{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "what-150-watts-buys",
 "title": "What 150 Watts Buys",
 "dek": "Turn a 600-watt card down by a quarter and one production model pays two percent of its speed; the other never notices.",
 "published": "2026-08-28",
 "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/what-150-watts-buys/",
  "md_url": "https://research.strata2signal.com/what-150-watts-buys/index.md",
  "md_sha": "ea29b4b3a30493aa2bba6ffb28597b652f6e88047c770731eb44d83abd0876cc",
  "html_sha": "c34f5e4642bdace78d97e0860e18baaa09a7db6ecd4e7024446bbd894aa0c33c",
  "anchors_sha": "6b76c91a816cad1a8b1abad84702efc640a3bad3900c50894164501bd8740fe6"
 },
 "short": {
  "paragraph": "Turn a 600-watt card down by a quarter and one production model pays two percent of its speed; the other never notices.",
  "from_dek": true,
  "counts": {
   "words": 3969,
   "minutes": 18,
   "tables": 2,
   "kit": true
  },
  "bullets": [
   {
    "text": "the dense arm at concurrency 16 buys back 138 W",
    "figure": "138 W",
    "cite": "the-answer-in-one-paragraph",
    "quote": "The answer, in one paragraph On the dense arm - the one actually pinned to the ceiling - the 600 → 450 W cap costs 2.0% of throughput and buys back 138 W, 11 °C, and a 28% jump in efficiency.",
    "span": [
     6110,
     6116
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-dial-and-who-notices",
   "heading": "The dial, and who notices",
   "level": 2,
   "span": [
    787,
    1599
   ],
   "chunks": [
    [
     787,
     1599
    ]
   ],
   "chars": 812,
   "digest": "the article explores the impact of turning a graphics card's power dial down from 600 watts to 450 watts on two different workloads to see which one notices the change.",
   "digest_skipped": null
  },
  {
   "id": "five-words-this-page-leans-on",
   "heading": "Five words this page leans on",
   "level": 2,
   "span": [
    1599,
    2751
   ],
   "chunks": [
    [
     1599,
     2751
    ]
   ],
   "chars": 1152,
   "digest": "this section defines key terms used in the article, including power limit (cap), SM clock, dense vs MoE, busy window, and tok/s per watt.",
   "digest_skipped": null
  },
  {
   "id": "the-moment-this-page-is-about",
   "heading": "The moment this page is about",
   "level": 2,
   "span": [
    2751,
    3525
   ],
   "chunks": [
    [
     2751,
     3525
    ]
   ],
   "chars": 774,
   "digest": "the study uses a 96 GB workstation card defaulting to 600 W, aiming for a 450 W budget for a UPS. measurements were taken on 2026-08-25 using specific driver and software versions.",
   "digest_skipped": null
  },
  {
   "id": "what-we-ran-and-the-rule-behind-every-number",
   "heading": "What we ran, and the rule behind every number",
   "level": 2,
   "span": [
    3525,
    5926
   ],
   "chunks": [
    [
     3525,
     5926
    ]
   ],
   "chars": 2401,
   "digest": "two workloads were tested: gemma4:26b (MoE) served by vLLM at FP8, which brushes 450 W, and mistral-small3.2:24b (dense, Q4) served by Ollama, which averages 594.6 W and 87 °C.",
   "digest_skipped": null
  },
  {
   "id": "the-answer-in-one-paragraph",
   "heading": "The answer, in one paragraph",
   "level": 2,
   "span": [
    5926,
    7265
   ],
   "chunks": [
    [
     5926,
     7265
    ]
   ],
   "chars": 1339,
   "digest": "on the dense arm, a 600 to 450 W cap costs 2.0% of throughput and buys back 138 W, 11 °C, and a 28% jump in efficiency. the MoE arm costs no measurable throughput.",
   "digest_skipped": null
  },
  {
   "id": "the-mechanism-checked-rather-than-inferred",
   "heading": "The mechanism, checked rather than inferred",
   "level": 2,
   "span": [
    7265,
    9074
   ],
   "chunks": [
    [
     7265,
     9074
    ]
   ],
   "chars": 1809,
   "digest": "the cap works by pulling the clock down. for the dense mistral (Q4) at concurrency 16, the SM clock dropped by 18.7%, resulting in a 2.0% throughput loss.",
   "digest_skipped": null
  },
  {
   "id": "the-instrument-that-half-failed-counted-rather-than-asserted",
   "heading": "The instrument that half-failed, counted rather than asserted",
   "level": 2,
   "span": [
    9074,
    10726
   ],
   "chunks": [
    [
     9074,
     10726
    ]
   ],
   "chars": 1652,
   "digest": "the card's power-cap flag worked correctly, but the cumulative counter failed to advance during flagged clamping. a zero flag is evidence, but a zero counter is not.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-say",
   "heading": "What this page does not say",
   "level": 2,
   "span": [
    10726,
    12342
   ],
   "chunks": [
    [
     10726,
     12342
    ]
   ],
   "chars": 1616,
   "digest": "the article clarifies it does not provide a full curve, does not claim MoE is inherently cap-proof, makes no quality claims, and acknowledges client-side noise.",
   "digest_skipped": null
  },
  {
   "id": "what-it-means",
   "heading": "What it means",
   "level": 2,
   "span": [
    12342,
    13120
   ],
