strata→signal research search all research →

The hardware roster · one card

RTX 3090 Ti (24 GB)

24 GB·NVIDIA·graphics card·211 figures, from 3 exhibits · named in 17 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.

Narrow these 211 figuresevery modelmodel: FLUX.2 [klein] 4B · 13 of 211 figuresmodel: gemma4:26b · 44 of 211 figuresmodel: klein-4b-32st · 4 of 211 figuresmodel: klein-4b-8st · 4 of 211 figuresmodel: klein-4b-8st-pixel4 · 4 of 211 figuresmodel: Mistral Small 3.2 · 67 of 211 figuresmodel: Qwen3.8-Flash-Next · 27 of 211 figuresmodel: not stated · 48 of 211 figuresevery power cappower cap: 300 W · 69 of 211 figurespower cap: 350 W · 55 of 211 figurespower cap: 400 W · 28 of 211 figurespower cap: 450 W · 42 of 211 figurespower cap: not stated · 17 of 211 figuresevery exhibitexhibit: Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench · 27 of 211 figuresexhibit: RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench · 171 of 211 figuresexhibit: Two 3090s out, one 3090 Ti in · 13 of 211 figures
Narrow by model

FLUX.2 [klein] 4B, as this shelf’s pages spell itFLUX.2 klein 4BMistral Small 3.2, as this shelf’s pages spell itmistral-small3.2:24b

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

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

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.

