RTX 3090 (24 GB)

24 GB·NVIDIA·graphics card·344 figures, from 10 exhibits · named in 43 exhibits

This roster is in beta and still being tested: every figure is proved against the page it came from, and the way they are read and shown is still changing.

Language 128
Pictures 28
Training 3
Power, temperature, and what it costs 82
Memory 2
Everything else this shelf measured for it 101

Narrow these 344 figuresevery modelmodel: a 26-billion-parameter mixture-of-experts writer · 1 of 344 figuresmodel: a dense 24-billion-parameter model · 1 of 344 figuresmodel: ACE-Step 1.5 · 5 of 344 figuresmodel: FLUX.2 [klein] 4B · 15 of 344 figuresmodel: gemma4:12b · 35 of 344 figuresmodel: gemma4:26b · 68 of 344 figuresmodel: HiDream-O1-Image-Dev · 4 of 344 figuresmodel: Kandinsky 5.0 Image Lite · 4 of 344 figuresmodel: klein-4b-32st · 6 of 344 figuresmodel: klein-4b-8st · 6 of 344 figuresmodel: klein-4b-8st-pixel4 · 6 of 344 figuresmodel: Krea 2 Turbo · 4 of 344 figuresmodel: Mistral Small 3.2 · 76 of 344 figuresmodel: mistral-nemo:12b-instruct-2407 · 40 of 344 figuresmodel: mistral-small3.2:24b, gemma4:26b · 3 of 344 figuresmodel: qwen3.5:122b · 3 of 344 figuresmodel: the same 26-billion writer at 4-bit · 1 of 344 figuresmodel: Z-Image base · 2 of 344 figuresmodel: Z-Image Turbo · 5 of 344 figuresmodel: not stated · 59 of 344 figuresevery power cappower cap: 250 W · 76 of 344 figurespower cap: 280 W · 5 of 344 figurespower cap: 300 W · 137 of 344 figurespower cap: 350 W · 50 of 344 figurespower cap: a 3090 at 250 W, the workstation card at 420 W · 3 of 344 figurespower cap: 280 W, 80 per cent of the card's 350 W rating · 1 of 344 figurespower cap: 250 W · 300 W · 1 of 344 figurespower cap: stated in words, not as a number · 8 of 344 figurespower cap: not stated · 63 of 344 figuresevery exhibitexhibit: A Rig Your Friend Already Owns · 15 of 344 figuresexhibit: A Short History of Mistral · 29 of 344 figuresexhibit: Teaching a Music Model Chopin in Five Minutes · 2 of 344 figuresexhibit: The cost to purchase and run a very capable home AI rig · 29 of 344 figuresexhibit: Ollama and vLLM on one RTX 3090 · 35 of 344 figuresexhibit: RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench · 105 of 344 figuresexhibit: The Ceiling Is Not the Corpus · 2 of 344 figuresexhibit: The quantization ladder · 40 of 344 figuresexhibit: Two 3090s out, one 3090 Ti in · 7 of 344 figuresexhibit: Two 3080s against one 3090, and the cap decides · 80 of 344 figures

Narrow by model any a 26-billion-parameter mixture-of-experts writer 1 a dense 24-billion-parameter model 1 ACE-Step 1.5 5 FLUX.2 [klein] 4B 15 gemma4:12b 35 gemma4:26b 68 HiDream-O1-Image-Dev 4 Kandinsky 5.0 Image Lite 4 klein-4b-32st 6 klein-4b-8st 6 klein-4b-8st-pixel4 6 Krea 2 Turbo 4 Mistral Small 3.2 76 mistral-nemo:12b-instruct-2407 40 mistral-small3.2:24b, gemma4:26b 3 qwen3.5:122b 3 the same 26-billion writer at 4-bit 1 Z-Image base 2 Z-Image Turbo 5 not stated 59

ACE-Step 1.5, as this shelf’s pages spell itACE-Step 1.5 · ACE-Step 1.5, LoRA adapters, three composersFLUX.2 [klein] 4B, as this shelf’s pages spell itFLUX.2 klein 4B · FLUX.2 klein 4B, the sketch model · FLUX.2 klein-4B · FLUX.2 klein-4B (our sketch artist)gemma4:12b, as this shelf’s pages spell itGemma 4 12Bgemma4:26b, as this shelf’s pages spell itgemma4:26b · gemma4:26b (MoE)HiDream-O1-Image-Dev, as this shelf’s pages spell ithidream-o1-devKandinsky 5.0 Image Lite, as this shelf’s pages spell itkandinsky5-liteklein-4b-32st, as this shelf’s pages spell itklein-4b-32st · klein-4b-32stklein-4b-8st, as this shelf’s pages spell itklein-4b-8st · klein-4b-8stklein-4b-8st-pixel4, as this shelf’s pages spell itklein-4b-8st-pixel4 · klein-4b-8st-pixel4Krea 2 Turbo, as this shelf’s pages spell itkrea2-turboMistral Small 3.2, as this shelf’s pages spell itmistral-small3.2:24b · mistral-small3.2:24b (dense)qwen3.5:122b, as this shelf’s pages spell itqwen3.5:122b, 10 billion active per token, 75.7 GiBZ-Image base, as this shelf’s pages spell itz-image-base (undistilled, 768 only)Z-Image Turbo, as this shelf’s pages spell itz-image-turbo

Grouped by the model tag the model roster knows, where it knows one; the page’s own words for it stay on every row.

Narrow by power cap any 250 W 76 280 W 5 300 W 137 350 W 50 a 3090 at 250 W, the workstation card at 420 W 3 280 W, 80 per cent of the card's 350 W rating 1 250 W · 300 W 1 stated in words, not as a number 8 not stated 63

250 W, as this shelf’s pages spell it250 W · 250 watts · matched at 250 W350 W, as this shelf’s pages spell it350 W · 350 W (the board's default)stated in words, not as a number, as this shelf’s pages spell itmeasured before this workshop capped its cards · the lower cap

Grouped by the watts each page states — one control per wattage, however its page spelled it — and the page’s own words stay on every row. A string naming two wattages is a pair and not a spelling, so it keeps a control of its own.

Narrow by exhibit any A Rig Your Friend Already Owns 15 A Short History of Mistral 29 Teaching a Music Model Chopin in Five Minutes 2 The cost to purchase and run a very capable home AI rig 29 Ollama and vLLM on one RTX 3090 35 RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench 105 The Ceiling Is Not the Corpus 2 The quantization ladder 40 Two 3090s out, one 3090 Ti in 7 Two 3080s against one 3090, and the cap decides 80

The fourth facet plan §1 promised and the first edition put only on the shelf: a reader who arrived from an article wants that article’s rows, not all of this card’s. Each count under “where it appears” links the control that narrows to it.

The rail is CSS and nothing else — no script runs on this page. It narrows by model and by power cap independently, and the line above says what each one is set to; the size of the two together is the one number it cannot count. A group or a band with nothing left in it goes with its rows, so what you are left looking at is what matched. A narrowed view is not part of the address, so it cannot be linked — each card’s own page is a URL, and that is the thing to share.

Language — decode, prefill and first token

Grouped by model and by what was measured, and inside a group sorted by power cap ascending — alphabetical and numeric, never by value. The heading of a group carries everything that does not change inside it; each row carries the value and the conditions that do. A condition the page does not state reads not stated, which is a third thing and not an uncapped card. A † means the same reading is printed in a second place on its exhibit — one reading, two printings, not two figures.

