{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "tokens-a-second-in-comfyui",
 "title": "Tokens a second in ComfyUI, v0.45.0 against v0.40.0, on one RTX 3090",
 "dek": "ComfyUI's built-in language model, timed at two ComfyUI versions from one Qwen3.5 9B file on one GeForce RTX 3090 (24 GB) capped at 250 W, to test a post that says the newer version writes up to twice as fast: with the draft head on, as both versions ship, v0.45.0 wrote 164.5 tokens a second against 81.9 for v0.40.0, so the claim holds at the line, 2.01 times at the middle of ten prompts with four of the ten under 2.00; with the draft head off it is 102.4 against 60.7; and with it on, the text differed from the draft-off text on 10 prompts of 10 at both versions while repeating itself exactly when the same requests ran again in the same order.",
 "published": "2026-11-15",
 "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/tokens-a-second-in-comfyui/",
  "md_url": "https://research.strata2signal.com/tokens-a-second-in-comfyui/index.md",
  "md_sha": "f406ef1cdd84a91d77bcd8a88acaea1a528e41fd68d32586644bd50155aafc31",
  "html_sha": "5b1c029d8ebf47a770e5e53da2615f1776160e35bacff222b0685a540541ccf1",
  "anchors_sha": "5569d0ff78257fcd2b540bbadcb9056f459cfdfa6c5e3257951f8ae5fa99ba88"
 },
 "short": {
  "paragraph": "On one NVIDIA GeForce RTX 3090 (24 GB) capped at 250 W, ComfyUI v0.45.0 wrote 164.5 tokens a second from qwen3.59bw4a8gptq.safetensors with its draft head on (the text node's mtp setting on auto, the default) and 102.4 with it off; v0.40.0 wrote 81.9 and 60.7. Each figure is the middle of ten prompts, sampling off, up to 2,048 tokens a text. Version against version with the draft head on, the way both ship: ×2.01 at the middle, ×1.70 on the lowest prompt, ×2.38 on the highest. So the post's \"up to twice as fast\" holds on this card, at the line: four of the ten prompts came in under 2.00. With the draft head off in both, v0.45.0 is worth ×1.69, nearly the same on every prompt (every change between the two versions but the draft head, the post's new kernel among them, not split); the draft head adds ×1.62 at v0.45.0 and ×1.36 at v0.40.0, more on code and tables and less on dialogue. In the post's own sampler setting the figure is ×1.82, reported and deciding nothing. With the draft head on, the text differed from the draft-off text on 10 prompts of 10 at both versions, and it repeated itself exactly, at 30 of 30 places when the session ran again in the same order; why, this bench did not read. On 48 closed questions each of the four states scored 33. With the model loaded the card read 7,406 to 7,458 MiB in use between requests; of 466 requests none crashed. One card and one file. A second bench timed a stand-in turn of our game RealKeep on the same card, two short narrations and a quick sketch: 15.60 s with the writer's draft head off and 15.32 s with it on, 1.8 per cent shorter. The bet was written down before the run.",
  "from_dek": false,
  "counts": {
   "words": 3963,
   "minutes": 18,
   "tables": 4,
   "kit": false
  },
  "bullets": [
   {
    "text": "the registration for the test was frozen at 23:14:35 UTC",
    "figure": "23:14:35 UTC",
    "cite": "what-the-post-says-and-what-we-counted",
    "quote": "The registration was frozen at 23:14:35 UTC on 2026-11-13; the first request went 33 seconds later and the last ended at 00:41:10 UTC on 2026-11-14, 86 minutes after the first.",
    "span": [
     4212,
     4225
    ]
   },
   {
    "text": "the RTX 3090 reached a temperature of 68 °C",
    "figure": "68 °C",
    "cite": "memory-heat-and-crashes",
    "quote": "It reached 68 °C and its fan 59 per cent.",
    "span": [
     19377,
     19383
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "what-the-post-says-and-what-we-counted",
   "heading": "What the post says, and what we counted",
   "level": 2,
   "span": [
    2693,
    4791
   ],
   "chunks": [
    [
     2693,
     4791
    ]
   ],
   "chars": 2098,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-draft-head-and-where-its-switch-is",
   "heading": "The draft head, and where its switch is",
   "level": 2,
   "span": [
    4791,
    6299
   ],
   "chunks": [
    [
     4791,
     6299
    ]
   ],
   "chars": 1508,
   "digest": "a draft head, or MTP, is a small extra part of a checkpoint that guesses tokens. the switch is a drop-down on the text node. the authors used the TextGenerate node at both versions with specific settings to test the draft head on auto and off.",
   "digest_skipped": null
  },
  {
   "id": "the-table",
   "heading": "The table",
   "level": 2,
   "span": [
    6299,
    8069
   ],
   "chunks": [
    [
     6299,
     8069
    ]
   ],
   "chars": 1770,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "at-the-line-the-ten-prompts",
   "heading": "At the line: the ten prompts",
   "level": 2,