   "chunks": [
    [
     12342,
     13120
    ]
   ],
   "chars": 778,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "same-dial-other-lane-measured-on-the-production-card-the",
   "heading": "Same dial, other lane — measured on the production card, the day this page shipped",
   "level": 2,
   "span": [
    13120,
    19253
   ],
   "chunks": [
    [
     13120,
     19253
    ]
   ],
   "chars": 6133,
   "digest": "on the production card, the heavy painter flux1-dev costs 19.0% at 450 W. the serving lanes barely notice the cap, with gemma being flat to a tenth of a percent.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    19253,
    22095
   ],
   "chunks": [
    [
     19253,
     22095
    ]
   ],
   "chars": 2842,
   "digest": "key takeaways include that a power cap is a clock cap, dense models pay while sparse ones do not, and one should trust the flag rather than the counter.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    22095,
    23442
   ],
   "chunks": [
    [
     22095,
     23442
    ]
   ],
   "chars": 1347,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    23442,
    24120
   ],
   "chunks": [
    [
     23442,
     24120
    ]
   ],
   "chars": 678,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-b244c902",
   "q": "how does a power cap affect a graphics card's performance?",
   "a": "A power cap functions by lowering the card's clock speed until the power draw fits the set ceiling. The card does not refuse work, but instead pulls the clock down to obey the limit.",
   "cites": [
    "five-words-this-page-leans-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-23c7855d",
   "q": "what specific workloads were used for the power cap measurements?",
   "a": "The tests compared a gemma4:26b MoE model served by vLLM at FP8 against a mistral-small3.2:24b dense model (Q4) served by Ollama.",
   "cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-717aec61",
   "q": "what throughput is lost when capping the dense model at 450 w?",
   "a": "Capping the dense model from 600 W to 450 W results in a 2.0% loss of throughput. At concurrency 16, the rate moves from 94.6 to 92.7 tok/s per stream.",
   "cites": [
    "the-answer-in-one-paragraph"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a4b80ad3",
   "q": "how does the clock speed change for the dense model under a cap?",
   "a": "For the dense mistral model at concurrency 16, the SM clock drops from 2764 MHz to 2246 MHz, representing a −18.7% change.",
   "cites": [
    "the-mechanism-checked-rather-than-inferred"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-b784a61b",
   "q": "what discrepancy was found in the card's power-cap reporting?",
   "a": "The power-cap flag correctly identified when the card was clamped, but the cumulative counter failed to advance, remaining at the same value even after sixty seconds of flagged clamping.",
   "cites": [
    "the-instrument-that-half-failed-counted-rather-than-asserted"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-13288621",
   "q": "how do different image models respond to a 450 w power limit?",
   "a": "The heavy painter, flux1-dev, sees a 19.0% cost in throughput, whereas lighter models like sdxl (base) show negligible changes, reading as noise around zero.",
   "cites": [
    "same-dial-other-lane-measured-on-the-production-card-the"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "one-3090-against-one-3090-ti",
   "why": "shares ground with § The answer, in one paragraph · § The two models want different caps"
  },
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § The answer, in one paragraph · § Power, heat, and the fifty watts"
  },
  {
   "slug": "two-hours-on-battery",
   "why": "shares ground with § The moment this page is about · § Two hours at 12 tok/s — the test that named the page"
  }
 ],
 "thanks": "",
 "kit": {
  "url": "https://research.strata2signal.com/what-150-watts-buys/data/",
  "licence": "CC BY 4.0",
  "files": [
   "NOTES.md",
   "README.md",
   "noise-floor.md",
   "reference/ollama-q4-mistral-p2k-20260825T105243Z/power.csv",
   "reference/ollama-q4-mistral-p2k-20260825T105243Z/summary.json",
   "reference/ollama-q4-mistral-p512-20260825T104841Z/power.csv",
   "reference/ollama-q4-mistral-p512-20260825T104841Z/summary.json",
   "rungs/ollama-q4-mistral-450W-p512-c1/power.csv",
   "rungs/ollama-q4-mistral-450W-p512-c1/raw.jsonl",
   "rungs/ollama-q4-mistral-450W-p512-c1/summary.json",
   "rungs/ollama-q4-mistral-450W-p512-c1/throttle-delta.json",
   "rungs/ollama-q4-mistral-450W-p512-c1/throttle.csv",
   "rungs/ollama-q4-mistral-450W-p512-c16/power.csv",
   "rungs/ollama-q4-mistral-450W-p512-c16/raw.jsonl",
   "rungs/ollama-q4-mistral-450W-p512-c16/summary.json",
   "rungs/ollama-q4-mistral-450W-p512-c16/throttle-delta.json",
   "rungs/ollama-q4-mistral-450W-p512-c16/throttle.csv",
   "rungs/ollama-q4-mistral-600W-p512-c1/power.csv",
   "rungs/ollama-q4-mistral-600W-p512-c1/raw.jsonl",
   "rungs/ollama-q4-mistral-600W-p512-c1/summary.json",
