The hardware roster · one identity
RTX 3090 (24 GB)
24 GB·NVIDIA·graphics card·154 figures · 10 exhibits
Narrow by model
Narrow by power cap
The rail is CSS and nothing else — no script runs on this page. Two things it therefore cannot do, said rather than discovered: a chosen filter is not part of the address, so a narrowed view cannot be linked or bookmarked; and it cannot tell you when nothing matches, which is why the count rides on every control and why a band whose rows are all filtered out keeps its heading and empties.
Language — decode, prefill and first token
Sorted by model, then by power cap ascending — alphabetical and numeric, never by value. Every condition the page states is printed; a condition it does not state reads not stated, which is a third thing and not an uncapped card. The columns a narrow screen cannot hold are printed under the row rather than scrolled to.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 137 | decode, tokens a second | a 26-billion-parameter mixture-of-experts writer | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026, 85 receipts |
| a 26-billion-parameter mixture-of-experts writermeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026, 85 receipts | |||||
| not statednot statednot statedone question at a timehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 53 | decode, tokens a second | a dense 24-billion-parameter model | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026, 85 receipts |
| a dense 24-billion-parameter modelmeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026, 85 receipts | |||||
| not statednot statednot statedone question at a timehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 66.3 | decode per conversation, tokens a second | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out4two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 126.7 | decode summed across conversations, tokens a second | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out1two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 195.5 | decode summed across conversations, tokens a second | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out2two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 265.2 | decode summed across conversations, tokens a second | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out4two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 125.6 | decode, tokens a second | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M131,072256 tokens outnot statedthe same reading this page also prints in power-heat-and-the-fifty-wattstwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 126.12 tokens a second | decode, tokens a second | gemma4:26b | 250 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| gemma4:26b250 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot statedthe same reading this page also prints in what-a-power-cap-actually-costsa-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1,832 | joules per 1,000 tokens | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 72.1 | decode per conversation, tokens a second | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out4two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 134.8 | decode summed across conversations, tokens a second | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out1two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 210.4 | decode summed across conversations, tokens a second | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out2two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 288.4 | decode summed across conversations, tokens a second | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens out4two-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 136.1 | decode, tokens a second | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M131,072256 tokens outnot statedthe same reading this page also prints in power-heat-and-the-fifty-wattstwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1,772 | joules per 1,000 tokens | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 133.6 | decode with an unquantised cache, tokens a second | gemma4:26b | the lower cap | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bthe lower capollama · 0.32.132026-09-17 | |||||
| Q4_K_M · f16131,072-token window256 tokens outnot statedthe same reading this page also prints in what-the-cache-type-coststwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 129.4 tok/s, 100 % on the cards | decode and residency out of the box, the server's own placement | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens outnot statedtwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 133.6 tok/s, 20,995 MiB | decode at the full window, and the memory it used | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_M · f16, unquantisedits full 131,072-token window256 tokens outnot statedthe same reading this page also prints in how-much-can-each-one-holdtwo-used-3080s-priced · what-the-cache-type-costs · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 124.7 tok/s, 19,415 MiB | decode at the full window, and the memory it used | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_M · q4_0its full 131,072-token window256 tokens outnot statedtwo-used-3080s-priced · what-the-cache-type-costs · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 125.6 tok/s, 20,129 MiB | decode at the full window, and the memory it used | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_M · q8_0, the setting both rigs are measured atits full 131,072-token window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · what-the-cache-type-costs · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 127.1 / 195.8 / 261.4 tokens a second | aggregate decode at one, two and four streams | gemma4:26b (MoE) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statedone, two and four simultaneous streamsa-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 126.12 | decode, tokens a second | gemma4:26b (MoE) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedthe same reading this page also prints in what-we-measured-on-two-cards-in-one-boxa-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +8.2% | decode against the 250 W rung | gemma4:26b (MoE) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 136.49 | decode, tokens a second | gemma4:26b (MoE) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 33.1 | decode, tokens a second | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M49,152256 tokens outnot statedtwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 33.3 | decode, tokens a second | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M65,536256 tokens outnot statedthe same reading this page also prints in power-heat-and-the-fifty-wattstwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 34.13 tokens a second | decode, tokens a second | mistral-small3.2:24b | 250 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b250 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| Q4_K_Mnot stated256 tokens outnot statedthe same reading this page also prints in what-a-power-cap-actually-costsa-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 7,076 | joules per 1,000 tokens | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 47.8 | decode, tokens a second | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M65,536256 tokens outnot statedthe same reading this page also prints in power-heat-and-the-fifty-wattstwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 6,059 | joules per 1,000 tokens | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| — | decode against the 250 W rung | mistral-small3.2:24b (dense) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedthe baseline rung — the page prints an em dash, not a figurea-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 34.13 | decode, tokens a second | mistral-small3.2:24b (dense) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedthe same reading this page also prints in what-we-measured-on-two-cards-in-one-boxa-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +41.2% | decode against the 250 W rung | mistral-small3.2:24b (dense) | 300 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)300 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 48.18 | decode, tokens a second | mistral-small3.2:24b (dense) | 300 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)300 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| +46.6% | decode against the 250 W rung | mistral-small3.2:24b (dense) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 50.02 | decode, tokens a second | mistral-small3.2:24b (dense) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 98 ms / 167 ms | first token on a prompt it has already seen, dense / MoE | mistral-small3.2:24b, gemma4:26b | 250 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b250 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 72.8 | decode once the window is full, tokens a second | not stated | 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_MOne 131,072-token window with about 98,000 tokens really in it256 tokens outnot statedtwo-used-3080s-priced · what-about-a-long-document · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 45.12 s | reading 98,024 tokens, seconds | not stated | 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_MOne 131,072-token window with about 98,000 tokens really in it256 tokens outnot statedtwo-used-3080s-priced · what-about-a-long-document · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 46.64 s | time to the first token, seconds | not stated | 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_MOne 131,072-token window with about 98,000 tokens really in it256 tokens outnot statedtwo-used-3080s-priced · what-about-a-long-document · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| about 3,600 | prompt reading, tokens a second | not stated | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026 |
| not statedmeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026 | |||||
| not statednot stated2,667 tokens in 0.73 snot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 1.01 s | time to the first token | not stated | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026 |
| not statedmeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026 | |||||
| not statednot stateda 2,000-token promptone question at a timehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 0.08 s | time to the first token | not stated | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026 |
| not statedmeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026 | |||||
| not statednot stateda 40-token promptone question at a timehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 0.30 s | time to the first token | not stated | measured before this workshop capped its cards | the model server of that day · version not stated | 24 August 2026 |
| not statedmeasured before this workshop capped its cardsthe model server of that day · version not stated24 August 2026 | |||||
| not statednot stateda 400-token promptone question at a timehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 71.48 | decode, tokens a second, split, seats stopped | qwen3.5:122b, 10 billion active per token, 75.7 GiB | a 3090 at 250 W, the workstation card at 420 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| qwen3.5:122b, 10 billion active per token, 75.7 GiBa 3090 at 250 W, the workstation card at 420 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 38.22 | decode, tokens a second, split, seats up | qwen3.5:122b, 10 billion active per token, 75.7 GiB | a 3090 at 250 W, the workstation card at 420 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| qwen3.5:122b, 10 billion active per token, 75.7 GiBa 3090 at 250 W, the workstation card at 420 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 328 | aggregate decode, eight questions at once | the same 26-billion writer at 4-bit | measured before this workshop capped its cards | a batching runtime · version not stated | 22 August 2026 |
| the same 26-billion writer at 4-bitmeasured before this workshop capped its cardsa batching runtime · version not stated22 August 2026 | |||||
| 4-bitnot statednot statedeight questions at oncehome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
The same card, the same question, power cap apart
decode, tokens a second
- 34.13 — 250 W a-short-history-of-mistral · what-a-power-cap-actually-costs
- 48.18 — 300 W a-short-history-of-mistral · what-a-power-cap-actually-costs
- 50.02 — 350 W (the board's default) a-short-history-of-mistral · what-a-power-cap-actually-costs
Every row above is the same card and the same mistral-small3.2:24b (dense). The only thing that moved is the power cap. That is why nothing on this page is ranked, and why a condition is part of a row rather than a heading above one.