value	what moves inside this group
gemma4:12b the page says: Gemma 4 12B decode per request, tokens a secondtok/s300 W · 2026-10-01 (UTC)
82.693tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token windowthe 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step1
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
85.007tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokensthe 536-token prompt with each request's seeded tag, 256 tokens out per request · num_batch: 5121
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
41.428tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token windowthe 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step16
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
24.449tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokensthe 536-token prompt with each request's seeded tag, 256 tokens out per request · num_batch: 51216
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
79.677tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token windowthe 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step2
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
76.43tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokensthe 536-token prompt with each request's seeded tag, 256 tokens out per request · num_batch: 5122
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
72.183tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token windowthe 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step4
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
51.713tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokensthe 536-token prompt with each request's seeded tag, 256 tokens out per request · num_batch: 5124
the page's caption: at four and at eight the sixteen-slot runner's three scored runs decoded at two speeds (52.026, 51.713 and 32.831 tokens a second at four; 23.816, 32.485 and 23.814 at eight), and each median above is one of them. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
58.778tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token windowthe 536-token prompt with each request's seeded tag, 256 tokens out per request · 512 tokens per step8
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
23.816tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokensthe 536-token prompt with each request's seeded tag, 256 tokens out per request · num_batch: 5128
the page's caption: at four and at eight the sixteen-slot runner's three scored runs decoded at two speeds (52.026, 51.713 and 32.831 tokens a second at four; 23.816, 32.485 and 23.814 at eight), and each median above is one of them. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
gemma4:12b the page says: Gemma 4 12B decode, tokens a second (scored)tok/s300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
85.216tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
78.126tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:12b the page says: Gemma 4 12B first token, ms (recorded, not scored)ms300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
1,018.1ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
the page's own words: what this reading is not: a first-token comparison. Ollama's 1,018 ms is about 761 ms of its own load step, on a model already loaded; the page's first-token comparison is arm C's at one request. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
226.0ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
the page's own words: what this reading is not: a first-token comparison. Ollama's 1,018 ms is about 761 ms of its own load step, on a model already loaded; the page's first-token comparison is arm C's at one request. answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:12b the page says: Gemma 4 12B the wait for the first token, 95th percentile, msms300 W · 2026-10-01 (UTC) · batch or steps not stated
212.999ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token window1
at one request a run holds one request, so its 95th percentile is that request's wait (the page's caption). answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
441.866ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokens1
at one request a run holds one request, so its 95th percentile is that request's wait (the page's caption). answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
20,935.24ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokens16
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
3,333.177ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token window16
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
1,068.577ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokens2
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
417.849ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token window2
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
2,244.823ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokens4
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
822.939ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token window4
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
1,641.068ms	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.30.0a 5,120-token window8
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
5,365.722ms	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.34.416 slots of 5,120 tokens8
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § Requests at once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from ollama-and-vllm-on-one-rtx-3090 · section id #requests-at-once
gemma4:12b the page says: Gemma 4 12B total speed of one answer, tokens a second (send to last token)tok/s300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
63.865tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
73.364tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:26b the page says: gemma4:26b (MoE) aggregate decode at one, two and four streamstok/sone ollama build · version not stated · 250 W · one, two and four simultaneous streams · 2026-09-16 (UTC) · quantisation, context and batch or steps not stated
127.1 / 195.8 / 261.4 tokens a second
from A Short History of Mistral § What a power cap actually costs typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
gemma4:26b the page says: gemma4:26b (MoE) decode against the 250 W rungone ollama build · version not stated · 350 W (the board's default) · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
+8.2%
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
gemma4:26b decode and residency out of the box, the server's own placementQ4_K_M · ollama · 0.32.13 · a 4,096 window · 256 tokens out · 2026-09-17 · power cap and concurrency not stated
129.4 tok/s, 100 % on the cards
from Two 3080s against one 3090, and the cap decides § How much can each one hold? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
gemma4:26b decode at the full window, and the memory it usedollama · 0.32.13 · 256 tokens out · concurrency not stated
135.249, 19,415 MiB	Q4_K_M · q4_0350 W131,072not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
136.618, 20,129 MiB†	Q4_K_M · q8_0350 W131,072not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
146.308 tok/s, 20,995 MiB	Q4_K_M · f16 cache350 W131,072not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
133.6 tok/s, 20,995 MiB†	Q4_K_M · f16, unquantisednot statedits full 131,072-token window2026-09-17
from Two 3080s against one 3090, and the cap decides § What the cache type costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-the-cache-type-costs
124.7 tok/s, 19,415 MiB	Q4_K_M · q4_0not statedits full 131,072-token window2026-09-17
from Two 3080s against one 3090, and the cap decides § What the cache type costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-the-cache-type-costs
125.6 tok/s, 20,129 MiB†	Q4_K_M · q8_0, the setting both rigs are measured atnot statedits full 131,072-token window2026-09-17
from Two 3080s against one 3090, and the cap decides § What the cache type costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-the-cache-type-costs
gemma4:26b decode per conversation, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · a 4,096 window · 256 tokens out · 4 · 2026-09-17
66.3tok/s	250 W
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
72.1tok/s	300 W
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
gemma4:26b decode summed across conversations, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · 256 tokens out
126.650tok/s	250 Wa 4,096-token window1not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
126.7tok/s	250 Wa 4,096 window12026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
195.5tok/s	250 Wa 4,096 window22026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
195.522tok/s	250 Wa 4,096-token window2not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
265.164tok/s	250 Wa 4,096-token window4not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
265.2tok/s	250 Wa 4,096 window42026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
134.8tok/s	300 Wa 4,096 window12026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
134.754tok/s	300 Wa 4,096-token window1not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
210.4tok/s	300 Wa 4,096 window22026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
210.419tok/s	300 Wa 4,096-token window2not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
288.4tok/s	300 Wa 4,096 window42026-09-17
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
288.438tok/s	300 Wa 4,096-token window4not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
136.274tok/s	350 Wa 4,096-token window1not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
214.039tok/s	350 Wa 4,096-token window2not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
294.636tok/s	350 Wa 4,096-token window4not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #three-more-arms
gemma4:26b decode with an unquantised cache, tokens a secondtok/sQ4_K_M · f16 · ollama · 0.32.13 · the lower cap · 131,072-token window · 256 tokens out · 2026-09-17 · concurrency not stated
133.6tok/s†
from Two 3080s against one 3090, and the cap decides § How much can each one hold? typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
gemma4:26b decode, tokens a secondtok/sconcurrency not stated
125.6tok/s†	gemma4:26bQ4_K_Mollama · 0.32.13250 W131,072256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § How fast does each one write? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
126.12 tokens a second†	gemma4:26bnot statedone ollama build · version not stated250 Wnot stated256 tokens outthe small hours of 2026-09-16 (UTC)
from A Short History of Mistral § What we measured on two cards in one box read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-short-history-of-mistral · section id #what-we-measured-on-two-cards-in-one-box
126.12tok/s†	gemma4:26b (MoE)not statedone ollama build · version not stated250 Wnot statednot stated2026-09-16 (UTC)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
136.1tok/s†	gemma4:26bQ4_K_Mollama · 0.32.13300 W131,072256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § How fast does each one write? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
136.139tok/s	gemma4:26bQ4_K_Mollama · 0.32.13300 W131,072256 tokens out2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
136.618tok/s†	gemma4:26bQ4_K_Mollama · 0.32.13350 W131,072256 tokens out2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
136.49tok/s	gemma4:26b (MoE)not statedone ollama build · version not stated350 W (the board's default)not statednot stated2026-09-16 (UTC)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
gemma4:26b joules per 1,000 tokensJ / 1k tokQ4_K_M · ollama · 0.32.13 · 256 tokens out · concurrency not stated
1,832J / 1k tok	250 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
1,771.71J / 1k tok	300 Wnot stated2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
1,772J / 1k tok	300 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