   "rungs/ollama-q4-mistral-600W-p512-c1/throttle-delta.json",
   "rungs/ollama-q4-mistral-600W-p512-c1/throttle.csv",
   "rungs/ollama-q4-mistral-600W-p512-c16/power.csv",
   "rungs/ollama-q4-mistral-600W-p512-c16/raw.jsonl",
   "rungs/ollama-q4-mistral-600W-p512-c16/summary.json",
   "rungs/ollama-q4-mistral-600W-p512-c16/throttle-delta.json",
   "rungs/ollama-q4-mistral-600W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-450W-p512-c1/power.csv",
   "rungs/vllm-fp8-gemma-450W-p512-c1/raw.jsonl",
   "rungs/vllm-fp8-gemma-450W-p512-c1/summary.json",
   "rungs/vllm-fp8-gemma-450W-p512-c1/throttle-delta.json",
   "rungs/vllm-fp8-gemma-450W-p512-c1/throttle.csv",
   "rungs/vllm-fp8-gemma-450W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-450W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-450W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-450W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-450W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-600W-p512-c1/power.csv",
   "rungs/vllm-fp8-gemma-600W-p512-c1/raw.jsonl",
   "rungs/vllm-fp8-gemma-600W-p512-c1/summary.json",
   "rungs/vllm-fp8-gemma-600W-p512-c1/throttle-delta.json",
   "rungs/vllm-fp8-gemma-600W-p512-c1/throttle.csv",
   "rungs/vllm-fp8-gemma-600W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-600W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-600W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-600W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-600W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-sustR1-450W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-sustR1-450W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-sustR1-450W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-sustR1-450W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-sustR1-450W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-sustR1-600W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-sustR1-600W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-sustR1-600W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-sustR1-600W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-sustR1-600W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-sustR2-450W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-sustR2-450W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-sustR2-450W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-sustR2-450W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-sustR2-450W-p512-c16/throttle.csv",
   "rungs/vllm-fp8-gemma-sustR2-600W-p512-c16/power.csv",
   "rungs/vllm-fp8-gemma-sustR2-600W-p512-c16/raw.jsonl",
   "rungs/vllm-fp8-gemma-sustR2-600W-p512-c16/summary.json",
   "rungs/vllm-fp8-gemma-sustR2-600W-p512-c16/throttle-delta.json",
   "rungs/vllm-fp8-gemma-sustR2-600W-p512-c16/throttle.csv",
   "ladder/PREREG-CAP-LADDER-0827.md",
   "ladder/sample.json",
   "ladder/rows.jsonl",
   "ladder/llm-rows.jsonl",
   "ladder/power-600W.csv",
   "ladder/power-500W.csv",
   "ladder/power-450W.csv",
   "ladder/perf-600W.txt",
   "ladder/perf-500W.txt",
   "ladder/perf-450W.txt",
   "ladder/ARM-500W-NOTES.md",
   "ladder/ARM-450W-NOTES.md",
   "ladder/provenance-ladder.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 1,
  "digests_written": 11,
  "digests_kept": 10,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 2,
   "length": 1,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 2
  },
  "new_noun_tokens_checked": [
   "0.161",
   "0.206",
   "11",
   "138",
   "16",
   "18.7%",
   "19.0%",
   "2.0%",
   "2026-08-25",
   "2246",
   "24b",
   "26b",
   "2764",
   "28%",
   "450",
   "594.6",
   "600",
   "87",
   "92.7",
   "94.6",
   "96",
   "C",
   "FP8",
   "GB",
   "MHz",
   "MoE",
   "Ollama",
   "Q4",
   "SM",
   "UPS",
   "W",
   "flux1-dev",
   "gemma4",
   "mistral-small3.2"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "for a dense workload like Q4, a 600 to 450 W cap trades 2.0% of throughput for 138 W and 11 °C. for sparse MoE that never reaches the ceiling, the cap is a no-op.",
    "detail": "figures absent: ['2.0']"
   },
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "the dense mistral at concurrency 16 saw an SM clock drop of −18.7%",
    "detail": "sm is in the cited section only as code or in a table cell"
   },
   {
    "kind": "bullet",
    "reason": "identifier",
    "item": "the heavy painter flux1-dev at 450 W costs +19.0%",
    "detail": "flux1-dev is in the cited section only as code or in a table cell"
   },
   {
    "kind": "chip",
    "reason": "length",
    "item": "what limitations are placed on the conclusions regarding MoE models?",
    "detail": "q not lowercase"
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "Capping the dense arm at 450 W buys back 138 W and 11 °C, while providing a 28% jump in efficiency, moving from 0.161 to 0.206 tok/s per watt.",
    "detail": "figures absent from cited spans: ['0.161', '0.206']"
   }
  ]
 },
 "generated_utc": "2026-09-10T04:14:07Z"
}