valuewhat moves inside this group
gemma4:26b decode at the full window, and the memory it usedollama · 0.32.13 · 350 W · 131,072 · 256 tokens out · concurrency and measurement window not stated
159.410 tok/s, 21,008 MiBQ4_K_M · f16 cache
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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
146.780, 19,428 MiBQ4_K_M · q4_0
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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
147.823, 20,142 MiB†Q4_K_M · q8_0
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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
gemma4:26b decode summed across conversations, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · a 4,096-token window · 256 tokens out · measurement window not stated
145.817tok/s300 W1
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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
228.601tok/s300 W2
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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
327.124tok/s300 W4
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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
147.201tok/s350 W1
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
231.908tok/s350 W2
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
337.044tok/s350 W4
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
146.996tok/s400 W1
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
232.647tok/s400 W2
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
337.842tok/s400 W4
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
147.364tok/s450 W1
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
231.240tok/s450 W2
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
337.620tok/s450 W4
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armsread 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 #three-more-arms
gemma4:26b decode, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · 256 tokens out · concurrency not stated
145.908tok/s†300 W131,0722026-09-17 (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 Ti, 300 W (67 %)
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 limitread 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
145.908tok/s†300 Wnot statednot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
147.823tok/s†350 W131,0722026-09-17 (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 Ti, 350 W (78 %)
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 limitread 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
147.823tok/s†350 Wnot statednot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
147.599tok/s400 Wnot statednot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
147.560tok/s†450 W131,0722026-09-17 (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 Ti, 450 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 limitread 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
147.560tok/s†450 Wnot statednot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
gemma4:26b joules per 1,000 tokensJ / 1k tokQ4_K_M · ollama · 0.32.13 · 256 tokens out · context and concurrency not stated
1,648.85J / 1k tok†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 Ti, 300 W (67 %)
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 limitread 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
1,648.85J / 1k tok†300 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
1,909.96J / 1k tok†350 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 Ti, 350 W (78 %)
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 limitread 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
1,909.96 ⚑†350 Wnot stated
the page marks this cell ⚑ and explains it below the table: between 350 and 400 W the mixture drew 28.4 W less at identical core clocks and identical decode, because the fan read 49 per cent at one rung and 0 at the other. The bench flags the step and corrects nothing
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
1,723.81 ⚑400 Wnot stated
the page marks this cell ⚑ and explains it below the table: between 350 and 400 W the mixture drew 28.4 W less at identical core clocks and identical decode, because the fan read 49 per cent at one rung and 0 at the other. The bench flags the step and corrects nothing
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
1,915.13J / 1k tok†450 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 Ti, 450 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 limitread 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
1,915.13J / 1k tok†450 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
Mistral Small 3.2 the page says: mistral-small3.2:24b decode, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · concurrency not stated
42.969tok/s†300 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 Ti, 300 W (67 %)
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 limitread 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
42.969tok/s†300 Wnot stated256 tokens outnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
42.969tok/s†300 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 limitread 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
53.354tok/s†350 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 Ti, 350 W (78 %)
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 limitread 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
53.354tok/s†350 Wnot stated256 tokens outnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
53.354tok/s†350 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 limitread 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
56.936tok/s†400 Wnot stated256 tokens outnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
56.936tok/s†400 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 again in §The two models want different caps