1,758.13J / 1k tok	350 Wnot stated2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b (dense) decode against the 250 W rungone ollama build · version not stated · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
—	250 W
the baseline rung — the page prints an em dash, not a figure from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
+41.2%	300 W
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
+46.6%	350 W (the board's default)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
Mistral Small 3.2 decode, tokens a secondtok/sconcurrency not stated
33.1tok/s	mistral-small3.2:24bQ4_K_Mollama · 0.32.13250 W49,152256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § How fast does each one write? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
33.3tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13250 W65,536256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § How fast does each one write? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
34.13 tokens a second†	mistral-small3.2:24bQ4_K_Mone ollama build · version not stated250 Wnot stated256 tokens outthe small hours of 2026-09-16 (UTC)
from A Short History of Mistral § What we measured on two cards in one box read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-short-history-of-mistral · section id #what-we-measured-on-two-cards-in-one-box
33.265tok/s	mistral-small3.2:24bQ4_K_Mollama · 0.32.13250 Wnot statedLadder rungnot stated
the cap table at the top of this section does not carry this rung: it prints a 3090 at 300 and 350 W and the Ti at 300, 350 and 450, and this reading is printed nowhere else on the page from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
34.13tok/s†	mistral-small3.2:24b (dense)not statedone ollama build · version not stated250 Wnot statednot stated2026-09-16 (UTC)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
47.8tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13300 W65,536256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § How fast does each one write? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
47.838tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13300 W65,536256 tokens out2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
48.18tok/s	mistral-small3.2:24b (dense)not statedone ollama build · version not stated300 Wnot statednot stated2026-09-16 (UTC)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
47.838tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13300 Wnot statedsingle-window repeatnot stated
the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
46.880tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13350 W65,536256 tokens out2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
46.880tok/s†	mistral-small3.2:24bQ4_K_Mollama · 0.32.13350 Wnot statedLadder rungnot stated
the same decode the cap table at the top of this section prints for this card at this cap — one reading, two printings, and both rows cite this one section id from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
50.02tok/s	mistral-small3.2:24b (dense)not statedone ollama build · version not stated350 W (the board's default)not statednot stated2026-09-16 (UTC)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b joules per 1,000 tokensJ / 1k tokQ4_K_M · ollama · 0.32.13 · 256 tokens out · concurrency not stated
7,076J / 1k tok	250 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
6,059J / 1k tok	300 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
6,059.20J / 1k tok	300 Wnot stated2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
5,842.92J / 1k tok	350 Wnot stated2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
mistral-nemo:12b-instruct-2407 card memory growth at the 4,096-token question load, MiBMiBOllama · 0.32.13 · 300 W · the 4,096-token question load · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
10,121MiB	Q6_K · cache at 16 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
12,795MiB	Q8_0 · cache at 16 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
6,583MiB	Q3_K_M · cache at 16 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
7,827MiB	Q4_K_M · cache at 16 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
8,939MiB	Q5_K_M · cache at 16 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
mistral-nemo:12b-instruct-2407 first token, msmsOllama · 0.32.13 · 300 W · 4,096 · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
481.8ms	Q4_K_M · cache at 8 bits
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
510.56ms	Q3_K_M · cache at 8 bits
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
516.45ms	Q5_K_M · cache at 8 bits
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
518.25ms	Q6_K · cache at 8 bits
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
548.22ms	Q8_0 · cache at 8 bits
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
mistral-nemo:12b-instruct-2407 joules per 1,000 tokensJ / 1k tokOllama · 0.32.13 · 300 W · 4,096 · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
2,948.31J / 1k tok	Q4_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
3,345.62J / 1k tok	Q5_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
3,439.41J / 1k tok	Q3_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
4,102.0J / 1k tok	Q6_K · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
4,293.87J / 1k tok	Q8_0 · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
mistral-nemo:12b-instruct-2407 largest window held whole, tokenstokensOllama · 0.32.13 · 300 W · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC · context not stated
131,072, topped out	Q6_K · cache at 8 bits
the ladder stops at the first window that does not fit whole; topped out is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
131,072, topped out	Q5_K_M · cache at 8 bits
the ladder stops at the first window that does not fit whole; topped out is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
131,072, topped out	Q4_K_M · cache at 8 bits
the ladder stops at the first window that does not fit whole; topped out is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
131,072, topped out	Q3_K_M · cache at 8 bits
the ladder stops at the first window that does not fit whole; topped out is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
98,304tokens	Q8_0 · cache at 8 bits
the ladder stops at the first window that does not fit whole; topped out is the page's word for a level that held 131,072, the model's advertised window, and never hit a ceiling from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
mistral-nemo:12b-instruct-2407 writing speed, tokens a secondtok/sOllama · 0.32.13 · 300 W · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
60.41tok/s	Q8_0 · cache at 8 bits32,768
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
60.545 (0.033)	Q8_0 · cache at 8 bits4,096
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
64.83tok/s	Q6_K · cache at 8 bits32,768
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
65.724 (0.441)	Q6_K · cache at 8 bits4,096
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
74.135tok/s	Q3_K_M · cache at 8 bits32,768
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
74.894 (0.398)	Q3_K_M · cache at 8 bits4,096
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
78.158tok/s	Q5_K_M · cache at 8 bits32,768
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
78.942 (0.068)	Q5_K_M · cache at 8 bits4,096
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
87.234tok/s	Q4_K_M · cache at 8 bits32,768
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
88.203 (0.252)	Q4_K_M · cache at 8 bits4,096
the median of three scored runs; the page's spread gate (three runs more than 15 per cent apart print no headline) and its throttling flag fired on no window from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
qwen3.5:122b the page says: qwen3.5:122b, 10 billion active per token, 75.7 GiB decode, tokens a second, split, seats stoppedtok/sone ollama build · version not stated · a 3090 at 250 W, the workstation card at 420 W · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
71.48tok/s
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
qwen3.5:122b the page says: qwen3.5:122b, 10 billion active per token, 75.7 GiB decode, tokens a second, split, seats uptok/sone ollama build · version not stated · a 3090 at 250 W, the workstation card at 420 W · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
38.22tok/s
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
a 26-billion-parameter mixture-of-experts writer decode, tokens a secondtok/sthe model server of that day · version not stated · measured before this workshop capped its cards · one question at a time · 24 August 2026, 85 receipts · quantisation, context and batch or steps not stated
137tok/s
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
a dense 24-billion-parameter model decode, tokens a secondtok/sthe model server of that day · version not stated · measured before this workshop capped its cards · one question at a time · 24 August 2026, 85 receipts · quantisation, context and batch or steps not stated
53tok/s
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
mistral-small3.2:24b, gemma4:26b first token on a prompt it has already seen, dense / MoEone ollama build · version not stated · 250 W · 256 tokens out · the small hours of 2026-09-16 (UTC) · quantisation, context and concurrency not stated
98 ms / 167 ms
from A Short History of Mistral § What we measured on two cards in one box read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-short-history-of-mistral · section id #what-we-measured-on-two-cards-in-one-box
not stated decode once the window is full, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · concurrency not stated
72.8tok/s	300 WOne 131,072-token window with about 98,000 tokens really in it256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § What about a long document? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-about-a-long-document
73.327tok/s	350 W98,000 tokens in a 131,072-token windownot statednot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #three-more-arms
not stated prompt reading, tokens a secondtok/sconcurrency not stated
2,270.547tok/s	Q4_K_Mollama · 0.32.13350 W98,000 tokens in a 131,072-token windownot statednot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #three-more-arms
About 3,600	not statedthe model server of that day · version not statedmeasured before this workshop capped its cardsnot stated2,667 tokens in 0.73 s24 August 2026
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
not stated reading 98,024 tokens, secondssQ4_K_M · ollama · 0.32.13 · 300 W · One 131,072-token window with about 98,000 tokens really in it · 256 tokens out · 2026-09-17 · concurrency not stated
45.12 s
from Two 3080s against one 3090, and the cap decides § What about a long document? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-about-a-long-document
not stated time to the first tokensthe model server of that day · version not stated · measured before this workshop capped its cards · one question at a time · 24 August 2026 · quantisation and context not stated
1.01 s	a 2,000-token prompt
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
0.08 s	a 40-token prompt
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
0.30 s	a 400-token prompt
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
not stated time to the first token, secondssQ4_K_M · ollama · 0.32.13 · 300 W · One 131,072-token window with about 98,000 tokens really in it · 256 tokens out · 2026-09-17 · concurrency not stated
46.64 s
from Two 3080s against one 3090, and the cap decides § What about a long document? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-about-a-long-document
the same 26-billion writer at 4-bit aggregate decode, eight questions at oncetok/s4-bit · a batching runtime · version not stated · measured before this workshop capped its cards · eight questions at once · 22 August 2026 · context and batch or steps not stated
328tok/s
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect

