 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 7,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Largest window held whole, tokens"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the ladder stops at the first window that does not fit whole; *topped out* is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-88-203-0-252",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "88.203 (0.252)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 4,096, tokens/s (run range, tokens/s)"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/first-token-ms-481-8",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "481.8",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "First token at 4,096, ms"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/board-w-259-65",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W",
   "value": "259.65",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 4,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Board W at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/joules-per-1-000-tokens-2-948-31",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "2,948.31",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 5,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Joules per 1,000 tokens at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-87-234",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "87.234",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "32,768",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 6,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 32,768, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/largest-window-held-whole-tokens-131-072-topped-out-3",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, tokens",
   "value": "131,072, topped out",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24273,
    24366
   ],
   "quote": "| Q4_K_M | 6.96 | 88.203 (0.252) | 481.8 | 259.65 | 2,948.31 | 87.234 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 7,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Largest window held whole, tokens"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the ladder stops at the first window that does not fit whole; *topped out* is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-74-894-0-398",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "74.894 (0.398)",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 4,096, tokens/s (run range, tokens/s)"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/first-token-ms-510-56",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "first token, ms",
   "value": "510.56",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "First token at 4,096, ms"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/board-w-257-59",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "board W",
   "value": "257.59",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 4,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Board W at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/joules-per-1-000-tokens-3-439-41",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "joules per 1,000 tokens",
   "value": "3,439.41",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "4,096",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 5,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Joules per 1,000 tokens at 4,096"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/writing-speed-tokens-a-second-74-135",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "writing speed, tokens a second",
   "value": "74.135",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "32,768",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 6,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Writing speed at 32,768, tokens/s"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/largest-window-held-whole-tokens-131-072-topped-out-4",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "largest window held whole, tokens",
   "value": "131,072, topped out",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 8 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "not stated",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    24367,
    24461
   ],
   "quote": "| Q3_K_M | 5.67 | 74.894 (0.398) | 510.56 | 257.59 | 3,439.41 | 74.135 | 131,072, topped out |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 7,
    "header_span": [
     23731,
     23966
    ],
    "header": "| Level | Download, GiB | Writing speed at 4,096, tokens/s (run range, tokens/s) | First token at 4,096, ms | Board W at 4,096 | Joules per 1,000 tokens at 4,096 | Writing speed at 32,768, tokens/s | Largest window held whole, tokens |",
    "head_cell": "Largest window held whole, tokens"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the ladder stops at the first window that does not fit whole; *topped out* is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/card-memory-growth-at-the-4-096-token-question-load-mib-12-795",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card memory growth at the 4,096-token question load, MiB",
   "value": "12,795",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25132,
    25164
   ],
   "quote": "| Q8_0 | 12,419 | 12,795 | 376 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Card memory growth at the 4,096-token question load, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/growth-minus-download-mib-376",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "growth minus download, MiB",
   "value": "376",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q8_0 · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25132,
    25164
   ],
   "quote": "| Q8_0 | 12,419 | 12,795 | 376 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Growth minus download, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own subtraction of the file's size from the card's growth, printed in its table",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/card-memory-growth-at-the-4-096-token-question-load-mib-10-121",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card memory growth at the 4,096-token question load, MiB",
   "value": "10,121",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25165,
    25196
   ],
   "quote": "| Q6_K | 9,590 | 10,121 | 531 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Card memory growth at the 4,096-token question load, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/growth-minus-download-mib-531",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "growth minus download, MiB",
   "value": "531",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q6_K · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25165,
    25196
   ],
   "quote": "| Q6_K | 9,590 | 10,121 | 531 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Growth minus download, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own subtraction of the file's size from the card's growth, printed in its table",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/card-memory-growth-at-the-4-096-token-question-load-mib-8-939",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card memory growth at the 4,096-token question load, MiB",
   "value": "8,939",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25197,
    25229
   ],
   "quote": "| Q5_K_M | 8,323 | 8,939 | 616 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Card memory growth at the 4,096-token question load, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/growth-minus-download-mib-616",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "growth minus download, MiB",
   "value": "616",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q5_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25197,
    25229
   ],
   "quote": "| Q5_K_M | 8,323 | 8,939 | 616 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Growth minus download, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own subtraction of the file's size from the card's growth, printed in its table",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/card-memory-growth-at-the-4-096-token-question-load-mib-7-827",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card memory growth at the 4,096-token question load, MiB",
   "value": "7,827",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25230,
    25262
   ],
   "quote": "| Q4_K_M | 7,131 | 7,827 | 696 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Card memory growth at the 4,096-token question load, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/growth-minus-download-mib-696",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "growth minus download, MiB",
   "value": "696",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q4_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25230,
    25262
   ],
   "quote": "| Q4_K_M | 7,131 | 7,827 | 696 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Growth minus download, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own subtraction of the file's size from the card's growth, printed in its table",
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/card-memory-growth-at-the-4-096-token-question-load-mib-6-583",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "card memory growth at the 4,096-token question load, MiB",
   "value": "6,583",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25263,
    25295
   ],
   "quote": "| Q3_K_M | 5,801 | 6,583 | 782 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 2,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Card memory growth at the 4,096-token question load, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": null,
   "record_published": null
  },
  {
   "id": "rtx-3090/the-quantization-ladder/growth-minus-download-mib-782",
   "identity": "rtx-3090",
   "road": "page-table",
   "kind": "figure",
   "metric": "growth minus download, MiB",
   "value": "782",
   "non_figure": null,
   "conditions": {
    "model": "mistral-nemo:12b-instruct-2407",
    "quant": "Q3_K_M · cache at 16 bits",
    "runtime": "Ollama",
    "runtime_version": "0.32.13",
    "cap": "300 W",
    "context": "the 4,096-token question load",
    "batch_or_steps": "519 tokens as this model counts it · asks for 256 tokens",
    "concurrency": "one request at a time",
    "window": "2026-10-15 between 05:34 and 11:01 UTC"
   },
   "figure_definition": "v3",
   "status": "published",
   "superseded_by": null,
   "source": {
    "kind": "page",
    "slug": "the-quantization-ladder",
    "section": "speed-and-memory",
    "md_sha": "5d85deae2497e8b04ad8c4f54baf78c95f6918828099f3f8b8f43ae2660f9ad1"
   },
   "cite": "speed-and-memory",
   "span": [
    25263,
    25295
   ],
   "quote": "| Q3_K_M | 5,801 | 6,583 | 782 |",
   "identity_basis": "section",
   "identity_token": "RTX 3090",
   "identity_cite": "speed-and-memory",
   "identity_span": [
    24464,
    24556
   ],
   "cell": {
    "column": 3,
    "header_span": [
     24997,
     25110
    ],
    "header": "| Level | Download, MiB | Card memory growth at the 4,096-token question load, MiB | Growth minus download, MiB |",
    "head_cell": "Growth minus download, MiB"
   },
   "conditions_cites": [
    "how-it-was-measured",
    "the-five-files"
   ],
   "same_reading_as": null,
   "later_reading_of": null,
   "note": "the page's own subtraction of the file's size from the card's growth, printed in its table",
   "record_published": null
  }
 ]
}