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 limitread 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
57.396tok/s†450 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 Ti, 450 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 limitread 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
57.396tok/s†450 Wnot stated256 tokens outnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
57.396tok/s†450 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 limitread 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 joules per 1,000 tokensJ / 1k tokQ4_K_M · ollama · 0.32.13 · 256 tokens out · context and concurrency not stated
6,902.38J / 1k tok†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 Ti, 300 W (67 %)
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 limitread 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
6,902.38J / 1k tok†300 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
6,343.81J / 1k tok†350 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
6,343.81J / 1k tok†350 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 Ti, 350 W (78 %)
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 limitread 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
6,713.05J / 1k tok400 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
7,295.49J / 1k tok†450 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 Ti, 450 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 limitread 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
7,295.49J / 1k tok†450 Wnot stated
the page's own words for this table's context: both models at their headline windows. It does not print the two window sizes here, so this row does not either
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
Qwen3.8-Flash-Next writing speed, tokens a secondtok/sIQ3_S · 300 W · one request at a time · 2026-10-09 (09:32Z to 10:51Z, UTC) · context not stated
13.993tok/sllama.cpp · 3cf03257128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
17.681tok/sllama.cpp · b11514128K (128,037 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowthe third figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
81.452tok/sthe Strata engine · 0.1.39128K (128,037 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
19.620tok/sllama.cpp · 3cf0325732K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
21.257tok/sllama.cpp · b1151432K (32,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowthe second figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
92.571tok/sthe Strata engine · 0.1.3932K (32,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
22.490tok/sllama.cpp · 3cf032574K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
22.657tok/sllama.cpp · b115144K (4,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowthe first figure of three in the page's sentence, which reads them in the order 4K / 32K / 128K; beside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
98.291tok/sthe Strata engine · 0.1.394K (4,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
not stated decode once the window is full, tokens a secondtok/sQ4_K_M · ollama · 0.32.13 · 98,000 tokens in a 131,072-token window · batch or steps, concurrency and measurement window not stated
80.056tok/s350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armstyped 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
79.740tok/s450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armstyped 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/sQ4_K_M · ollama · 0.32.13 · 98,000 tokens in a 131,072-token window · batch or steps, concurrency and measurement window not stated
2,549.966 tokens a second350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armstyped 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
2,628.756tok/s450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armstyped 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

Where one condition moved, and nothing else

core temperature peak — power cap apart

The same question

  • 58 °C — 300 W Renders
  • 60 °C — 350 W Renders
  • 61 °C — 400 W Renders
  • 62 °C — 450 W Renders

Every row above is the same card and the same question. 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.

core temperature plateau — power cap apart

The same question

  • 56 °C at 0.7 min — 300 W Renders
  • 55 °C at 0.4 min — 350 W Renders
  • 59 °C at 0.3 min — 400 W Renders
  • 57 °C at 0.3 min — 450 W Renders

Pictures — seconds per render

3 rows — same rule as above.

valuewhat moves inside this group
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B the exhibit's own margin over a 3090 on the heavy plate at a shared capquantisation, runtime, context, batch or steps, concurrency and measurement window not stated
8.5 per cent300 W
both boards held at 300 W. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else
from Two 3090s out, one 3090 Ti in § Why this card, and why 350 Wtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-3090s-out-one-3090-ti-in · section id #why-this-card-and-why-350-w
9.5350 W
both boards held at 350 W, which is a 3090's own limit and 78 per cent of this Ti's. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else
from Two 3090s out, one 3090 Ti in § Why this card, and why 350 Wtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-3090s-out-one-3090-ti-in · section id #why-this-card-and-why-350-w
not stated the exhibit's own margin over a 3090 on one picture, run as they comeComfyUI · 0.21.1 · 8 steps, 512 × 512, three at a time · quantisation, power cap, context, concurrency and measurement window not stated
9.3 %
each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W — which is why this row states no cap; the page prints the same margin as 0.907× the seconds
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § At a glancetyped 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

