The hardware roster · one identity
RTX PRO 6000 Blackwell (96 GB)
96 GB·NVIDIA·graphics card·67 figures · 24 exhibits
Narrow by model
Narrow by power cap
The rail is CSS and nothing else — no script runs on this page. Two things it therefore cannot do, said rather than discovered: a chosen filter is not part of the address, so a narrowed view cannot be linked or bookmarked; and it cannot tell you when nothing matches, which is why the count rides on every control and why a band whose rows are all filtered out keeps its heading and empties.
Language — decode, prefill and first token
Sorted by model, then by power cap ascending — alphabetical and numeric, never by value. Every condition the page states is printed; a condition it does not state reads not stated, which is a third thing and not an uncapped card. The columns a narrow screen cannot hold are printed under the row rather than scrolled to.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 94.6 → 92.7 tok/s per stream | decode per stream | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.152026-08-25 | |||||
| Q432768256 tokens out16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| p50 20.6 → 22.2 ms and p95 20.8 → 23.8 ms | time to first token, p50 and p95 | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | three requests, 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.15three requests, 2026-08-25 | |||||
| Q432768not stated1what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 205.85 tokens a second | decode, tokens a second | gemma4:26b | 420 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| gemma4:26b420 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| gemma read 207.3 / 207.1 / 207.2 tok/s at 600 / 500 / 450 W | decode, the serving lane | gemma4:26b | 600 / 500 / 450 W | Ollama · 0.32.15 | the median of 6–7 per rung, 2026-08-27 |
| gemma4:26b600 / 500 / 450 WOllama · 0.32.15the median of 6–7 per rung, 2026-08-27 | |||||
| not stated32768128 tokens out1what-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 202–214 tok/s | decode after the upgrade | gemma4:26b | not stated | ollama · 0.32.13 | the service restarted at 2026-08-16T01:36:27Z |
| gemma4:26bnot statedollama · 0.32.13the service restarted at 2026-08-16T01:36:27Z | |||||
| not statednot statednot statednot statedthe-free-speed-wasnt-free · the-day-the-box-got-faster · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 124.975 | decode at a deeper context tier | gemma4:26b | not stated | ollama · 0.32.9 | 2026-08 |
| gemma4:26bnot statedollama · 0.32.92026-08 | |||||
| not stateda much deeper context tiernot statednot statedthe-free-speed-wasnt-free · the-day-the-box-got-faster · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 138.6 tok/s over 1,982 live rulings | production median decode | gemma4:26b | not stated | ollama · 0.32.9 | 2026-08-15 |
| gemma4:26bnot statedollama · 0.32.92026-08-15 | |||||
| not statednot statednot statednot statedthe-free-speed-wasnt-free · the-day-the-box-got-faster · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 139.55 (n=6, range 137.75–150.49) | sealed bench median decode | gemma4:26b | not stated | ollama · 0.32.9 | 2026-08-14 |
| gemma4:26bnot statedollama · 0.32.92026-08-14 | |||||
| not statednot statednot statednot statedthe-free-speed-wasnt-free · the-day-the-box-got-faster · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 92.77 tokens a second | decode, tokens a second | mistral-small3.2:24b | 420 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b420 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| Q4_K_Mnot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| mistral 95.0 / 94.6 / 93.5 at the same three | decode, the serving lane | mistral-small3.2:24b | 600 / 500 / 450 W | Ollama · 0.32.15 | the median of 6–7 per rung, 2026-08-27 |
| mistral-small3.2:24b600 / 500 / 450 WOllama · 0.32.15the median of 6–7 per rung, 2026-08-27 | |||||
| not stated32768128 tokens out1what-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 95.0 / 94.6 / 93.5 tokens a second | writing speed, single warm calls | mistral-small3.2:24b | 600 / 500 / 450 W | not stated | 2026-08-27 |
| mistral-small3.2:24b600 / 500 / 450 Wnot stated2026-08-27 | |||||
| Q4_K_Mnot stated128 tokens outnot stateda-short-history-of-mistral · what-does-it-do-on-our-own-machines · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 28 ms / 162 ms | first token on a prompt it has already seen, dense / MoE | mistral-small3.2:24b, gemma4:26b | 420 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b420 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 170.24 → 170.09 tok/s | sustained decode | MoE gemma (FP8), sustained | 600 → 450 W | vLLM · 0.27.1 | averaged over both interleaved repeats, 2026-08-25 |
| MoE gemma (FP8), sustained600 → 450 WvLLM · 0.27.1averaged over both interleaved repeats, 2026-08-25 | |||||
| FP8not statednot stated16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 207.7 tok/s | writing speed | the chat seat | not stated | not stated | the whole run took 35 seconds, on 2026-08-26 |