The same card, the same question, power cap apart
energy per 1,000 tokens
- 6,680 J — 250 W a-short-history-of-mistral · what-a-power-cap-actually-costs
- 5,673 J — 300 W a-short-history-of-mistral · what-a-power-cap-actually-costs
- 5,318 J — 350 W (the board's default) a-short-history-of-mistral · what-a-power-cap-actually-costs
Every row above is the same card and the same mistral-small3.2:24b (dense). The only thing that moved is the power cap. That is why nothing on this page is ranked, and why a condition is part of a row rather than a heading above one.
Pictures — seconds per render
28 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 2.86 | seconds per render, median of its cell | FLUX.2 klein 4B | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| FLUX.2 klein 4B320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated1,024, 4 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 1.65 | seconds per render, median of its cell | FLUX.2 klein 4B | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| FLUX.2 klein 4B320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated768, 4 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 0.88 | seconds per render, median of its cell | FLUX.2 klein 4B, the sketch model | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| FLUX.2 klein 4B, the sketch model320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated512, 4 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 2.86 s (2.84–3.17) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | not stated |
| FLUX.2 klein-4Bnot statednot statednot stated | |||||
| not statednot stated1024×1024, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 0.88 s (0.88–1.19) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | not stated |
| FLUX.2 klein-4Bnot statednot statednot stated | |||||
| not statednot stated512×512, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.65 s (1.64–1.96) | seconds per render, median (min–max) | FLUX.2 klein-4B | not stated | not stated | not stated |
| FLUX.2 klein-4Bnot statednot statednot stated | |||||
| not statednot stated768×768, steps held at 4, six renders per cellnot stateda-rig-your-friend-already-owns · the-sketch-ladder-the-number-the-game-lives-on · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.65 s | median seconds per render, mixed resolutions | FLUX.2 klein-4B (our sketch artist) | not stated | not stated | not stated |
| FLUX.2 klein-4B (our sketch artist)not statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 3.4× | the exhibit's own multiplier against the 96 GB card | FLUX.2 klein-4B (our sketch artist) | not stated | not stated | not stated |
| FLUX.2 klein-4B (our sketch artist)not statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
| one 3090, 2.17× | the exhibit's own multiplier, out of the box, against the pair | gemma4:26b | not stated | ollama · 0.32.13 | not stated |
| gemma4:26bnot statedollama · 0.32.13not stated | |||||
| Q4_K_Ma 4,096 window256 tokens outnot statedthe cap is not stated because the page states it is not one cap: the cell's own words are that each rig was measured at one cap only here, and they are not the same captwo-used-3080s-priced · at-a-glance · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 11.62 | seconds per render, median of its cell | hidream-o1-dev | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| hidream-o1-dev320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated1,024, 32 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 6.19 | seconds per render, median of its cell | hidream-o1-dev | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| hidream-o1-dev320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated512, 32 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 7.34 s | median seconds per render, mixed resolutions | hidream-o1-dev | not stated | not stated | not stated |
| hidream-o1-devnot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| † | the exhibit's own multiplier against the 96 GB card | hidream-o1-dev | not stated | not stated | not stated |
| hidream-o1-devnot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 49.89 | seconds per render, median of its cell | kandinsky5-lite | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| kandinsky5-lite320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated1,024, 32 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 11.43 | seconds per render, median of its cell | kandinsky5-lite | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| kandinsky5-lite320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated512, 32 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 28.95 s | median seconds per render, mixed resolutions | kandinsky5-lite | not stated | not stated | not stated |
| kandinsky5-litenot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 4.7× | the exhibit's own multiplier against the 96 GB card | kandinsky5-lite | not stated | not stated | not stated |
| kandinsky5-litenot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 16.09 | seconds per render, median of its cell | krea2-turbo | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| krea2-turbo320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated1,024, 8 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 5.34 | seconds per render, median of its cell | krea2-turbo | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| krea2-turbo320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated512, 8 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 9.62 s | median seconds per render, mixed resolutions | krea2-turbo | not stated | not stated | not stated |