Power, temperature, and what it costs

79 rows — same rule as above.

valuewhat moves inside this group
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B the exhibit's own margin in core temperature over a 3090 at a shared cap350 W · Ten minutes of continuous drawing · quantisation, runtime, context, concurrency and measurement window not stated
sixteen degrees cooler at the plateau and fourteen at the peak
a 55 °C plateau and a 60 °C peak against a 3090's 71 and 74, both cards held at 350 W. The page prints the companion's own caveat beside it — those rows did not do equal work. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else
from Two 3090s out, one 3090 Ti in § Why this card, and why 350 Wtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-3090s-out-one-3090-ti-in · section id #why-this-card-and-why-350-w
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
median of 18.74 W and a mean of 43.04
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 armstyped 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
gemma4:26b mean drawWQ4_K_M · ollama · 0.32.13 · 256 tokens out · 2026-09-17 (14:54 to 23:46 UTC) · context and concurrency not stated
240.58 W300 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 300 W (67 %)
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 limitread 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
282.64 W350 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 350 W (78 %)
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 limitread 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
283.01 W450 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 450 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 limitread 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 step against the rung below it in wattsQ4_K_M · ollama · 0.32.13 · 256 tokens out · context, concurrency and measurement window not stated
—300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
+1.31 %350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
−0.15 %400 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
−0.03 %450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
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
6.88300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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
6.96450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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 · 2026-09-17 (14:54 to 23:46 UTC) · context and concurrency not stated
960 MHz300 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 300 W (67 %)
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 limitread 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
1,305 MHz350 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 350 W (78 %)
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 limitread 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
1,965 MHz450 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 450 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 limitread 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 draw as a share of the cap%Q4_K_M · ollama · 0.32.13 · context and concurrency not stated
98.9 %†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 Ti, 300 W (67 %)
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 limitread 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
98.9 %†300 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 limitread 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 %†350 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 Ti, 350 W (78 %)
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 limitread 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 %†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 limitread 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
95.5 %400 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 limitread 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
93.1 %†450 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 Ti, 450 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 limitread 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
93.1 %†450 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 limitread 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 mean drawWQ4_K_M · ollama · 0.32.13 · context and concurrency not stated
296.64 W†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 Ti, 300 W (67 %)
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 limitread 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
296.64 W†300 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 limitread 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
338.05 W†350 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 Ti, 350 W (78 %)
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 limitread 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
338.05 W†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 limitread 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
382.16 W400 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 limitread 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
418.76 W†450 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 Ti, 450 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 limitread 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
418.76 W†450 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 limitread 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 · Ladder rung · context, concurrency and measurement window not stated
10,251 MHz300 W
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 limitread 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
10,251 MHz350 W
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 limitread 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
10,251 MHz400 W
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 limitread 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
10,251 MHz450 W
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 limitread 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 · Ladder rung · context, concurrency and measurement window not stated
652.2 GB/s300 W
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 limitread 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
809.8 GB/s350 W
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 limitread 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
864.1 GB/s400 W
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 limitread 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
871.1 GB/s450 W
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 limitread 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 · Ladder rung · context, concurrency and measurement window not stated
984.1 GB/s300 W
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 limitread 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
984.1 GB/s350 W
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 limitread 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
984.1 GB/s400 W
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 limitread 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
984.1 GB/s450 W
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 limitread 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 step against the rung below it in wattsQ4_K_M · ollama · 0.32.13 · 256 tokens out · context, concurrency and measurement window not stated
—300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
+24.17 %350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
+6.71 %400 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
+0.81 %450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two models want different capsread 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 #the-two-models-want-different-caps
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
2.03300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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
2.71450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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
Qwen3.8-Flash-Next prompt rate, tokens a second, the engine's clocktok/sIQ3_S · 300 W · one request at a time · 2026-10-09 (09:32Z to 10:51Z, UTC) · context not stated
226.1tok/sllama.cpp · 3cf03257128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
242.3tok/sllama.cpp · b11514128K (128,037 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
2,535.0tok/sthe Strata engine · 0.1.39128K (128,037 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
232.3tok/sllama.cpp · 3cf0325732K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
244.9tok/sllama.cpp · b1151432K (32,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
2,671.6tok/sthe Strata engine · 0.1.3932K (32,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
239.7tok/sllama.cpp · 3cf032574K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
255.2tok/sllama.cpp · b115144K (4,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
2,143.1tok/sthe Strata engine · 0.1.394K (4,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
not stated core temperature peak°CComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
58 °C300 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 § Rendersread 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
60 °C350 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 § Rendersread 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
61 °C400 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 § Rendersread 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
62 °C450 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 § Rendersread 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, concurrency and measurement window not stated
56 °C at 0.7 min300 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 § Rendersread 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
55 °C at 0.4 min350 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 § Rendersread 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
59 °C at 0.3 min400 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 § Rendersread 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
57 °C at 0.3 min450 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 § Rendersread 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 images drawn in ten minutesimagesComfyUI · 0.21.1 · Ten minutes of continuous drawing · quantisation, context, concurrency and measurement window not stated
320images300 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 § Rendersread 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
331images350 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 § Rendersread 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
347images400 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 § Rendersread 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
348images450 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 § Rendersread 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
292.0 W300 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 § Rendersread 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
316.2 W350 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 § Rendersread 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
355.4 W400 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 § Rendersread 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
357.1 W450 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 § Rendersread 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
1.95300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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
2.07450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingsread 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 card, newordered 15 September 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
$2,119.99
the record behind this figure is public — the listing its page prints in full, read in the window that page namesthe page prints the listing title it was bought under: EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming, 24G-P5-4985-KR, 24GB GDDR6X, iCX3, ARGB LED, Backplate, Free eLeash
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What it costtyped 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 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
1.811s300 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 § Rendersread 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
1.715s350 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 § Rendersread 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
1.676s400 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 § Rendersread 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
1.672s450 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 § Rendersread 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 memory clock's own reported maximumMHzquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
10,501MHz
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The questiontyped 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