| the chat seatnot statednot statedthe whole run took 35 seconds, on 2026-08-26 | |||||
| not statednot statednot statedalonethe-typist-and-the-developer · one-card-both-jobs-at-once · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 51.6 tok/s | writing speed | the chat seat | not stated | not stated | the whole run took 35 seconds, on 2026-08-26 |
| the chat seatnot statednot statedthe whole run took 35 seconds, on 2026-08-26 | |||||
| not statednot statednot statedwhile pictures developthe-typist-and-the-developer · one-card-both-jobs-at-once · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 207.5 – 207.7 | writing speed, spread | the chat seat | not stated | not stated | the whole run took 35 seconds, on 2026-08-26 |
| the chat seatnot statednot statedthe whole run took 35 seconds, on 2026-08-26 | |||||
| not statednot statednot statedalonethe-typist-and-the-developer · one-card-both-jobs-at-once · carried by hand, proved against that page’s own row · identity from the same section | |||||
The same card, the same question, power cap apart
busy-window draw
- 419.0 — 450 W what-150-watts-buys · the-mechanism-checked-rather-than-inferred
- 412.8 — 600 W what-150-watts-buys · the-mechanism-checked-rather-than-inferred
Every row above is the same card and the same MoE gemma (FP8), sustained. The only thing that moved is the power cap. That is why nothing on this page is ranked, and why a condition is part of a row rather than a heading above one.
The same card, the same question, concurrency apart
seconds per sketch
- 1.54 s — alone the-typist-and-the-developer · one-card-both-jobs-at-once
- 1.82 s — while tokens stream the-typist-and-the-developer · one-card-both-jobs-at-once
Every row above is the same card and the same the game's own sketch graph. The only thing that moved is the concurrency. That is why nothing on this page is ranked, and why a condition is part of a row rather than a heading above one.
Pictures — seconds per render
13 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 41.0 to 42.2 seconds per painting by the render engine's own clock | seconds per painting | FLUX.1-dev | not stated | ComfyUI · version not stated | not stated |
| FLUX.1-devnot statedComfyUI · version not statednot stated | |||||
| not statednot stated2048×1280, 40 stepsnot statedsix-worlds-painted-large · for-the-technically-curious · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 0.82 s (0.80–0.96) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | August 23rd |
| FLUX.2 klein-4Bnot statednot statedAugust 23rd | |||||
| not statednot stated1024×1024, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 0.26 s (0.26–0.40) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | August 23rd |
| FLUX.2 klein-4Bnot statednot statedAugust 23rd | |||||
| not statednot stated512×512, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 0.48 s (0.47–0.61) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | August 23rd |
| FLUX.2 klein-4Bnot statednot statedAugust 23rd | |||||
| not statednot stated768×768, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 0.48 s | median seconds per render, mixed resolutions | FLUX.2 klein-4B (our sketch artist) | not stated | not stated | August 23rd |
| FLUX.2 klein-4B (our sketch artist)not statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 2.53 s | median seconds per render, mixed resolutions | hidream-o1-dev | not stated | not stated | August 23rd |
| hidream-o1-devnot statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 6.17 s | median seconds per render, mixed resolutions | kandinsky5-lite | not stated | not stated | August 23rd |
| kandinsky5-litenot statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.97 s | median seconds per render, mixed resolutions | krea2-turbo | not stated | not stated | August 23rd |
| krea2-turbonot statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.54 s | seconds per sketch | the game's own sketch graph | not stated | not stated | the whole run took 35 seconds, on 2026-08-26 |
| the game's own sketch graphnot statednot statedthe whole run took 35 seconds, on 2026-08-26 | |||||
| not statednot statedsixteen-passalonethe-typist-and-the-developer · one-card-both-jobs-at-once · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 1.82 s | seconds per sketch | the game's own sketch graph | not stated | not stated | the whole run took 35 seconds, on 2026-08-26 |
| the game's own sketch graphnot statednot statedthe whole run took 35 seconds, on 2026-08-26 | |||||
| not statednot statedsixteen-passwhile tokens streamthe-typist-and-the-developer · one-card-both-jobs-at-once · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 0.77, 1.08, 1.51, 1.89 and 3.52 seconds | seconds per render at four step counts | the shipped painter | not stated | not stated | not stated |
| the shipped painternot statednot statednot stated | |||||
