{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "one-3090-out-one-3090-ti-in",
 "title": "One 3090 out, one 3090 Ti in, and the cap that decided it",
 "dek": "On 2026-09-25 (UTC) an EVGA GeForce RTX 3090 XC3 Ultra Gaming left the workstation, the computer behind the workshop's public services, and an EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming took its place at 350 W, not the old card's 300, because at 300 W our earlier bench had it behind a 3090. On the same computer and test, nine days apart in a different slot, it writes text about 9.5 per cent faster; its first live replies, eight long ones, agree. Whether it runs cooler here is not settled: its first hour read warmer, on a fresh boot.",
 "published": "2026-09-25",
 "series": [
  "notes"
 ],
 "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/one-3090-out-one-3090-ti-in/",
  "md_url": "https://research.strata2signal.com/one-3090-out-one-3090-ti-in/index.md",
  "md_sha": "c2dd13e4c62678c1816f8e7ca8ca169ab6b7f1e398f3e52871fa8dc2dd9c1631",
  "html_sha": "8ff1b3e603a51dd7b0362570cd7f744b41dc3b2c9cb1e866218075811adf10a4",
  "anchors_sha": "a95aa31fde925ff90531c862cda519768e90f11c0958b4df65554336444aae37"
 },
 "short": {
  "paragraph": "The workstation's small language model, mistral-small3.2:24b, moved from a 3090 at 300 W to a 3090 Ti at 350 W, and the cap is the story: held to 300 W, two-thirds of what it ships with, this 3090 Ti was the slower card on that model; at 350 W it led. In service, 350 W against 300, it is about 9.5 per cent faster (8.1 to 11.2 across the runs) on about the same energy per thousand tokens, for 7 to 13 per cent more board power. Beyond that one dense model it is the more efficient card: it idles lower (18.74 W against 22.26 on the render computer, a model loaded), draws less and finishes sooner when drawing pictures (316.2 W against 325.8, 9.1 per cent less time per image), and spends 10.1 per cent less energy on the mixture model. \"Runs much cooler\" held while drawing pictures on the render computer, the workshop's picture machine, against a used 2021 3090, and has not shown here.",
  "from_dek": false,
  "counts": {
   "words": 3186,
   "minutes": 14,
   "tables": 2,
   "kit": false
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "why-the-swap",
   "heading": "Why the swap",
   "level": 2,
   "span": [
    3060,
    3939
   ],
   "chunks": [
    [
     3060,
     3939
    ]
   ],
   "chars": 879,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-the-earlier-bench-said",
   "heading": "What the earlier bench said",
   "level": 2,
   "span": [
    3939,
    5305
   ],
   "chunks": [
    [
     3939,
     5305
    ]
   ],
   "chars": 1366,
   "digest": "render computer files suggested a 300 W swap would be a downgrade. at 300 W, a 3090 Ti ran 42.969 tokens a second, while a 3090 ran 47.838. at 350 W, the Ti ran 53.354, using 8.1 per cent less energy than at 300 W.",
   "digest_skipped": null
  },
  {
   "id": "the-same-computer-the-same-script",
   "heading": "The same computer, the same script",
   "level": 2,
   "span": [
    5305,
    11113
   ],
   "chunks": [
    [
     5305,
     11113
    ]
   ],
   "chars": 5808,
   "digest": "both cards ran the workshop's cap ladder on the workstation using one program and one frozen 1,023-token prompt. the 3090 Ti's runs spread 0.1 to 0.8 per cent at each cap, while the old card's 300 W baseline spreads 2.1 per cent.",
   "digest_skipped": null
  },
  {
   "id": "what-these-cards-serve",
   "heading": "What these cards serve",
   "level": 2,
   "span": [
    11113,
    13347
   ],
   "chunks": [
    [
     11113,
     13347
    ]
   ],
   "chars": 2234,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-seat-live",
   "heading": "The seat, live",
   "level": 2,
   "span": [
    13347,
    14726
   ],
   "chunks": [
    [
     13347,
     14726
    ]
   ],
   "chars": 1379,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-swap-minute-by-minute-utc",
   "heading": "The swap, minute by minute (UTC)",
   "level": 2,
   "span": [
    14726,
    15215
   ],
   "chunks": [
    [
     14726,
     15215
    ]
   ],
   "chars": 489,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "cooler",
   "heading": "Cooler?",
   "level": 2,
   "span": [
    15215,
    19412
   ],
   "chunks": [
    [
     15215,
     19412
    ]
   ],
   "chars": 4197,