Where one condition moved, and nothing else

decode, tokens a second — power cap apart

The same mistral-small3.2:24b (dense)

34.13 — 250 W What a power cap actually costs
48.18 — 300 W What a power cap actually costs
50.02 — 350 W (the board's default) What a power cap actually costs

Every row above is the same card and the same mistral-small3.2:24b (dense). 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.

energy per 1,000 tokens — power cap apart

The same mistral-small3.2:24b (dense)

6,680 J — 250 W What a power cap actually costs
5,673 J — 300 W What a power cap actually costs
5,318 J — 350 W (the board's default) What a power cap actually costs

Pictures — seconds per render

28 rows — same rule as above.

value	what moves inside this group
FLUX.2 [klein] 4B the page says: FLUX.2 klein-4B (our sketch artist) median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
1.65 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
FLUX.2 [klein] 4B the page says: FLUX.2 klein-4B seconds per render, median (min–max)squantisation, runtime, power cap, context, concurrency and measurement window not stated
2.86 s (2.84–3.17)	1024×1024, steps held at 4, six renders per cell
from A Rig Your Friend Already Owns § The sketch ladder — the number the game lives on read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-sketch-ladder-the-number-the-game-lives-on
0.88 s (0.88–1.19)	512×512, steps held at 4, six renders per cell
from A Rig Your Friend Already Owns § The sketch ladder — the number the game lives on read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-sketch-ladder-the-number-the-game-lives-on
1.65 s (1.64–1.96)	768×768, steps held at 4, six renders per cell
from A Rig Your Friend Already Owns § The sketch ladder — the number the game lives on read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-sketch-ladder-the-number-the-game-lives-on
FLUX.2 [klein] 4B seconds per render, median of its cellsthe night of 26 August 2026 · quantisation, runtime, power cap, context and concurrency not stated
2.86s	FLUX.2 klein 4B1,024, 4 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
0.88s	FLUX.2 klein 4B, the sketch model512, 4 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
1.65s	FLUX.2 klein 4B768, 4 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
FLUX.2 [klein] 4B the page says: FLUX.2 klein-4B (our sketch artist) the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
3.4×
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
gemma4:26b the exhibit's own multiplier, out of the box, against the pairQ4_K_M · ollama · 0.32.13 · a 4,096 window · 256 tokens out · power cap, concurrency and measurement window not stated
one 3090, 2.17×
the cap is not stated because the page states it is not one cap: the cell's own words are that each rig was measured at one cap only here, and they are not the same cap from Two 3080s against one 3090, and the cap decides § At a glance typed in from that page by hand, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #at-a-glance
HiDream-O1-Image-Dev the page says: hidream-o1-dev median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
7.34 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
HiDream-O1-Image-Dev the page says: hidream-o1-dev seconds per render, median of its cellsthe night of 26 August 2026 · quantisation, runtime, power cap, context and concurrency not stated
11.62s	1,024, 32 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
6.19s	512, 32 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
HiDream-O1-Image-Dev the page says: hidream-o1-dev the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
†
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Kandinsky 5.0 Image Lite the page says: kandinsky5-lite median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
28.95 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Kandinsky 5.0 Image Lite the page says: kandinsky5-lite seconds per render, median of its cellsthe night of 26 August 2026 · quantisation, runtime, power cap, context and concurrency not stated
49.89s	1,024, 32 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
11.43s	512, 32 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
Kandinsky 5.0 Image Lite the page says: kandinsky5-lite the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
4.7×
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Krea 2 Turbo the page says: krea2-turbo median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
9.62 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Krea 2 Turbo the page says: krea2-turbo seconds per render, median of its cellsthe night of 26 August 2026 · quantisation, runtime, power cap, context and concurrency not stated
16.09s	1,024, 8 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
5.34s	512, 8 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
Krea 2 Turbo the page says: krea2-turbo the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
4.9×
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Z-Image base the page says: z-image-base (undistilled, 768 only) median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
33.01 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Z-Image base the page says: z-image-base (undistilled, 768 only) the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
4.2×
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Z-Image Turbo the page says: z-image-turbo median seconds per render, mixed resolutionssidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
4.44 s
from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question
Z-Image Turbo the page says: z-image-turbo seconds per render, median of its cellsthe night of 26 August 2026 · quantisation, runtime, power cap, context and concurrency not stated
8.08s	1,024, 8 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
2.10s	512, 8 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
4.44s	768, 8 steps
the page states no cap for this arm: it reads the card at 320 W at the wall that night, above the 280 W it recommends, and says a capped card draws each of these a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
Z-Image Turbo the page says: z-image-turbo the exhibit's own multiplier against the 96 GB cardidentical mixed resolution-and-step sets on every rig · quantisation, runtime, power cap, context, concurrency and measurement window not stated
3.9×
the exhibit prints this ratio; the roster computes none from A Rig Your Friend Already Owns § The table that answers the question read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from a-rig-your-friend-already-owns · section id #the-table-that-answers-the-question

Training

3 rows — same rule as above.

value	what moves inside this group
ACE-Step 1.5 peak memory, the trainer's own allocatorGiBSide-Step · version not stated · 250 W · 331 tracks five times over, 415 small corrections in all · quantisation, context, concurrency and measurement window not stated
5.7 GiB
from The Ceiling Is Not the Corpus § Forty minutes on a card bought over five years ago read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-ceiling-is-not-the-corpus · section id #forty-minutes-on-a-card-bought-over-five-years-ago
ACE-Step 1.5 peak memory, two instrumentsSide-Step · version not stated · 250 watts · ten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1) · quantisation, context, concurrency and measurement window not stated
5.7 gigabytes by the trainer's own bookkeeping; the card itself showed 6.8 in use
from Teaching a Music Model Chopin in Five Minutes § Five minutes of training, honestly accounted typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from chopin-in-five-minutes · section id #five-minutes-of-training-honestly-accounted
ACE-Step 1.5 seconds per optimiser stepsSide-Step · version not stated · 250 W · 331 tracks five times over, 415 small corrections in all · quantisation, context, concurrency and measurement window not stated
5.73s
from The Ceiling Is Not the Corpus § Forty minutes on a card bought over five years ago read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-ceiling-is-not-the-corpus · section id #forty-minutes-on-a-card-bought-over-five-years-ago

Power, temperature, and what it costs

82 rows — same rule as above.