| not statednot stated768², at 4, 8, 12, 16 and 32not statedthis section dates nothing — the page's first run is dated 2026-07-10 and this ladder is the second, undated where it sitsfifty-seven-milliseconds · finding-four-the-argument-with-our-own-settings-and-what-it · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 7.84 s | median seconds per render, mixed resolutions | z-image-base (undistilled, 768 only) | not stated | not stated | August 23rd |
| z-image-base (undistilled, 768 only)not statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.13 s | median seconds per render, mixed resolutions | z-image-turbo | not stated | not stated | August 23rd |
| z-image-turbonot statednot statedAugust 23rd | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
Training
2 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 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 | peak memory, two instruments | ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42 | not stated | Side-Step · version not stated | not stated |
| ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42not statedSide-Step · version not statednot stated | |||||
| not statednot stated400 in allnot statedten-minutes-with-living-artists · ten-minutes-on-the-big-card · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 1.515 seconds per update | seconds per update | ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42 | not stated | Side-Step · version not stated | 10 minutes 8 seconds |
| ACE-Step 1.5, rank 64, alpha 128, bf16, seed 42not statedSide-Step · version not stated10 minutes 8 seconds | |||||
| not statednot stated400 in allnot statedten-minutes-with-living-artists · ten-minutes-on-the-big-card · carried by hand, proved against that page’s own row · identity from the same section | |||||
Power, temperature, and what it costs
33 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 428.6 | busy-window draw | dense mistral (Q4) | 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated1the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads busy W @450what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 448.4 | busy-window draw | dense mistral (Q4) | 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads busy W @450what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2409 | SM clock | dense mistral (Q4) | 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated1the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads SM MHz @450what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2246 | SM clock | dense mistral (Q4) | 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads SM MHz @450what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 520.8 | busy-window draw | dense mistral (Q4) | 600 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated1the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads busy W @600what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 586.5 | busy-window draw | dense mistral (Q4) | 600 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads busy W @600what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2737 | SM clock | dense mistral (Q4) | 600 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated1the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads SM MHz @600what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2764 | SM clock | dense mistral (Q4) | 600 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads SM MHz @600what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 586.5 → 448.4 W | busy-window draw | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 0.161 → 0.206 tok/s per watt | efficiency | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 83 → 72 °C | peak temperature | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.152026-08-25 | |||||
| Q432768not stated16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| the 600 → 450 W cap costs 2.0% | what the cap costs and buys | dense mistral (Q4) | 600 → 450 W | Ollama · 0.32.15 | a median of 48 requests at each cap, 2026-08-25 |
| dense mistral (Q4)600 → 450 WOllama · 0.32.15a median of 48 requests at each cap, 2026-08-25 | |||||
| Q432768256 tokens out16what-150-watts-buys · the-answer-in-one-paragraph · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +8.4% | what a 450 W cap costs | flux-2-klein-4b (the limner) | 450 W, against the same-day 600 W rung | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux-2-klein-4b (the limner)450 W, against the same-day 600 W rungnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 401.7 W | busy draw, p50 | flux-2-klein-4b (the limner) | 600 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux-2-klein-4b (the limner)600 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +19.0% | what a 450 W cap costs | flux1-dev (the heavy painter) | 450 W, against the same-day 600 W rung | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-dev (the heavy painter)450 W, against the same-day 600 W rungnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +12.0% at 512 px, +19.2% at 768, +20.4% at 1024 | what a 450 W cap