   "digest": "a 3090 Ti peaked at 60 °C during ten minutes of continuous drawing at 350 W, while a 3090 peaked at 74 °C. at idle on the render computer, a 3090 Ti drew 18.74 W with a model loaded against a 3090's 22.26 W.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-cannot-say",
   "heading": "What this page cannot say",
   "level": 2,
   "span": [
    19412,
    20151
   ],
   "chunks": [
    [
     19412,
     20151
    ]
   ],
   "chars": 739,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    20151,
    20861
   ],
   "chunks": [
    [
     20151,
     20861
    ]
   ],
   "chars": 710,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    20861,
    21778
   ],
   "chunks": [
    [
     20861,
     21778
    ]
   ],
   "chars": 917,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-4384c905",
   "q": "what motivated the hardware replacement?",
   "a": "An operator proposed the swap based on expectations that a 3090 Ti runs much cooler and runs dense models over 10 per cent faster than a 3090.",
   "cites": [
    "why-the-swap"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-d8f1e909",
   "q": "how did the earlier benchmark compare performance at 300 w?",
   "a": "At 300 W, the 3090 Ti ran the dense model at 42.969 tokens a second, while the 3090 ran it at 47.838 tokens a second, representing a loss of 10.2 per cent.",
   "cites": [
    "what-the-earlier-bench-said"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-677b2006",
   "q": "what variables differed during the workstation testing?",
   "a": "Differences included the day and hour, the slot, the presence of an RTX PRO 6000, device-selection settings, and a safety guard polling the battery backup and the card.",
   "cites": [
    "the-same-computer-the-same-script"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-6d21676a",
   "q": "how are the two cards utilized in current services?",
   "a": "The 3090 Ti handles short calls and screens questions using a dense model, while the RTX PRO 6000 serves as the writer using Gemma 4 and a reranker.",
   "cites": [
    "what-these-cards-serve"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-22918c4b",
   "q": "how long was the workstation offline during the swap?",
   "a": "The workstation was down for between eleven and thirteen minutes, with the power-off occurring after 09:36:05 and the boot at 09:49:15.",
   "cites": [
    "the-swap-minute-by-minute-utc"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-fa1630e3",
   "q": "what limitations exist regarding the temperature data?",
   "a": "The article does not report memory temperature, only the chip's core, and lacks a same-day comparison or a sustained load measurement on the workstation.",
   "cites": [
    "what-this-page-cannot-say"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "two-3090s-out-one-3090-ti-in",
   "why": "shares ground with § Cooler? · § Why this card, and why 350 W"
  },
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § What the earlier bench said · § The setting this page would choose, and the dissent"
  },
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § Cooler? · § Renders: one 3080 against one 3090, seconds per image"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\n**EVGA** built the three boards and **NVIDIA** the chips, the driver and CUDA. The seat stands on **ollama** and **ggml** from **llama.cpp**; the model is Mistral AI's **Mistral Small 3.2** with Nomic's **nomic-embed-text**; Google's **Gemma 4** ran the mixture-of-experts control and writes the answers above, on ollama and on **vLLM**; the reranker is the **Sentence-Transformers** project's `ms-marco-MiniLM-L6-v2`; Black Forest Labs' **FLUX.2 klein 4B** drew the render rows on **ComfyUI** and **PyTorch**. **AMD**, **ASRock Rack**, **ARCTIC** and **CyberPower** made the parts in the picture, **OpenRGB** lit the card, and **systemd**'s journal held every live speed figure. None of them owed us anything. An operator proposed the swap, moved the cards and ran the first round; agents ran the ladder and did the arithmetic; an operator signed the numbers.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 8,
  "digests_kept": 3,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 4,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "hand-edit": 6
  },
  "new_noun_tokens_checked": [
   "0.1",
   "0.8",
   "023-token",