value	what moves inside this group
ACE-Step 1.5 wall clockSide-Step · version not stated · 250 watts · ten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1) · quantisation, context, concurrency and measurement window not stated
4 minutes 49 seconds by the trainer's own clock
from Teaching a Music Model Chopin in Five Minutes § Five minutes of training, honestly accounted typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from chopin-in-five-minutes · section id #five-minutes-of-training-honestly-accounted
ACE-Step 1.5 the page says: ACE-Step 1.5, LoRA adapters, three composers wall-clock, three composers250 W · training for three composers · quantisation, runtime, context, batch or steps and concurrency not stated
27 minutes 48 seconds
from The cost to purchase and run a very capable home AI rig § What the card is like for training typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-the-card-is-like-for-training
gemma4:12b the page says: Gemma 4 12B decode, tokens a second (the server's own clock)tok/s300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
85.553tok/s	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
78.118tok/s	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:12b the page says: Gemma 4 12B joules per 1,000 tokens, busy watts ÷ decode speed (arithmetic, not registered)J / 1k tok300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
3,402J / 1k tok	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
3,829J / 1k tok	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:26b dollars per token a second, at what was paid250 W · bought for $1,449.99 in May 2026 · quantisation, runtime, context, batch or steps and concurrency not stated
$11.50 per token a second
from A Short History of Mistral § Is it cheaper to run it yourself? typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from a-short-history-of-mistral · section id #is-it-cheaper-to-run-it-yourself
gemma4:26b the page says: gemma4:26b (MoE) draw while decodingWone ollama build · version not stated · 250 W · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
206.4 W
from A Short History of Mistral § What a power cap actually costs typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
gemma4:26b the page says: gemma4:26b (MoE) energy per 1,000 tokensJ / 1k tokone ollama build · version not stated · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
1,632 J	250 W
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
1,823 J	350 W (the board's default)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
gemma4:26b idle draw with the model resident, median and mean57 samples over thirty seconds each · quantisation, runtime, power cap, context, concurrency and measurement window not stated
22.26 and 52.76
the page's own reading of the pair: the median is the idle figure and the mean an upper bound containing a load from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #three-more-arms
gemma4:26b mean drawWQ4_K_M · ollama · 0.32.13 · 256 tokens out · concurrency not stated
230 W	250 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
241.53 W	300 Wnot stated2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
242 W	300 Weach rig's headline window2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
240.49 W	350 Wnot stated2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
gemma4:26b tokens a second per $100 of the price as readQ4_K_M · ollama · 0.32.13 · 256 tokens out · context, concurrency and measurement window not stated
7.24	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
7.27	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
Mistral Small 3.2 the page says: mistral-small3.2:24b core clock, medianMHzQ4_K_M · ollama · 0.32.13 · 256 tokens out · context and concurrency not stated
1,275 MHz	300 W2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
1,215 MHz	350 W2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b dollars per million output tokens, electricity only300 W · 18.34 cents a kilowatt-hour, the US residential average for June 2026 · quantisation, runtime, context, batch or steps and concurrency not stated
$0.29 per million
from A Short History of Mistral § Is it cheaper to run it yourself? typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from a-short-history-of-mistral · section id #is-it-cheaper-to-run-it-yourself
Mistral Small 3.2 the page says: mistral-small3.2:24b dollars per token a second, at what was paid300 W · bought for $1,449.99 in May 2026 · quantisation, runtime, context, batch or steps and concurrency not stated
$30.10
from A Short History of Mistral § Is it cheaper to run it yourself? typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from a-short-history-of-mistral · section id #is-it-cheaper-to-run-it-yourself
Mistral Small 3.2 the page says: mistral-small3.2:24b draw as a share of the cap%Q4_K_M · ollama · 0.32.13 · context and concurrency not stated
94.2 %	250 WLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
96.6 %†	300 W256 tokens out2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
96.6 %†	300 Wsingle-window repeatnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
78.1 %†	350 W256 tokens out2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
78.1 %†	350 WLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b (dense) draw while decoding, and the clockone ollama build · version not stated · 250 W · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
228.2 W — 91% of cap — clock pulled to 495–675 MHz
from A Short History of Mistral § What a power cap actually costs typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b (dense) energy per 1,000 tokensJ / 1k tokone ollama build · version not stated · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
6,680 J	250 W
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
5,673 J	300 W
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
5,318 J	350 W (the board's default)
from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b mean drawWQ4_K_M · ollama · 0.32.13 · concurrency not stated
235 W	250 Weach rig's headline window256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
235.38 W	250 Wnot statedLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
289.86 W†	300 Wnot stated256 tokens out2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
290 W	300 Weach rig's headline window256 tokens out2026-09-17
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
289.86 W†	300 Wnot statedsingle-window repeatnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
273.35 W†	350 Wnot stated256 tokens out2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
273.35 W†	350 Wnot statedLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b memory clock held under loadMHzQ4_K_M · ollama · 0.32.13 · context, concurrency and measurement window not stated
9,501 MHz	250 WLadder rung
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
9,501 MHz	300 Wsingle-window repeat
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
9,501 MHz	350 WLadder rung
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b streamed bandwidth while decodingGB/sQ4_K_M · ollama · 0.32.13 · context, concurrency and measurement window not stated
504.9 GB/s	250 WLadder rung
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
726.1 GB/s	300 Wsingle-window repeat
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
711.5 GB/s	350 WLadder rung
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b the bandwidth ceiling at the memory clock observedGB/sQ4_K_M · ollama · 0.32.13 · context, concurrency and measurement window not stated
912.1 GB/s	250 WLadder rung
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
912.1 GB/s	300 Wsingle-window repeat
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
912.1 GB/s	350 WLadder rung
the page derives this from the memory clock observed at that rung and the bus width and says so: a ceiling at an observed clock, never an achieved rate. A 3090's two bandwidth columns are computed from what that bench's files carry — its own per-arm harness predates the instrument — and the Ti's 384 bits is declared rather than probed from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b tokens a second per $100 of the price as readQ4_K_M · ollama · 0.32.13 · 256 tokens out · context, concurrency and measurement window not stated
2.54	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
2.49	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
Mistral Small 3.2 the page says: mistral-small3.2:24b what fifty watts are worth, denseQ4_K_M · ollama · 0.32.13 · 250 W · 256 tokens out · 2026-09-17 · context and concurrency not stated
43.8 per cent†
from Two 3080s against one 3090, and the cap decides § How fast does each one write? typed in from that page by hand, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
mistral-small3.2:24b, gemma4:26b energy per 1,000 tokens, dense / MoEone ollama build · version not stated · 250 W · 256 tokens out · the small hours of 2026-09-16 (UTC) · quantisation, context and concurrency not stated
6,680 J / 1,632 J
from A Short History of Mistral § What we measured on two cards in one box read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-short-history-of-mistral · section id #what-we-measured-on-two-cards-in-one-box
mistral-small3.2:24b, gemma4:26b watts while decoding, dense / MoEone ollama build · version not stated · 250 W · 256 tokens out · the small hours of 2026-09-16 (UTC) · quantisation, context and concurrency not stated
228.2 W / 206.4 W
from A Short History of Mistral § What we measured on two cards in one box read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from a-short-history-of-mistral · section id #what-we-measured-on-two-cards-in-one-box
not stated core clock while drawing, medianMHzComfyUI · 0.21.1 · 250 W · batch 3 · 2026-09-17 · quantisation, context and concurrency not stated
1,575MHz
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
not stated core temperature peak°CComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context and concurrency not stated
67 °C	250 W2026-09-17
the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
67 °C	250 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
74 °C	350 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated core temperature plateauComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context and concurrency not stated
65 °C at 0.5 min	250 W2026-09-17
the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
65 °C at 0.5 min	250 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
71 °C at 0.9 min	350 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated draw at the wallWthe night of 26 August 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
320 W
the page prints this beside the render table to say the card was NOT at the 280 W cap it recommends that night, and that a capped card would draw each of those pictures a few per cent slower, unmeasured from The cost to purchase and run a very capable home AI rig § What kind of performance to expect typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect
not stated energy at depth, joules per 1,000 tokensJ / 1k tokQ4_K_M · ollama · 0.32.13 · 300 W · One 131,072-token window with about 98,000 tokens really in it · 256 tokens out · 2026-09-17 · concurrency not stated
3,902J / 1k tok
from Two 3080s against one 3090, and the cap decides § What about a long document? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #what-about-a-long-document
not stated idle draw, the empty and the loaded window sharing one medianW350 W · quantisation, runtime, context, batch or steps, concurrency and measurement window not stated
22.39 W
the page prints this as its evidence that the gap between the median and the mean is a load spike inside the window, not a difference between an empty card and a loaded one from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more arms typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #three-more-arms
not stated images drawn in ten minutesimagesComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context and concurrency not stated
269images	250 W2026-09-17
the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
269images	250 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
311images	350 Wnot stated
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated median drawWComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
244.3 W	250 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
325.8 W	350 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated pictures a minute per $100 of the price as readQ4_K_M · ComfyUI · 0.21.1 · 8 steps, batch 3 · context, concurrency and measurement window not stated
2.05	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
2.11	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readings read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-the-240-buys-on-those-two-readings
not stated price, one cardquantisation, runtime, power cap, context, batch or steps and concurrency not stated
$1,094.44	August 2026's average completed sale
from The cost to purchase and run a very capable home AI rig § The cards, in May and in September read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
$1,449.99	paid by this workshop, May 2026
from The cost to purchase and run a very capable home AI rig § The cards, in May and in September read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
$1,829.99	13 September 2026
from The cost to purchase and run a very capable home AI rig § The cards, in May and in September read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
$1,849.99	13 September 2026
from The cost to purchase and run a very capable home AI rig § The cards, in May and in September read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
not stated price, one card, renewedquantisation, runtime, power cap, context, batch or steps and concurrency not stated
$1,879.99	read by hand at the store page on 17 September 2026 between 15:00 and 15:30 UTC
the replacement price of record, read by hand because that retailer serves this workshop's fetcher no price; exhibit fifty publishes the same reading of the same listing from The cost to purchase and run a very capable home AI rig § The cards, in May and in September typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
$1,879.99	read 17 September 2026 in the same half hour
the record behind this figure is public — the listing its page prints in full, read in the window that page namesread by hand at the store page, because that retailer serves no price to this workshop's fetcher; it replaces an earlier figure this workshop published on 13 September that an operator ruled inaccurate and struck from Two 3080s against one 3090, and the cap decides § What it costs, priced 17 September 2026 typed in from that page by hand, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #what-it-costs-priced-17-september-2026
$1,879.99	read by hand at the store page on 17 September 2026
the record behind this figure is public — the listing its page prints in full, read in the window that page namesthis 3090 has no price on record — it was already in the workshop — and this is a listing of the exact board, read by hand because on 18 September both retailers served no price to this workshop's fetcher from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What it cost typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #what-it-cost
not stated price, one card, usedthe past year's average completed sale · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
$908.00
WITHDRAWN BY ITS OWN EXHIBIT on 2026-09-18: the page quotes the row it used to print and says who withdrew the figure and why. The two totals derived from it, $2,256.87 and $3,164.87, went with it“An operator ruled it inaccurate on 2026-09-17 — a trailing twelve-month average of a market that has since climbed is not a price a reader can pay — and it is no longer a price of record on this shelf.” — home-inference-and-diffusion-rig-cost · the-cards-in-may-and-in-september from The cost to purchase and run a very capable home AI rig § The cards, in May and in September typed in from that page by hand, then checked against it word for word · the card is named in this row · read from home-inference-and-diffusion-rig-cost · section id #the-cards-in-may-and-in-september
not stated seconds an image, over ten minutes of continuous drawingsComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
2.156s	250 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
1.886s	350 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated the exhibit's own energy saving against the pairQ4_K_M · ollama · 0.32.13 · 300 W · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
40.8 per cent less per thousand tokens on the mixture at 300 W and 46.8 per cent less on the dense model
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
not stated the memory clock's own reported maximumMHzquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
9,751 MHz
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The question typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #the-question
not stated what fifty watts are worth to the single cardQ4_K_M · ollama · 0.32.13 · 250 W · 300 W · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
43.8 per cent on the dense model and 8.4 on the mixture†
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
not stated what the house cap costs280 W, 80 per cent of the card's 350 W rating · 23 August 2026 · quantisation, runtime, context, batch or steps and concurrency not stated
3 per cent of writing speed and 6 per cent of prompt reading
from The cost to purchase and run a very capable home AI rig § What kind of performance to expect typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from home-inference-and-diffusion-rig-cost · section id #what-kind-of-performance-to-expect