costs, by resolution | flux1-dev (the heavy painter) | 450 W, against the same-day 600 W rung | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-dev (the heavy painter)450 W, against the same-day 600 W rungnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated512 px, 768, 1024not statedthe table row above pools them — the page's own wordswhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +10.2% | what a 500 W cap costs | flux1-dev (the heavy painter) | 500 W, against the same-day 600 W rung | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-dev (the heavy painter)500 W, against the same-day 600 W rungnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 599.3 W — pegged with no cap to blame | busy draw, p50 | flux1-dev (the heavy painter) | 600 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-dev (the heavy painter)600 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| dragged from 2392 to 1927 MHz, and at 450 to 1796 MHz | the clock under the ladder | flux1-dev (the heavy painter) | 600 → 500 → 450 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-dev (the heavy painter)600 → 500 → 450 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot statednot statednot statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +12.1% | what a 450 W cap costs | flux1-schnell | 450 W, against the same-day 600 W rung | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-schnell450 W, against the same-day 600 W rungnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 449.6 W | busy draw, p50 | flux1-schnell | 600 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| flux1-schnell600 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 64–102 tok/s (median 94) | decode, wall-clock, under load | gemma4:12b | not stated | not stated | not stated |
| gemma4:12bnot statednot statednot stated | |||||
| not statednot statednot statedtaken while the full chair-trials bench shared this cardcove-voice-head-to-head · same-question-three-voices · carried by hand, proved against that page’s own row · identity from the same page | |||||
| $0.045 per million | dollars per million output tokens, electricity only | gemma4:26b | 420 W | not stated | 18.34 cents a kilowatt-hour, the US residential average for June 2026 |
| gemma4:26b420 Wnot stated18.34 cents a kilowatt-hour, the US residential average for June 2026 | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · is-it-cheaper-to-run-it-yourself · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 50–139 tok/s (median 75) | decode, wall-clock, under load | gemma4:26b | not stated | not stated | not stated |
| gemma4:26bnot statednot statednot stated | |||||
| not statednot statednot statedtaken while the full chair-trials bench shared this cardcove-voice-head-to-head · same-question-three-voices · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 594.6 W mean across its first four-minute leg, 87 °C peak across the eight minutes | mean draw and peak temperature | mistral-small3.2:24b | uncapped in that arm | not stated | 2026-08-25 |
| mistral-small3.2:24buncapped in that armnot stated2026-08-25 | |||||
| Q4_K_Mnot statednot statednot stateda-short-history-of-mistral · what-does-it-do-on-our-own-machines · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 3,538 J / 887 J | energy per 1,000 tokens, dense / MoE | mistral-small3.2:24b, gemma4:26b | 420 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b420 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 328.3 W / 182.5 W | watts while decoding, dense / MoE | mistral-small3.2:24b, gemma4:26b | 420 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b420 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 419.0 | busy-window draw | MoE gemma (FP8), sustained | 450 W | vLLM · 0.27.1 | 2026-08-25 |
| MoE gemma (FP8), sustained450 WvLLM · 0.27.12026-08-25 | |||||
| FP8not statednot stated16the 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 sowhat-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 412.8 | busy-window draw | MoE gemma (FP8), sustained | 600 W | vLLM · 0.27.1 | 2026-08-25 |
| MoE gemma (FP8), sustained600 WvLLM · 0.27.12026-08-25 | |||||
| FP8not statednot stated16the 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 capwhat-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2797 | SM clock | MoE gemma (FP8), sustained | 600 W | vLLM · 0.27.1 | 2026-08-25 |
| MoE gemma (FP8), sustained600 WvLLM · 0.27.12026-08-25 | |||||
| FP8not statednot stated16the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads SM MHz @600what-150-watts-buys · the-mechanism-checked-rather-than-inferred · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 37–68 tok/s (median 55) | decode, wall-clock, under load | qwen3.6:27b | not stated | not stated | not stated |
| qwen3.6:27bnot statednot statednot stated | |||||