   "span": [
    8069,
    11049
   ],
   "chunks": [
    [
     8069,
     11049
    ]
   ],
   "chars": 2980,
   "digest": "the ten prompts showed varying speed increases. the middle of the ten was 2.009. the prompt rewrite gained most at ×2.381, while verse gained least at ×1.700. the draft head's value varied by prompt, with tokens a round ranging from 2.15 to 3.41 at v0.45.0.",
   "digest_skipped": null
  },
  {
   "id": "what-moves-the-figure",
   "heading": "What moves the figure",
   "level": 2,
   "span": [
    11049,
    16589
   ],
   "chunks": [
    [
     11049,
     16589
    ]
   ],
   "chars": 5540,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "with-the-draft-head-on-it-is-a-different-text",
   "heading": "With the draft head on, it is a different text",
   "level": 2,
   "span": [
    16589,
    19147
   ],
   "chunks": [
    [
     16589,
     19147
    ]
   ],
   "chars": 2558,
   "digest": "with sampling off, the draft-on text is not the draft-off text. the two versions do not write each other's text. while 48 closed questions showed no gross change in score, the draft-on text differed in length on several prompts.",
   "digest_skipped": null
  },
  {
   "id": "memory-heat-and-crashes",
   "heading": "Memory, heat and crashes",
   "level": 2,
   "span": [
    19147,
    19691
   ],
   "chunks": [
    [
     19147,
     19691
    ]
   ],
   "chars": 544,
   "digest": "the RTX 3090 used between 7,406 and 7,458 MiB of memory between requests. the card reached 68 °C. of 466 requests, none crashed or were thrown out.",
   "digest_skipped": null
  },
  {
   "id": "what-it-would-mean-for-a-turn-of-realkeep",
   "heading": "What it would mean for a turn of RealKeep",
   "level": 2,
   "span": [
    19691,
    23110
   ],
   "chunks": [
    [
     19691,
     23110
    ]
   ],
   "chars": 3419,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-say",
   "heading": "What this page does not say",
   "level": 2,
   "span": [
    23110,
    24316
   ],
   "chunks": [
    [
     23110,
     24316
    ]
   ],
   "chars": 1206,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    24316,
    26536
   ],
   "chunks": [
    [
     24316,
     26536
    ]
   ],
   "chars": 2220,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    26536,
    27171
   ],
   "chunks": [
    [
     26536,
     27171
    ]
   ],
   "chars": 635,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    27171,
    27487
   ],
   "chunks": [
    [
     27171,
     27487
    ]
   ],
   "chars": 316,
   "digest": "the article cites a blog post from 2026-11-11 and ComfyUI releases and issue 17115.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-471df4fe",
   "q": "how does the article define the speed increase claim?",
   "a": "The article evaluates a claim that ComfyUI can generate tokens up to twice as fast. The authors set a strict bar of a 2.00 ratio at the middle of ten prompts' ratios, with the draft head on, to determine if the claim holds.",
   "cites": [
    "what-the-post-says-and-what-we-counted"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-32345ed7",
   "q": "how much faster is v0.45.0 compared to v0.40.0 with the draft head on?",
   "a": "With the draft head on, the middle of the ten prompts' ratios for v0.45.0 over v0.40.0 is ×2.01. The lowest ratio observed was ×1.70 and the highest was ×2.38.",
   "cites": [
    "the-table"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-ea501499",
   "q": "which prompt type showed the highest speed increase in v0.45.0?",
   "a": "The prompt rewrite, which reworks an example image prompt, achieved the highest speed increase with a ratio of 2.381 over v0.40.0.",
   "cites": [
    "at-the-line-the-ten-prompts"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-aebaeab3",
   "q": "what portion of the speed increase comes from the draft head?",
   "a": "The speed increase is composed of changes from everything but the draft head and the draft head itself. In v0.45.0, the draft head provided a middle ratio of 1.62, while the changes excluding the draft head provided a middle ratio of 1.69.",
   "cites": [
    "what-moves-the-figure"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § Memory, heat and crashes · § At a glance"
  },
  {
   "slug": "six-cards-the-same-questions",
   "why": "shares ground with § Memory, heat and crashes · § Heat, and how hard the fans worked"
  },
  {
   "slug": "flash-next-on-three-cards",
   "why": "shares ground with § At the line: the ten prompts · § Writing the answer"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to ComfyUI's maintainers, and to the post's author, rattus128, for a post that names its versions and its switch, which made it testable. The runs stood on **ComfyUI** with **comfy-kitchen** and **comfy-aimdo**, Alibaba's **Qwen3.5 9B** as **Comfy-Org** publishes it on **Hugging Face**, Black Forest Labs' **FLUX.2 [klein] 4B** for the sketch, an **NVIDIA** card and driver, **PyTorch**, **CUDA** and Astral's **uv**.\n\nA small human team owns the card, gave the go and ruled on the calls put to it; AI agents did the rest, this page included, under that team's rulings.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 10,
