{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "the-same-eighteen-pictures-card-by-card",
 "title": "The same eighteen pictures, card by card",
 "dek": "Held to the same 300 W, at 2048×1280 on the print lab's own picture model, each card in its own computer, an RTX PRO 6000 Blackwell Workstation Edition 96 GB does each sampling step 1.6 times as fast as a ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB, down from 2.8 times at their stock 600 W and 360 W. Over ten minutes at 1024×1024 on the same model at 8 steps, a GeForce RTX 5090 Laptop 24 GB, on the 150 W its Dynamic Boost held through that run, spent the same board energy per picture as the workstation card at 300 W, 663 against 667 J, each in its own computer. On a GeForce RTX 3080 Ti Founders Edition 12 GB, the smallest of the three picture models sat whole on the card in every cell and the other two paged their weights in every cell that ran, down to 68.7 per cent resident. In one computer, on one checkout of the picture software, both held to 300 W on a 512×512 recipe, a 5080 did a 3090 Ti's work in 0.69 of the time for 0.64 of the board energy; at their stock limits, against the second 3090 Ti in that computer, 0.74 of the time for 0.53 of the energy.",
 "published": "2026-09-26",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/the-same-eighteen-pictures-card-by-card/",
  "md_url": "https://research.strata2signal.com/the-same-eighteen-pictures-card-by-card/index.md",
  "md_sha": "9c059975b94ed606de6acd3a369832614ca8add22e0f4df055481654e3d5eff5",
  "html_sha": "24799f123cb52db1e9471a4ff6f3200a08dddda019ccbbd06a85d9a040add786",
  "anchors_sha": "67cc45d78323d6520ecf211ed09af50e3297271019670f7c68b0ea07559ca36b"
 },
 "short": {
  "paragraph": "Four cards ran the same eighteen kinds of picture on one ComfyUI stack in four computers: every cross-card figure is \"same software, different computer\". One FLUX.2 [klein] 4B picture at 1024×1024 and 8 steps, at stock: 1.51 s on the workstation card, 3.59 on the 5080, 4.50 on the laptop, 5.73 on the 3080 Ti. On the 12 GB card klein 4B in bf16 sat whole in every cell; Z-Image-Turbo in bf16 and FLUX.1 [dev] in fp8 paged, down to 68.7 per cent resident. Ten minutes of that klein 4B picture, board energy only: 252 at 667 J from the workstation card at 300 W, 164 at 977.6 to 991.6 J from the 5080 at 300 W or stock, 132 at 663 J from the laptop at 150 W, 101 at 1,908.4 J from the 3080 Ti at its stock 350 W beside an idle board. In one computer at 300 W a 5080 did a 3090 Ti's work in 0.69 of the time for 0.64 of the energy.",
  "from_dek": false,
  "counts": {
   "words": 6511,
   "minutes": 30,
   "tables": 10,
   "kit": true
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "who-ran-where",
   "heading": "Who ran where",
   "level": 2,
   "span": [
    2856,
    5177
   ],
   "chunks": [
    [
     2856,
     5177
    ]
   ],
   "chars": 2321,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "words-this-page-leans-on",
   "heading": "Words this page leans on",
   "level": 2,
   "span": [
    5177,
    7710
   ],
   "chunks": [
    [
     5177,
     7710
    ]
   ],
   "chars": 2533,
   "digest": "this section defines terms used in the article, such as a sampling step, seconds per step, a cell, the ninety-second rule, board power, joules per picture, and stock, cap and posture. it also explains the label same software, different computer.",
   "digest_skipped": null
  },
  {
   "id": "the-grid-and-the-one-thing-it-holds-fixed",
   "heading": "The grid, and the one thing it holds fixed",
   "level": 2,
   "span": [
    7710,
    9102
   ],
   "chunks": [
    [
     7710,
     9102
    ]
   ],
   "chars": 1392,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "seven-things-the-grid-found",
   "heading": "Seven things the grid found",
   "level": 2,
   "span": [
    9102,
    24931
   ],
   "chunks": [
    [
     9102,
     24931
    ]
   ],
   "chars": 15829,
   "digest": "the grid findings cover model residency on a 12 GB card, seconds per sampling step at various power settings, joules per step, ten-minute drawing performance, the impact of a 300 W cap, scaling costs between image sizes, and the ninety-second rule.",
   "digest_skipped": null
  },
  {
   "id": "three-cards-in-one-computer-one-checkout-one-recipe",
   "heading": "Three cards in one computer, one checkout, one recipe",
   "level": 2,
   "span": [
    24931,
    28301
   ],
   "chunks": [
    [
     24931,
     28301
    ]
   ],
   "chars": 3370,
   "digest": "three cards, including a ZOTAC GeForce RTX 5080 Solid CORE OC 16 GB and two 3090 Tis, took turns in the same x16 slot of a bench computer. the test used an 8-step recipe at 300 W and at stock.",