Everything else this shelf measured for it

69 rows — same rule as above.

valuewhat moves inside this group
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B milliseconds a sleeve, maximumms350 W · 2026-09-18 20:42 → 20:49 · quantisation, runtime, context, batch or steps and concurrency not stated
2,125ms
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 instrumentsread 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, minimumms350 W · 2026-09-18 20:42 → 20:49 · quantisation, runtime, context, batch or steps and concurrency not stated
2,060ms
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 instrumentsread 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, maximumsComfyUI · version not stated · 350 W · a batch of three, on the same 32-step graph the open set hangs · 2026-09-18 20:41 → 22:57 · quantisation, context and concurrency not stated
15.52s
a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, 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 instrumentsread 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, medians350 W · a batch of three, on the same 32-step graph the open set hangs · quantisation, context and concurrency not stated
15.39snot stated2026-09-17
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 instrumentsread 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
15.47sComfyUI · version not stated2026-09-18 20:41 → 22:57
a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, 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 instrumentsread 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, minimumsComfyUI · version not stated · 350 W · a batch of three, on the same 32-step graph the open set hangs · 2026-09-18 20:41 → 22:57 · quantisation, context and concurrency not stated
15.42s
a live figure, read off the service's own journal rather than from a bench: the first creations drawn on the new card after the move, 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 instrumentsread 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 margin over a 3090 at the 300 W setting nearest the old machine's cap350 W · quantisation, runtime, context, batch or steps, concurrency and measurement window not stated
12 per cent
not a matched-cap comparison and the page says so in the same clause: this Ti at the 350 W the new machine runs, against a 3090 at the 300 W setting nearest the old machine's 280 W. the reading underneath is the companion bench's (one-3090-against-one-3090-ti), and this page's own arithmetic on it is published here and nowhere else
from Two 3090s out, one 3090 Ti in § Why this card, and why 350 Wtyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-3090s-out-one-3090-ti-in · section id #why-this-card-and-why-350-w
FLUX.2 [klein] 4B the page says: FLUX.2 klein 4B the exhibit's own margin over the old machine at its own cap350 W · a batch of three, on the same 32-step graph the open set hangs · quantisation, runtime, context, concurrency and measurement window not stated
11.5 to 18.1 per cent faster
a live median of seven creations on this Ti at 350 W read against the bracketed range for the old machine at 280 W, so the answer is a range: one instrument on each side, which is the join this page exists to make
from Two 3090s out, one 3090 Ti in § Before and after, on the same instrumentstyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · 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 margin under the old painter's median350 W · quantisation, runtime, context, batch or steps, concurrency and measurement window not stated
17 and 19 per cent under the old median
two sleeves against a median of sixteen on the old card at 280 W, and the page says in the same sentence that two sleeves is not a median of its own
from Two 3090s out, one 3090 Ti in § Before and after, on the same instrumentstyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · 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 spread of the live creations above the bench median350 W · a batch of three, on the same 32-step graph the open set hangs · quantisation, runtime, context, concurrency and measurement window not stated
between 0.2 and 0.9 per cent above the bench's median
the page's own reason for quoting a bench figure at all: seven creations off a live service landed within a per cent of what two timed submits predicted
from Two 3090s out, one 3090 Ti in § Before and after, on the same instrumentstyped in from that page by hand, then checked against it word for word · the card is named elsewhere on that page · read from two-3090s-out-one-3090-ti-in · section id #before-and-after-on-the-same-instruments
gemma4:26b 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
131,072 whole, 20,142 MiB
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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
gemma4:26b the cap flag, the driver's own reportQ4_K_M · ollama · 0.32.13 · 256 tokens out · 2026-09-17 (14:54 to 23:46 UTC) · context and concurrency not stated
Active in 83.3 %300 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 300 W (67 %)
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 limitread 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
0 %350 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 350 W (78 %)
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 limitread 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
0 %450 W
the cap is the source table's COLUMN HEADING and is inverted into this row's condition: the header cell reads a 3090 Ti, 450 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 limitread 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 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
+7.8 per cent on the Ti
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache coststyped 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-fits-and-what-the-cache-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
5.332 s300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
5.130 s350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
5.015 s400 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
5.015 s450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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 seconds a finished image, median of two timed submitssComfyUI · 0.21.1 · batch 3 · quantisation, context, concurrency and measurement window not stated
1.455 s300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.397 s350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.363 s400 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.369 s450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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 seconds a finished image, median of two timed submitssComfyUI · 0.21.1 · batch 3 · quantisation, context, concurrency and measurement window not stated
1.436 s300 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.379 s350 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.351 s400 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
1.350 s450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Rendersread 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
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, concurrency and measurement window not stated
98,304 whole, 23,482 MiB
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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, 20,602 MiB
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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 same error, at all four caps
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What fits, and what the cache costsread 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 streamed as a share of the ceiling%Q4_K_M · ollama · 0.32.13 · Ladder rung · context, concurrency and measurement window not stated
66.3 %300 W
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 limitread 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
82.3 %350 W
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 limitread 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
87.8 %400 W
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 limitread 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