  "digests_kept": 5,
  "chips_written": 7,
  "chips_kept": 4,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 8
  },
  "new_noun_tokens_checked": [
   "1",
   "1.62",
   "1.69",
   "1.70",
   "1.700",
   "1.8",
   "14",
   "15.32",
   "15.60",
   "17115",
   "2.00",
   "2.009",
   "2.01",
   "2.15",
   "2.38",
   "2.381",
   "2026-11-11",
   "23",
   "3.41",
   "3090",
   "35",
   "4.83",
   "406",
   "458",
   "466",
   "48",
   "68",
   "7",
   "900",
   "91.19",
   "C",
   "ComfyUI",
   "LLM",
   "MTP",
   "MiB",
   "Qwen3.x",
   "RTX",
   "TextGenerate",
   "UTC",
   "v0.40.0",
   "v0.45.0"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "the article examines a claim that ComfyUI can double LLM token rates. testing on an RTX 3090 with Qwen3.x models between v0.40.0 and v0.45.0, the authors set a bar of 2.00. they measured tokens a seco",
    "detail": "figures absent: ['3090']"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 47,
    "text_sha": "6b41b406779c",
    "figure": "2.01",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "Version against version with the draft head on, as both ship, the middle of the ten prompts' ratios is ×2.01",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: Version against version with the draft head on, as both ship, the middle of the ten prompts' ratios is ×2.01. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 226,
    "text_sha": "cb4c34a5f772",
    "section": "the-table",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "It holds at the line, and the next table shows how near.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: It holds at the line, and the next table shows how near. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 255,
    "text_sha": "b317a6804a85",
    "section": "what-moves-the-figure",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "Broadly, the prompts that kept more tokens a round gained more from the draft head, as the next chart shows; the order is not exact.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: Broadly, the prompts that kept more tokens a round gained more from the draft head, as the next chart shows; the order is not exact. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 217,
    "text_sha": "d12cce7ef911",
    "section": "what-it-would-mean-for-a-turn-of-realkeep",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "the draft head was worth ×1.27 (×1.19 to ×1.32 over the twelve)",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: the draft head was worth ×1.27 (×1.19 to ×1.32 over the twelve). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 204,
    "text_sha": "67590e7ba2f3",
    "section": "what-this-page-does-not-say",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "The only short texts timed are the turn bench's narrations of about 74 tokens, at v0.45.0 alone.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: The only short texts timed are the turn bench's narrations of about 74 tokens, at v0.45.0 alone. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 305,
    "text_sha": "253c7af28ed2",
    "chip_id": "c-656e8da0",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "That ratio holds every change between the two versions apart from the draft head; the post names several, the new kernel among them, and nothing here splits them.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: That ratio holds every change between the two versions apart from the draft head; the post names several, the new kernel among them, and nothing here splits them. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 255,
    "text_sha": "0465ad6b9743",
    "chip_id": "c-1e36b984",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "With sampling off the model takes its top choice at every step.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: With sampling off the model takes its top choice at every step. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 220,
    "text_sha": "505c9bb77f23",
    "chip_id": "c-531960fc",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-15",
    "why": "these seconds are the card's busy time for one scene change and not what a player waits for.",
    "detail": "removed from this pack by hand on 2026-11-15 (UTC), after the writer ran and after the gates had counted: these seconds are the card's busy time for one scene change and not what a player waits for. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-11-15T03:02:37Z"
}