   "digest_skipped": null
  },
  {
   "id": "the-tables",
   "heading": "The full tables",
   "level": 2,
   "span": [
    28301,
    29027
   ],
   "chunks": [
    [
     28301,
     29027
    ]
   ],
   "chars": 726,
   "digest": "the full tables, containing twenty-nine tables with various headings and conditions, are located in the data kit.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-cannot-say",
   "heading": "What this page cannot say",
   "level": 2,
   "span": [
    29027,
    31552
   ],
   "chunks": [
    [
     29027,
     31552
    ]
   ],
   "chars": 2525,
   "digest": "this section clarifies the limitations of the data, noting that ratios between grid cards include the computer, and that certain factors like power at the wall socket or room temperature were not measured.",
   "digest_skipped": null
  },
  {
   "id": "how-it-was-measured",
   "heading": "How it was measured",
   "level": 2,
   "span": [
    31552,
    35768
   ],
   "chunks": [
    [
     31552,
     35768
    ]
   ],
   "chars": 4216,
   "digest": "the measurement process involved a harness on each computer, a fixed ComfyUI stack, and specific model workflows. power, temperature, and clocks were read twice a second, and the order of operations for each card is described.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    35768,
    37500
   ],
   "chunks": [
    [
     35768,
     37500
    ]
   ],
   "chars": 1732,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "more-on-this-site",
   "heading": "More on this site",
   "level": 2,
   "span": [
    37500,
    38288
   ],
   "chunks": [
    [
     37500,
     38288
    ]
   ],
   "chars": 788,
   "digest": "this section provides links to related content on the site, including a history of Black Forest Labs, comparisons of different cards, and information on the hardware roster.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    38288,
    39682
   ],
   "chunks": [
    [
     38288,
     39682
    ]
   ],
   "chars": 1394,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-794353be",
   "q": "how are the different hardware setups identified?",
   "a": "The article identifies the specific cards, computers, processors, system memory, and driver versions used for each test in a detailed table.",
   "cites": [
    "who-ran-where"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-eade9d2a",
   "q": "what definitions govern the timing and power metrics?",
   "a": "The text defines terms such as sampling steps, seconds per step, board power, joules per picture, and the ninety-second rule used to manage long renders.",
   "cites": [
    "words-this-page-leans-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-20797d3e",
   "q": "what happens when desktop cards are limited to 300 w?",
   "a": "Limiting cards to 300 W increases the time taken for steps on the workstation card by 1.69 to 1.73 times, while the 5080 sees minimal changes.",
   "cites": [
    "seven-things-the-grid-found"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a774a26a",
   "q": "how much energy is consumed per sampling step?",
   "a": "The energy cost varies by card and model; for example, at 2048×1280, the RTX PRO 6000 at 600 W uses 210.5 J per klein step.",
   "cites": [
    "seven-things-the-grid-found"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-654f51f4",
   "q": "what results came from testing three cards in one slot?",
   "a": "Testing three different cards in the same x16 slot showed the 5080 completed the 8-step recipe in 0.69 of the EVGA 3090 Ti's time at 300 W.",
   "cites": [
    "three-cards-in-one-computer-one-checkout-one-recipe"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "a-laptop-asked-the-desktops-questions",
   "why": "shares ground with § What this page cannot say · § The tables"
  },
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § What this page cannot say · § What this page does not measure"
  },