Memory — what fits, and at which context

2 rows — same rule as above.

value	what moves inside this group
qwen3.5:122b the page says: qwen3.5:122b, 10 billion active per token, 75.7 GiB the fraction of a 122B model on card, and aggregate decode, seats upone ollama build · version not stated · a 3090 at 250 W, the workstation card at 420 W · 2026-09-16 (UTC) · quantisation, context, batch or steps and concurrency not stated
39.5%
split across the 3090 and the workstation card; the page's own words from A Short History of Mistral § What a power cap actually costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
not stated card board power, the model residentW250 W · nothing generating · 2026-09-17 03:15 to 03:52 UTC · quantisation, runtime, context and concurrency not stated
53.1 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts

Everything else this shelf measured for it

101 rows — same rule as above.

value	what moves inside this group
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B milliseconds a sleeve, maximumms280 W · 2026-09-11 14:33 → 2026-09-18 01:38 · quantisation, runtime, context, batch or steps and concurrency not stated
2,899ms
a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B milliseconds a sleeve, medianms280 W · 2026-09-11 14:33 → 2026-09-18 01:38 · quantisation, runtime, context, batch or steps and concurrency not stated
2,551ms
a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B milliseconds a sleeve, minimumms280 W · 2026-09-11 14:33 → 2026-09-18 01:38 · quantisation, runtime, context, batch or steps and concurrency not stated
2,497ms
a live figure, read off the service's own journal: the sleeve painter draws a cover for a groove posted to the board, at the clock range this row's own window cell prints from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B seconds a batch of three, mediansa batch of three, on the same 32-step graph the open set hangs · 2026-09-17 · quantisation, runtime, context and concurrency not stated
18.89s	250 W
a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
17.48s	300 W
a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
17.01s	350 W
a bench figure, and the page derives it rather than reading it: the companion bench publishes seconds per finished image and no spread, and this column is that cell times three — the batch the print lab actually draws. The page says so in its own dek above the table from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B the exhibit's own bracket for the old machine's 280 W cap, seconds a batch of three280 W · a batch of three, on the same 32-step graph the open set hangs · quantisation, runtime, context, concurrency and measurement window not stated
17.5 to 18.9 seconds per batch
a bracket and not a reading: the bench ran 250 and 300 W rungs and no 280 W one, and the page rounds the two rows either side of the old machine's cap rather than interpolating between them from Two 3090s out, one 3090 Ti in § Before and after, on the same instruments typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
gemma4:12b the page says: Gemma 4 12B board power while busy, WW300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
289.88W	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
299.14W	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:12b the page says: Gemma 4 12B of which Ollama's load step, msmsgemma4:12b-it-qat, a 4-bit GGUF · Ollama · 0.32.13 · 300 W · 16 slots of 5,120 tokens · One 536-token prompt (series-p512), 256 tokens out · num_batch: 512 · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
761.1ms
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:12b the page says: Gemma 4 12B spread of the ten runs300 W · one request at a time · 2026-09-28 from 11:29:59 to 11:33:35 UTC
0.41 %	gemma4:12b-it-qat, a 4-bit GGUFOllama · 0.32.1316 slots of 5,120 tokensOne 536-token prompt (series-p512), 256 tokens out · num_batch: 512
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
0.50 %	google/gemma-4-12B-it-qat-w4a16-ct at revision 1d2c2d7fvLLM · 0.27.1a 4,352-token windowOne 536-token prompt (series-p512), 256 tokens out · our kit's own serve line
answers differ at this level: the page's label on every speed row, because the two servers agreed on 63 of the six-way's 108 items against the 97 per cent it had bet on (P-6) from Ollama and vLLM on one RTX 3090 § One request at a time read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from ollama-and-vllm-on-one-rtx-3090 · section id #one-request-at-a-time
gemma4:26b core temp max°CQ4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
63 °C	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
47 °C	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
gemma4:26b fan max%Q4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
42 %	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
0 %	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
gemma4:26b largest window held whole, and the memory it used, the layer count namedQ4_K_M · ollama · 0.32.13 · 256 tokens out · 2026-09-17 · power cap, context and concurrency not stated
131,072 whole, 20,129 MiB used
from Two 3080s against one 3090, and the cap decides § How much can each one hold? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
gemma4:26b largest window held whole, placement left to the serverQ4_K_M · ollama · 0.32.13 · 256 tokens out · power cap, context and concurrency not stated
131,072 whole, 20,129 MiB	not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
131,072 whole	2026-09-17
from Two 3080s against one 3090, and the cap decides § How much can each one hold? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
gemma4:26b the cap flag, the driver's own reportQ4_K_M · ollama · 0.32.13 · 256 tokens out · context and concurrency not stated
Active in 100 %	300 W2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Active in 83.3 %	350 W2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
gemma4:26b the exhibit's own change from its 300 W rung to its 350 W rungQ4_K_M · ollama · 0.32.13 · 350 W · context, batch or steps, concurrency and measurement window not stated
0.35 per cent
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
gemma4:26b the exhibit's own margin over the pair, mixtureQ4_K_M · ollama · 0.32.13 · 131,072 · 256 tokens out · 2026-09-17 · concurrency not stated
9.4 per cent	250 W
from Two 3080s against one 3090, and the cap decides § How fast does each one write? typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
18.0 per cent	300 W
from Two 3080s against one 3090, and the cap decides § How fast does each one write? typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
gemma4:26b the exhibit's own margin over the pair, one caller and fourQ4_K_M · ollama · 0.32.13 · 300 W · a 4,096 window · 256 tokens out · 2026-09-17 · concurrency not stated
7.8 per cent
from Two 3080s against one 3090, and the cap decides § The full concurrency ladder typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #the-full-concurrency-ladder
gemma4:26b tok/stok/sQ4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
125.6tok/s†	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
136.1tok/s†	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
gemma4:26b what an unquantised cache buys at the full windowQ4_K_M · f16, unquantised · ollama · 0.32.13 · 350 W · 131,072 · batch or steps, concurrency and measurement window not stated
+7.1 on a 3090
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
gemma4:26b what an unquantised cache needs at the full windowQ4_K_M · f16 · ollama · 0.32.13 · 131,072-token window · 2026-09-17 · power cap, batch or steps and concurrency not stated
20,995 MiB, which is 515 more than two 10 GB cards have between them†
from Two 3080s against one 3090, and the cap decides § How much can each one hold? typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
gemma4:26b what compressing the cache to 8 bits costsQ4_K_M · q8_0 · ollama · 0.32.13 · its full 131,072-token window · 2026-09-17 · power cap, batch or steps and concurrency not stated
about 6 per cent
from Two 3080s against one 3090, and the cap decides § What the cache type costs typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #what-the-cache-type-costs
klein-4b-32st seconds a finished image, median of two timed submitssComfyUI · 0.21.1 · batch 3 · quantisation, context, concurrency and measurement window not stated
6.297 s	250 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #renders
5.827 s	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
5.670 s	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
klein-4b-32st the page says: klein-4b-32st seconds a finished image, median, the single cardsComfyUI · 0.21.1 · batch 3 · 17 September 2026, 17:11 and 17:24 UTC · quantisation, context and concurrency not stated
6.297 s	250 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
5.827 s	300 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
klein-4b-32st the page says: klein-4b-32st the exhibit's own ratio, the single card against one card of the pairComfyUI · 0.21.1 · matched at 250 W · batch 3 · 2026-09-17 · quantisation, context and concurrency not stated
1.16×
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
klein-4b-8st seconds a finished image, median of two timed submitssComfyUI · 0.21.1 · batch 3 · quantisation, context, concurrency and measurement window not stated
1.683 s	250 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #renders
1.560 s	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
1.509 s	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
klein-4b-8st the page says: klein-4b-8st seconds a finished image, median, the single cardsComfyUI · 0.21.1 · batch 3 · 17 September 2026, 17:11 and 17:24 UTC · quantisation, context and concurrency not stated
1.683 s	250 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
1.560 s	300 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
klein-4b-8st the page says: klein-4b-8st the exhibit's own ratio, the single card against one card of the pairComfyUI · 0.21.1 · matched at 250 W · batch 3 · 2026-09-17 · quantisation, context and concurrency not stated
1.23×
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
klein-4b-8st-pixel4 seconds a finished image, median of two timed submitssComfyUI · 0.21.1 · batch 3 · quantisation, context, concurrency and measurement window not stated
1.721 s	250 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #renders
1.579 s	300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
1.528 s	350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #renders
klein-4b-8st-pixel4 the page says: klein-4b-8st-pixel4 seconds a finished image, median, the single cardsComfyUI · 0.21.1 · batch 3 · 17 September 2026, 17:11 and 17:24 UTC · quantisation, context and concurrency not stated
1.721 s	250 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
1.579 s	300 W
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
klein-4b-8st-pixel4 the page says: klein-4b-8st-pixel4 the exhibit's own ratio, the single card against one card of the pairComfyUI · 0.21.1 · matched at 250 W · batch 3 · 2026-09-17 · quantisation, context and concurrency not stated
1.20×
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
Mistral Small 3.2 the page says: mistral-small3.2:24b a gate-shaped call, median / p95Q4_K_M · ollama · 0.32.13 · eight tokens out · 2026-09-17 · power cap and context not stated
594 ms / 683 ms	1 caller · 4 parallel slots
from Two 3080s against one 3090, and the cap decides § The slot trap, if you serve other people read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #the-slot-trap-if-you-serve-other-people
1,179 ms / 6,011 ms	4 callers · 4 parallel slots
from Two 3080s against one 3090, and the cap decides § The slot trap, if you serve other people read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #the-slot-trap-if-you-serve-other-people
Mistral Small 3.2 the page says: mistral-small3.2:24b core temp max°CQ4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
63 °C	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
57 °C	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
Mistral Small 3.2 the page says: mistral-small3.2:24b fan max%Q4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
51 %	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
0 %	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
Mistral Small 3.2 the page says: mistral-small3.2:24b largest window held whole, and the memory it used, the layer count namedQ4_K_M · ollama · 0.32.13 · 256 tokens out · power cap, context and concurrency not stated
98,304 whole, 23,469 MiB used	2026-09-17