88.5 %450 W
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 limitread 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 in 100 %†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 Ti, 300 W (67 %)
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 limitread 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 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 limitread 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 out2026-09-17 (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 Ti, 350 W (78 %)
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 limitread 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 limitread 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 100 %400 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 limitread 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 %†450 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 Ti, 450 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 limitread 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 %†450 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 limitread 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
Qwen3.8-Flash-Next first word, secondssIQ3_S · 300 W · one request at a time · 2026-10-09 (09:32Z to 10:51Z, UTC) · context not stated
566.57sllama.cpp · 3cf03257128K (128,037 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
528.72sllama.cpp · b11514128K (128,037 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
51.05sthe Strata engine · 0.1.39128K (128,037 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
138.02sllama.cpp · 3cf0325732K (32,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
130.90sllama.cpp · b1151432K (32,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
12.14sthe Strata engine · 0.1.3932K (32,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
16.87sllama.cpp · 3cf032574K (4,038 tokens) · 34 of 48 layers' experts in system memory · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
15.84sllama.cpp · b115144K (4,038 tokens) · 8 of 64 threadsoff
“neither build ran the draft head” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card; it ran last in the window and found the model's weights file already in the computer's file cache
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
1.92sthe Strata engine · 0.1.394K (4,038 tokens) · 8,144 of 24,576 experts on the cardon
“the draft head in use with a verify window of up to 6 tokens” — § The 24 GB desktop card: both engines in one windowbeside live work on the computer's other card
from Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench § The 24 GB desktop card: both engines in one windowread 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 flash-next-why-the-first-word-comes-sooner · section id #the-24-gb-desktop-card-both-engines-in-one-window
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
300 W = 67 % · 350 W = 78 % · 400 W = 89 % · 450 W = 100 %
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two cards, named onceread 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
72 %300 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 § Rendersread 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
73 %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 § Rendersread 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
75 %400 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 § Rendersread 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
75 %450 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 § Rendersread 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 memory chips, their count and size, and where they sitquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
twelve 2 GB memory chips, every one face-up beside the processor
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an objecttyped 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 the board's own default limit, and the driver's maximumquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
450 W (max_limit 480 W)
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The two cards, named onceread 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
92.9 %300 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 § Rendersread 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
87.9 %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 § Rendersread 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
16.0 %400 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 § Rendersread 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
14.5 %450 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 § Rendersread 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 difference in weight and thicknessMeasured on 17 September 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
954 g heavier, 1.79 times the weight, and about three quarters of an inch thicker
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an objecttyped 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 the exhibit's own difference in what each card costs to leave onquantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
$6 a year on the medians and $16 on the means
at 18.34 cents a kilowatt-hour, the United States residential average for June 2026 as the Energy Information Administration publishes it; board power, so a wall meter would read more for both
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Three more armstyped 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 the exhibit's own gain at 400 W over 350 W, the 32-step graphComfyUI · 0.21.1 · 400 W · quantisation, context, batch or steps, concurrency and measurement window not stated
400 W wins 2.24 per cent
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The setting this page would choose, and the dissenttyped 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-setting-this-page-would-choose-and-the-dissent
not stated the exhibit's own margin in board power over a 3090quantisation, runtime, power cap, context, batch or steps, concurrency and measurement window not stated
17.7 per cent more board power on the mixture model and 53.2 more on the dense one
each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § What the $240 buys, on those two readingstyped 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-the-240-buys-on-those-two-readings
not stated the exhibit's own ratio, its seconds against a 3090'sComfyUI · 0.21.1 · batch 3 · quantisation, power cap, context, concurrency and measurement window not stated
0.907×, 0.884× and 0.884×
each card at the setting it ships with — a 3090 at 350 W and the Ti at 450 W — over the three certified graphs, in the order the render table prints them
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § Renderstyped 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 #renders
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
3.4 times
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The questiontyped 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 thickness, by tape measureMeasured on 17 September 2026 · quantisation, runtime, power cap, context, batch or steps and concurrency not stated
about 2.75 inches
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an objecttyped 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
2,157 g
from RTX 3090 vs RTX 3090 Ti: an inference and diffusion bench § The card itself, as an objecttyped 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

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.

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.

  • EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming 24 GB
  • EVGA GeForce RTX 3090 Ti FTW3 Ultra Gaming, 24G-P5-4985-KR, 24GB GDDR6X, iCX3, ARGB LED, Backplate, Free eLeash
  • GeForce RTX 3090 Ti
  • The GeForce RTX 3090 Ti
  • ZOTAC GeForce RTX 3090 Ti AMP Extreme Holo 24 GB
  • a 3090 Ti
  • one EVGA GeForce RTX 3090 Ti FTW3 Ultra 24 GB

elsewhere in the workshop

RuleSageRealKeepambleResearchstudioprint labthe whole family →
SharpSignalKilnmemoryagencyeaselbrainkeelforge

a strata→signal property · hello@strata2signal.com · say hello