| krea2-turbonot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 4.9× | the exhibit's own multiplier against the 96 GB card | krea2-turbo | not stated | not stated | not stated |
| krea2-turbonot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 33.01 s | median seconds per render, mixed resolutions | z-image-base (undistilled, 768 only) | not stated | not stated | not stated |
| z-image-base (undistilled, 768 only)not statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 4.2× | the exhibit's own multiplier against the 96 GB card | z-image-base (undistilled, 768 only) | not stated | not stated | not stated |
| z-image-base (undistilled, 768 only)not statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 8.08 | seconds per render, median of its cell | z-image-turbo | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| z-image-turbo320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated1,024, 8 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 2.10 | seconds per render, median of its cell | z-image-turbo | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| z-image-turbo320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated512, 8 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 4.44 | seconds per render, median of its cell | z-image-turbo | 320 W at the wall that night | not stated | the night of 26 August 2026 |
| z-image-turbo320 W at the wall that nightnot statedthe night of 26 August 2026 | |||||
| not statednot stated768, 8 stepsnot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 4.44 s | median seconds per render, mixed resolutions | z-image-turbo | not stated | not stated | not stated |
| z-image-turbonot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot stateda-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 3.9× | the exhibit's own multiplier against the 96 GB card | z-image-turbo | not stated | not stated | not stated |
| z-image-turbonot statednot statednot stated | |||||
| not statednot statedidentical mixed resolution-and-step sets on every rignot statedthe exhibit prints this ratio; the roster computes nonea-rig-your-friend-already-owns · the-table-that-answers-the-question · carried by hand, proved against that page’s own row · identity from the same section | |||||
Training
3 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 5.7 GiB | peak memory, the trainer's own allocator | ACE-Step 1.5 | 250 W | Side-Step · version not stated | not stated |
| ACE-Step 1.5250 WSide-Step · version not statednot stated | |||||
| not statednot stated331 tracks five times over, 415 small corrections in allnot statedthe-ceiling-is-not-the-corpus · forty-minutes-on-a-card-bought-over-five-years-ago · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 5.73 | seconds per optimiser step | ACE-Step 1.5 | 250 W | Side-Step · version not stated | not stated |
| ACE-Step 1.5250 WSide-Step · version not statednot stated | |||||
| not statednot stated331 tracks five times over, 415 small corrections in allnot statedthe-ceiling-is-not-the-corpus · forty-minutes-on-a-card-bought-over-five-years-ago · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 5.7 gigabytes by the trainer's own bookkeeping; the card itself showed 6.8 in use | peak memory, two instruments | ACE-Step 1.5 | 250 watts | Side-Step · version not stated | not stated |
| ACE-Step 1.5250 wattsSide-Step · version not statednot stated | |||||
| not statednot statedten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1)not statedchopin-in-five-minutes · five-minutes-of-training-honestly-accounted · carried by hand, proved against that page’s own row · identity from the same section | |||||
Power, temperature, and what it costs
32 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 4 minutes 49 seconds by the trainer's own clock | wall clock | ACE-Step 1.5 | 250 watts | Side-Step · version not stated | not stated |
| ACE-Step 1.5250 wattsSide-Step · version not statednot stated | |||||
| not statednot statedten passes over the nineteen recordings, fifty in all, LoRA rank 64 (alpha 128, dropout 0.1)not statedchopin-in-five-minutes · five-minutes-of-training-honestly-accounted · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 27 minutes 48 seconds | wall-clock, three composers | ACE-Step 1.5, LoRA adapters, three composers | 250 W | not stated | training for three composers |
| ACE-Step 1.5, LoRA adapters, three composers250 Wnot statedtraining for three composers | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · what-the-card-is-like-for-training · carried by hand, proved against that page’s own row · identity from the same section | |||||
| $11.50 per token a second | dollars per token a second, at what was paid | gemma4:26b | 250 W | not stated | bought for $1,449.99 in May 2026 |