from Two 3080s against one 3090, and the cap decides § How much can each one hold? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
98,304 whole, 23,469 MiB	not stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b largest window held whole, placement left to the serverQ4_K_M · ollama · 0.32.13 · 256 tokens out · power cap, context, concurrency and measurement window not stated
65,536 whole, 20,589 MiB
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b largest window held whole, the layer count namedQ4_K_M · ollama · 0.32.13 · 131,072 · 256 tokens out · power cap, concurrency and measurement window not stated
Refuses to load — the allocator's own out-of-memory error, in the file
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costs read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-fits-and-what-the-cache-costs
Mistral Small 3.2 the page says: mistral-small3.2:24b residency under the server's own placementQ4_K_M · ollama · 0.32.13 · 2026-09-17 · power cap and context not stated
56.2 %	eight tokens out4 parallel slots
from Two 3080s against one 3090, and the cap decides § The slot trap, if you serve other people read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #the-slot-trap-if-you-serve-other-people
65,536 whole	256 tokens outnot stated
from Two 3080s against one 3090, and the cap decides § How much can each one hold? read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #how-much-can-each-one-hold
Mistral Small 3.2 the page says: mistral-small3.2:24b streamed as a share of the ceiling%Q4_K_M · ollama · 0.32.13 · context, concurrency and measurement window not stated
55.4 %	250 WLadder rung
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
79.6 %	300 Wsingle-window repeat
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
78.0 %	350 WLadder rung
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b the cap flag, the driver's own reportQ4_K_M · ollama · 0.32.13 · context and concurrency not stated
Active 100 %	250 WLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Active in 100 % of samples†	300 W256 tokens out2026-09-17 (14:54 to 23:46 UTC)
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 300 W (86 % of its limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Active 100 %†	300 Wsingle-window repeatnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Active in 100 %†	350 W256 tokens out2026-09-18, 00:22 to 00:54 UTC
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090, 350 W (its own limit) from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Active 100 %†	350 WLadder rungnot stated
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b the exhibit's own change from its 300 W rung to its 350 W rungQ4_K_M · ollama · 0.32.13 · 350 W · context, batch or steps, concurrency and measurement window not stated
2.0 per cent slower
the page says in the same sentence that this is not a cap effect: a single-window repeat on a cold card against a ladder rung on a card eight degrees warmer, drawing 16.5 W less at 78.1 per cent of its cap from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What a power cap takes, and what each card does at its own limit typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #what-a-power-cap-takes-and-what-each-card-does-at-its-own
Mistral Small 3.2 the page says: mistral-small3.2:24b the exhibit's own margin over the pair, denseQ4_K_M · ollama · 0.32.13 · 300 W · 65,536 · 256 tokens out · 2026-09-17 · concurrency not stated
the dense model by 7.0
the page's own sentence gives the unit once for both halves: the mixture by 18.0 per cent and the dense model by 7.0 from Two 3080s against one 3090, and the cap decides § How fast does each one write? typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-used-3080s-priced · section id #how-fast-does-each-one-write
Mistral Small 3.2 the page says: mistral-small3.2:24b the exhibit's own margin over the Ti at the shared 300 W capQ4_K_M · ollama · 0.32.13 · 300 W · 65,536 · batch or steps, concurrency and measurement window not stated
11.3 %
the page's own reading of it, in the same cell: the two rows above reverse it, and one cap step is the whole difference from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § At a glance typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #at-a-glance
Mistral Small 3.2 the page says: mistral-small3.2:24b tok/stok/sQ4_K_M · ollama · 0.32.13 · each rig's headline window · 256 tokens out · 2026-09-17 · concurrency not stated
33.3tok/s†	250 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
47.8tok/s†	300 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
mistral-nemo:12b-instruct-2407 board WWOllama · 0.32.13 · 300 W · 4,096 · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
257.59W	Q3_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
259.65W	Q4_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
259.9W	Q8_0 · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
264.11W	Q5_K_M · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
270.81W	Q6_K · cache at 8 bits
from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
mistral-nemo:12b-instruct-2407 growth minus download, MiBMiBOllama · 0.32.13 · 300 W · the 4,096-token question load · 519 tokens as this model counts it · asks for 256 tokens · one request at a time · 2026-10-15 between 05:34 and 11:01 UTC
376MiB	Q8_0 · cache at 16 bits
the page's own subtraction of the file's size from the card's growth, printed in its table from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
531MiB	Q6_K · cache at 16 bits
the page's own subtraction of the file's size from the card's growth, printed in its table from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
616MiB	Q5_K_M · cache at 16 bits
the page's own subtraction of the file's size from the card's growth, printed in its table from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
696MiB	Q4_K_M · cache at 16 bits
the page's own subtraction of the file's size from the card's growth, printed in its table from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
782MiB	Q3_K_M · cache at 16 bits
the page's own subtraction of the file's size from the card's growth, printed in its table from The quantization ladder § What each step down bought: speed and memory read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the same section · read from the-quantization-ladder · section id #speed-and-memory
not stated card board power, cards emptyW250 W · nothing generating · 2026-09-17 03:15 to 03:52 UTC · quantisation, runtime, context and concurrency not stated
25.4 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts
not stated each cap as a fraction of the board's own limitquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
250 W = 71 % · 300 W = 86 % · 350 W = 100 %
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two cards, named once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #the-two-cards-named-once
not stated fan max%ComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
58 %	250 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
67 %	350 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated fan, maximum%ComfyUI · 0.21.1 · 250 W · Ten minutes of continuous drawing · 2026-09-17 · quantisation, context and concurrency not stated
58 %
the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
not stated images a minute, sustained over ten minutesimages/minComfyUI · 0.21.1 · 250 W · Ten minutes of continuous drawing · 2026-09-17 · quantisation, context and concurrency not stated
26.87 a minute
from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image typed in from that page by hand, then checked against it word for word · the card is named in this row · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
not stated memory chips, their count and size, and where they sitquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
twenty-four 1 GB chips, twelve on the front and twelve on the back
the page's own consequence: this is the heat a person feels through a 3090's backplate, and the reason that backplate is a cooling part from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an object typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #the-card-itself-as-an-object
not stated one card idle, by its own readingW280 W · 9 September 2026 · quantisation, runtime, context, batch or steps and concurrency not stated
about 24 W each by the second card's own reading over the same day
from The cost to purchase and run a very capable home AI rig § What the rig costs to run typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from home-inference-and-diffusion-rig-cost · section id #what-the-rig-costs-to-run
not stated seconds an image, mediansComfyUI · 0.21.1 · 250 W · Ten minutes of continuous drawing · 2026-09-17 · quantisation, context and concurrency not stated
2.156s
the page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as one from Two 3080s against one 3090, and the cap decides § Renders: one 3080 against one 3090, seconds per image read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #renders-one-3080-against-one-3090-seconds-per-image
not stated the board's own default limit, and the driver's maximumquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
350 W (the driver's max_limit is 366 W)
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two cards, named once read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #the-two-cards-named-once
not stated the cap flag, the share of samples it was set in%ComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
99.9 %	250 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
100 %	350 W
the page's own caveat: this is not a cooler comparison — the rows did different amounts of work at different caps, and the image counts are printed so a reader can see that from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renders read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #renders
not stated the embedding arm's spread, against the gate it was pre-registered underone ollama build · version not stated · 2026-09-16 (UTC) · quantisation, power cap, context, batch or steps and concurrency not stated
five batches spread 34.9% around their median where the gate allowed 15%
the arm voided itself against its own pre-registered gate, and the page's own rule is that a rung rejected by its own gate gets no ratio — so this shelf carries the spread and no ratio from it from A Short History of Mistral § What a power cap actually costs typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from a-short-history-of-mistral · section id #what-a-power-cap-actually-costs
not stated the exhibit's own ratio, one card against the tworead 17 September 2026 in the same half hour · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
2.51 times
the record behind this figure is public — the listing its page prints in full, read in the window that page names from Two 3080s against one 3090, and the cap decides § What it costs, priced 17 September 2026 typed in from that page by hand, then checked against it word for word · the card is named in the same section · read from two-used-3080s-priced · section id #what-it-costs-priced-17-september-2026
not stated the exhibit's own ratio, the mixture against the dense model300 W · quantisation, runtime, context, batch or steps, concurrency and measurement window not stated
2.8 times
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The question typed in from that page by hand, then checked against it word for word · the card is named in this row · read from one-3090-against-one-3090-ti · section id #the-question
not stated thickness, by tape measureMeasured on 17 September 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
about 2 inches
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an object typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #the-card-itself-as-an-object
not stated weight, on a kitchen scalegMeasured on 17 September 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
1,203 g
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an object typed in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from one-3090-against-one-3090-ti · section id #the-card-itself-as-an-object
not stated what the model itself costs to holdW250 W · nothing generating · 2026-09-17 03:15 to 03:52 UTC · quantisation, runtime, context and concurrency not stated
27.7 W
from Two 3080s against one 3090, and the cap decides § Power, heat, and the fifty watts read off that page’s own table by the roster’s reader, then checked against it word for word · the card is named in the table’s own heading · read from two-used-3080s-priced · section id #power-heat-and-the-fifty-watts