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § What this page cannot say · § What this page does not measure"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe cards are named in full in Who ran where and Three cards in one computer; the laptop is an **MSI Raider 16 Max HX**. **NVIDIA** for the silicon, the driver every reading came from and the Founders Edition board; **ZOTAC**, **EVGA** and **MSI** for the boards, and **Gigabyte** for one in the kit's tables. The pictures stand on **ComfyUI** and the **comfy-kitchen** and **comfy-aimdo** libraries beneath it, on **PyTorch** and **NVIDIA's CUDA**, on **Astral's uv** and on **Network UPS Tools**. The models are Black Forest Labs' **FLUX.2 [klein] 4B** and **FLUX.1 [dev]**, Alibaba's Tongyi Lab's **Z-Image-Turbo**, Alibaba Cloud's **Qwen 3 4B**, and Google's **T5-XXL** and OpenAI's **CLIP ViT-L/14**, the text encoders inside the FLUX.1 [dev] checkpoint; **Hugging Face** hosted every pinned file. None of them owed us anything.\n\nA small human team owns the cards, ruled the grid and its refusals before a picture was drawn, pasted every cap and signed the numbers; a fleet of AI agents built the harness, ran it and did the arithmetic under that team's rulings. Thanks to the people who ask, on every forum this workshop reads, what a picture costs on the card they have.\n\n*Measured on 2026-09-24 and 2026-09-25 by one pinned stack on four computers and one instrument in the fourth, with the label on every cross-computer figure.*",
 "kit": {
  "url": "https://research.strata2signal.com/the-same-eighteen-pictures-card-by-card/data/",
  "licence": "CC BY 4.0",
  "files": [
   "TABLES.md"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 9,
  "digests_kept": 7,
  "chips_written": 8,
  "chips_kept": 5,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 1,
   "profanity": 0,
   "identifier": 2,
   "bare_figure": 0,
   "house_voice": 0,
   "hand-edit": 4
  },
  "new_noun_tokens_checked": [
   "0.69",
   "1.69",
   "1.73",
   "103",
   "12",
   "138",
   "150",
   "16",
   "2048×1280",
   "210.5",
   "24",
   "265.9",
   "300",
   "3080",
   "3090",
   "3216c62e",
   "360",
   "5080",
   "5090",
   "600",
   "6000",
   "8-step",
   "83",
   "96",
   "Black",
   "C",
   "CORE",
   "ComfyUI",
   "EVGA",
   "Edition",
   "FE",
   "Forest",
   "Founders",
   "GB",
   "GPU",
   "GeForce",
   "J",
   "Labs",
   "Laptop",
   "MiB",
   "NVIDIA",
   "OC",
   "PRO",
   "RTX",
   "Solid",
   "Ti",
   "Ti's",
   "Tis",
   "W",
   "Workstation",
   "ZOTAC",
   "v0.34.0",
   "x16"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 300,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "a 3080 Ti FE reached a peak temperature of 83 °C",
    "detail": "fe is in the cited section only as code or in a table cell"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "a 5080 at 360 W had a mean draw of 265.9 W",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "a 5090 Laptop GPU 24 GB had a power budget of 138 to 150 W",
    "detail": "gpu is in the cited section only as code or in a table cell"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 171,
    "text_sha": "2d2ae153ef3a",
    "section": "the-grid-and-the-one-thing-it-holds-fixed",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the digest says the ComfyUI stack is held fixed across all tests; the page holds it fixed across the grid's cards, and the render bench is a second test",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the digest says the ComfyUI stack is held fixed across all tests; the page holds it fixed across the grid's cards, and the render bench is a second test. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 211,
    "text_sha": "2d11148f3625",
    "chip_id": "c-0c211acf",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the chip says the article cannot provide ratios between grid cards; the page gives them, each labelled same software, different computer",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the chip says the article cannot provide ratios between grid cards; the page gives them, each labelled same software, different computer. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 199,
    "text_sha": "07e8439ed097",
    "chip_id": "c-8a87c03b",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the chip states paging as what larger pictures do on any card; the page finds paging on the 12 GB card, and every model sat whole on the cards with 16 GB or more",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the chip states paging as what larger pictures do on any card; the page finds paging on the 12 GB card, and every model sat whole on the cards with 16 GB or more. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 184,
    "text_sha": "a95fcad4c414",
    "chip_id": "c-24a1af91",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the chip says the stack is held fixed for all tests; the page holds it fixed across the grid's cards only",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the chip says the stack is held fixed for all tests; the page holds it fixed across the grid's cards only. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-09-25T22:46:25Z"
}