| gemma4:26b250 Wnot statedbought for $1,449.99 in May 2026 | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · is-it-cheaper-to-run-it-yourself · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 230 W | mean draw | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 242 W | mean draw | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 206.4 W | draw while decoding | gemma4:26b (MoE) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 1,632 J | energy per 1,000 tokens | gemma4:26b (MoE) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 1,823 J | energy per 1,000 tokens | gemma4:26b (MoE) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| gemma4:26b (MoE)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 235 W | mean draw | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 43.8 per cent | what fifty watts are worth, dense | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot stated256 tokens outnot statedthe same reading this page also prints in power-heat-and-the-fifty-wattstwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| $0.29 per million | dollars per million output tokens, electricity only | mistral-small3.2:24b | 300 W | not stated | 18.34 cents a kilowatt-hour, the US residential average for June 2026 |
| mistral-small3.2:24b300 Wnot stated18.34 cents a kilowatt-hour, the US residential average for June 2026 | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · is-it-cheaper-to-run-it-yourself · carried by hand, proved against that page’s own row · identity from the same section | |||||
| $30.10 | dollars per token a second, at what was paid | mistral-small3.2:24b | 300 W | not stated | bought for $1,449.99 in May 2026 |
| mistral-small3.2:24b300 Wnot statedbought for $1,449.99 in May 2026 | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · is-it-cheaper-to-run-it-yourself · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 290 W | mean draw | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 228.2 W — 91% of cap — clock pulled to 495–675 MHz | draw while decoding, and the clock | mistral-small3.2:24b (dense) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 6,680 J | energy per 1,000 tokens | mistral-small3.2:24b (dense) | 250 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)250 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 5,673 J | energy per 1,000 tokens | mistral-small3.2:24b (dense) | 300 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)300 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 5,318 J | energy per 1,000 tokens | mistral-small3.2:24b (dense) | 350 W (the board's default) | one ollama build · version not stated | 2026-09-16 (UTC) |
| mistral-small3.2:24b (dense)350 W (the board's default)one ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot stateda-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 6,680 J / 1,632 J | energy per 1,000 tokens, dense / MoE | mistral-small3.2:24b, gemma4:26b | 250 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b250 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 228.2 W / 206.4 W | watts while decoding, dense / MoE | mistral-small3.2:24b, gemma4:26b | 250 W | one ollama build · version not stated | the small hours of 2026-09-16 (UTC) |
| mistral-small3.2:24b, gemma4:26b250 Wone ollama build · version not statedthe small hours of 2026-09-16 (UTC) | |||||
| not statednot stated256 tokens outnot stateda-short-history-of-mistral · what-we-measured-on-two-cards-in-one-box · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1,575 | core clock while drawing, median | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 67 °C | core temperature peak | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedthe page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as onetwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 65 °C at 0.5 min | core temperature plateau | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedthe page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as onetwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 269 | images drawn in ten minutes | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedthe page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as onetwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 43.8 per cent on the dense model and 8.4 on the mixture | what fifty watts are worth to the single card | not stated | 250 W · 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated250 W · 300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 3 per cent of writing speed and 6 per cent of prompt reading | what the house cap costs | not stated | 280 W, 80 per cent of the card's 350 W rating | not stated | 23 August 2026 |
| not stated280 W, 80 per cent of the card's 350 W ratingnot stated23 August 2026 | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · what-kind-of-performance-to-expect · carried by hand, proved against that page’s own row · identity from the same section | |||||
| 3,902 | energy at depth, joules per 1,000 tokens | not stated | 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_MOne 131,072-token window with about 98,000 tokens really in it256 tokens outnot statedtwo-used-3080s-priced · what-about-a-long-document · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 40.8 per cent less per thousand tokens on the mixture at 300 W and 46.8 per cent less on the dense model | the exhibit's own energy saving against the pair | not stated | 300 W | ollama · 0.32.13 | 2026-09-17 |
| not stated300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the same page | |||||