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 the control that narrows the table to them — a number with no way back to what it counts is the thing this shelf exists against.

A 3090, in its own words — named, and no figure carried from it — yet
A harness that rewrites itself — named, and no figure carried from it — yet
A laptop, asked the desktop's questions — named, and no figure carried from it — yet
A Rig Your Friend Already Owns — 15 figures on this page
A sealed 3090 Ti FTW3 Ultra, four years after EVGA left the business — named, and no figure carried from it — yet
A Short History of Mistral — 29 figures on this page
A short history of Ollama — named, and no figure carried from it — yet
A short history of Qwen — named, and no figure carried from it — yet
A short history of vLLM — named, and no figure carried from it — yet
Teaching a Music Model Chopin in Five Minutes — 2 figures on this page
Electricity per million tokens, card by card — named, and no figure carried from it — yet
Elo ratings, from chess to image models — named, and no figure carried from it — yet
Qwen3.8-Flash-Next on an RTX 5090 laptop, an RTX PRO 6000 Blackwell and an RTX 3090 Ti: the Strata engine and llama.cpp, the numbers — named, and no figure carried from it — yet
Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench — named, and no figure carried from it — yet
FLUX 3 Image Klein, ranked blind in a game's own painted style — named, and no figure carried from it — yet
FLUX 3 Image Klein, run again with its maker's prompt upsampler — named, and no figure carried from it — yet
Four homes for one reranker: what RuleSage's slowest stage cost at each stop — named, and no figure carried from it — yet
GPUs and orchids — named, and no figure carried from it — yet
Half an Hour with Dead Composers — named, and no figure carried from it — yet
The cost to purchase and run a very capable home AI rig — 29 figures on this page
How the speed tool works — named, and no figure carried from it — yet
llm-d 1.0, and what it means if you run one or two GPUs — named, and no figure carried from it — yet
Ollama and vLLM on one RTX 3090 — 35 figures on this page
Ollama or vLLM: which to run, and how — named, and no figure carried from it — yet
RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench — 105 figures on this page
One 3090 out, one 3090 Ti in, and the cap that decided it — named, and no figure carried from it — yet
Reading the answer instead of writing it — named, and no figure carried from it — yet
RTX 5080 vs RTX 3080 Ti — named, and no figure carried from it — yet
Six graphics cards, the same questions — named, and no figure carried from it — yet
Small open deciders, measured: a word counter kept up on our hardest question — named, and no figure carried from it — yet
Ten Minutes with Living Artists — named, and no figure carried from it — yet
The 3080 Rig, Priced — named, and no figure carried from it — yet
The one-card RTX 3090 rig, repriced — named, and no figure carried from it — yet
The Ceiling Is Not the Corpus — 2 figures on this page
The cost of intelligence, as a utility — named, and no figure carried from it — yet
The Instrument Travels — named, and no figure carried from it — yet
The quantization ladder — 40 figures on this page
The same eighteen pictures, card by card — named, and no figure carried from it — yet
Two 3090s out, one 3090 Ti in — 7 figures on this page
Two Hours at 12 tok/s — named, and no figure carried from it — yet
Two 3080s against one 3090, and the cap decides — 80 figures on this page
What 150 Watts Buys — named, and no figure carried from it — yet
What it costs to run a high-end AI rig at home — named, and no figure carried from it — yet

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.

A five-year-old RTX 3090 — 24 gigabytes of memory, power-capped at 250 watts
EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)
EVGA GeForce RTX 3090 XC3 Ultra Gaming 24 GB
EVGA's own new FTW3 Ultra, one seller, one left
GeForce RTX 3090 24 GB
GeForce RTX 3090 XC3 Ultra
NVIDIA GeForce RTX 3090 (23.6 GiB)
RTX 3090
RTX 3090 (24 GB)
RTX 3090 (24 GB) — a bench instance
RTX 3090 24G (same box)
RTX 3090 24G, single GPU (consumer desktop)
Two EVGA GeForce RTX 3090 XC3 Ultras, renewed, 24 GB of graphics memory each, 48 GB between them
a 3090
a ZOTAC Trinity OC
a several-year-old consumer RTX 3090
a used RTX 3090 (24 GB)
a used RTX 3090
an RTX 3090 purchased over five years ago
an EVGA K|NGP|N Hybrid
an EVGA XC3 Ultra
one 3090
one EVGA GeForce RTX 3090 XC3 Ultra 24 GB
the 936 GB/s NVIDIA publishes for a GeForce RTX 3090
the RTX 3090 a friend bought for gaming back in 2021
the cheapest new RTX 3090 of any brand