   "05",
   "09",
   "1",
   "10",
   "10.1",
   "10.2",
   "14",
   "15",
   "18.74",
   "2.1",
   "2026-09-25",
   "22.26",
   "24b",
   "26b",
   "300",
   "3090",
   "3090's",
   "350",
   "36",
   "4",
   "42.969",
   "420",
   "450",
   "47.838",
   "49",
   "49.01",
   "53.11",
   "53.354",
   "598.7",
   "60",
   "6000",
   "64",
   "71",
   "74",
   "8.1",
   "8.4",
   "80",
   "895.4",
   "Amble",
   "Beat",
   "C",
   "Gemma",
   "Lab",
   "PRO",
   "RTX",
   "RuleSage",
   "Ti",
   "Ti's",
   "W",
   "gemma4",
   "mistral-small3.2"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "on 2026-09-25, the workstation was down between eleven and thirteen minutes. the RTX PRO 6000 held a 420 W cap, while a 3090 Ti read back a 350 W cap and a 450 W default.",
    "detail": "figures absent: ['3090']"
   },
   {
    "kind": "digest",
    "reason": "figure",
    "item": "this page cannot say if a 3090 Ti is cooler on the workstation, the memory heat, the idle draw on the workstation, the slot, a same-day comparison, the live gain, or boards in general.",
    "detail": "figures absent: ['3090']"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "a 3090 Ti at 350 W was faster on 10.1 %",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "During ten minutes of continuous drawing at 350 W, the 3090 Ti peaked at 60 °C, which was 14 °C cooler than the 3090's peak of 74 °C.",
    "detail": "figures absent from cited spans: ['14']"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 55,
    "text_sha": "fd26f91348c7",
    "figure": "42.969",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the line drops its conditions: 42.969 tokens a second is this Ti at 300 W in a 65,536-token window on the render computer",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the line drops its conditions: 42.969 tokens a second is this Ti at 300 W in a 65,536-token window on the render computer. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 25,
    "text_sha": "49ed15680509",
    "figure": "60",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the line drops its conditions: the 60 °C peak is ten minutes of drawing at 350 W on the render computer",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the line drops its conditions: the 60 °C peak is ten minutes of drawing at 350 W on the render computer. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 252,
    "text_sha": "569e2cd7910e",
    "section": "why-the-swap",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the digest says the RTX PRO 6000 was swapped out; only the 3090 left, and the RTX PRO 6000 stays",
    "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 RTX PRO 6000 was swapped out; only the 3090 left, and the RTX PRO 6000 stays. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 214,
    "text_sha": "ccf95524a530",
    "section": "what-these-cards-serve",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the digest says these cards serve the print lab; the section says the print lab draws on the render computer, on neither card here",
    "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 these cards serve the print lab; the section says the print lab draws on the render computer, on neither card here. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 245,
    "text_sha": "29903ef693be",
    "section": "the-seat-live",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the digest gives 49.01 and 53.11 without saying which card wrote which (the old card 49.01 over 71 replies, this Ti 53.11 over eight); the section's own chip read it backwards",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the digest gives 49.01 and 53.11 without saying which card wrote which (the old card 49.01 over 71 replies, this Ti 53.11 over eight); the section's own chip read it backwards. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 209,
    "text_sha": "08b8246c8f7e",
    "chip_id": "c-ad78a04c",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-09-25",
    "why": "the chip gives the old card's 49.01 to the new card and calls 53.11 the old card's",
    "detail": "removed from this pack by hand on 2026-09-25 (UTC), after the writer ran and after the gates had counted: the chip gives the old card's 49.01 to the new card and calls 53.11 the old card's. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-09-25T17:55:38Z"
}