| not statednot statednot statedtaken while the full chair-trials bench shared this cardcove-voice-head-to-head · same-question-three-voices · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 291.4 W | busy draw, p50 | sdxl (albedobase) | 600 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| sdxl (albedobase)600 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 236.8 W | busy draw, p50 | sdxl (base) | 600 W | not stated | 2026-08-27, 16:19–18:14Z (UTC) |
| sdxl (base)600 Wnot stated2026-08-27, 16:19–18:14Z (UTC) | |||||
| not statednot stated139 sealed cells, n per row 24–31not statedwhat-150-watts-buys · same-dial-other-lane-measured-on-the-production-card-the · carried by hand, proved against that page’s own row · identity from the same page | |||||
Memory — what fits, and at which context
1 row — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 17.5 GB | 29.2 GB | 53.1 GB | resident VRAM, three precisions | qwen3.8:27b | not stated | not stated | not stated |
| qwen3.8:27bnot statednot statednot stated | |||||
| all three precisions32768not statednot statedthe-new-kid · the-other-precisions · carried by hand, proved against that page’s own row · identity from the same section | |||||
Everything else this shelf measured for it
2 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 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 | the exhibit's own cross-card ratio | ACE-Step 1.5 | not stated | Side-Step · version not stated | not stated |
| ACE-Step 1.5not statedSide-Step · version not statednot stated | |||||
| not statednot statednot statednot statedthe exhibit prints this comparison; the roster computes nonethe-ceiling-is-not-the-corpus · forty-minutes-on-a-card-bought-over-five-years-ago · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| refused — does not fit | a 122B model on this card alone, seats stopped | qwen3.5:122b, 10 billion active per token, 75.7 GiB | 420 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| qwen3.5:122b, 10 billion active per token, 75.7 GiB420 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedthe page's own refusal, carried as a refusal: 63.0% of the model would board and the rest would nota-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
Where it appears
Every released exhibit that names this identity, in the shelf’s own order. A count beside an exhibit is how many of its figures are carried above — and it links them, because a number with no way back to what it counts is the thing this shelf exists against.
- A Rig Your Friend Already Owns — 9 figures on this page
- A Short History of Mistral — 9 figures on this page
- Ask About This Page — named, and no figure carried from it — yet
- The August arrivals — named, and no figure carried from it — yet
- The chair trials — named, and no figure carried from it — yet
- It’s Not a Metaphor: A Language Model Is a Compression of Its Corpus — named, and no figure carried from it — yet
- The narrator’s chair, refused. — 3 figures on this page
- Fifty-Seven Milliseconds: What a Diffusion Step Actually Buys — 1 figure on this page
- The licence ledger — named, and no figure carried from it — yet
- Nine Worlds, One Dog — named, and no figure carried from it — yet
- The outside judges — named, and no figure carried from it — yet
- The seat trials — named, and no figure carried from it — yet
- Six Worlds for the Print Lab — 1 figure on this page
- Ten Minutes with Living Artists — 2 figures on this page
- The Ceiling Is Not the Corpus — 1 figure on this page
- The Dog, the Dice, and the Painter: How a Small World Draws Itself — named, and no figure carried from it — yet
- The free speed wasn’t free — 4 figures on this page
- qwen3.8:27b across four house benches: one seat filled, one floor missed — 1 figure on this page
- The open call — named, and no figure carried from it — yet
- The Typist and the Developer — 5 figures on this page
- Everyone on the payroll, three at the table — named, and no figure carried from it — yet
- Three New Voices at the Narrator's Chair — named, and no figure carried from it — yet
- The voice trials — named, and no figure carried from it — yet
- What 150 Watts Buys — 31 figures on this page
How this one is named here
The canonical name above is the full product name. Everything below is a spelling a published page or kit has actually used for it, kept so a reader who searches an older page’s words lands on this row. The set is closed and historical: no page cut after 2026-09-17 writes a memory class where it can write a card.
- 96 GB GDDR7 with ECC, 1792 GB/s
- 96 GB class
- 96 GB workstation card
- 96G VRAM
- 96G card
- RTX PRO 6000 Blackwell (96 GB)
- RTX PRO 6000 Blackwell (96 GB) — the served seat
- RTX PRO 6000 Blackwell Workstation Edition
- a 96 GB workstation card
- a 96 GB workstation-class card (95.0 GiB)
- a 96 GB-class workstation GPU
- a 96G VRAM workstation
- a 96G-VRAM box
- a single 96 GB-class workstation card
- a single workstation-class GPU
- our 96 GB card
- our 96 GB workstation-class card
- the 96 GB class
- the 96 GB workstation
- the 96G VRAM workstation
- the 96G workstation
- the inference workstation
- the lab's workstation-class card — 96 gigabytes of memory