| $1,449.99 | price, one card | not stated | not stated | not stated | paid by this workshop, May 2026 |
| not statednot statednot statedpaid by this workshop, May 2026 | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · the-cards-in-may-and-in-september · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| $1,829.99 | price, one card | not stated | not stated | not stated | 13 September 2026 |
| not statednot statednot stated13 September 2026 | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · the-cards-in-may-and-in-september · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| $1,849.99 | price, one card | not stated | not stated | not stated | 13 September 2026 |
| not statednot statednot stated13 September 2026 | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · the-cards-in-may-and-in-september · carried by hand, proved against that page’s own row · identity from the same section | |||||
| $908.00 | price, one card | not stated | not stated | not stated | the past year's average completed sale |
| not statednot statednot statedthe past year's average completed sale | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · the-cards-in-may-and-in-september · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| $1,879.99 | price, one card, renewed | not stated | not stated | not stated | read 17 September 2026 in the same half hour |
| not statednot statednot statedread 17 September 2026 in the same half hour | |||||
| not statednot statednot statednot statedthe record behind this figure is public — the listing its page prints in full, read in the window that page namesread by hand at the store page, because that retailer serves no price to this workshop's fetcher; it replaces an earlier figure this workshop published on 13 September that an operator ruled inaccurate and strucktwo-used-3080s-priced · what-it-costs-priced-17-september-2026 · carried by hand, proved against that page’s own row · identity from the row itself | |||||
Memory — what fits, and at which context
2 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 53.1 W | card board power, the model resident | not stated | 250 W | not stated | 2026-09-17 03:15 to 03:52 UTC |
| not stated250 Wnot stated2026-09-17 03:15 to 03:52 UTC | |||||
| not statednot statednothing generatingnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 39.5% | the fraction of a 122B model on card, and aggregate decode, seats up | qwen3.5:122b, 10 billion active per token, 75.7 GiB | a 3090 at 250 W, the workstation card at 420 W | one ollama build · version not stated | 2026-09-16 (UTC) |
| qwen3.5:122b, 10 billion active per token, 75.7 GiBa 3090 at 250 W, the workstation card at 420 Wone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedsplit across the 3090 and the workstation card; the page's own wordsa-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
Everything else this shelf measured for it
42 rows — same rule as above.
| value | what was measured | model | power cap | runtime | measurement window |
|---|---|---|---|---|---|
| 6.297 s | seconds a finished image, median, the single card | klein-4b-32st | 250 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-32st250 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.16× | the exhibit's own ratio, the single card against one card of the pair | klein-4b-32st | matched at 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| klein-4b-32stmatched at 250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 5.827 s | seconds a finished image, median, the single card | klein-4b-32st | 300 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-32st300 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.721 s | seconds a finished image, median, the single card | klein-4b-8st-pixel4 | 250 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-8st-pixel4250 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.20× | the exhibit's own ratio, the single card against one card of the pair | klein-4b-8st-pixel4 | matched at 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| klein-4b-8st-pixel4matched at 250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 1.579 s | seconds a finished image, median, the single card | klein-4b-8st-pixel4 | 300 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-8st-pixel4300 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.683 s | seconds a finished image, median, the single card | klein-4b-8st | 250 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-8st250 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 1.23× | the exhibit's own ratio, the single card against one card of the pair | klein-4b-8st | matched at 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| klein-4b-8stmatched at 250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 1.560 s | seconds a finished image, median, the single card | klein-4b-8st | 300 W | ComfyUI · 0.21.1 | 17 September 2026, 17:11 and 17:24 UTC |
| klein-4b-8st300 WComfyUI · 0.21.117 September 2026, 17:11 and 17:24 UTC | |||||
| not statednot statedbatch 3not statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 63 °C | core temp max | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 42 % | fan max | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 9.4 per cent | the exhibit's own margin over the pair, mixture | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M131,072256 tokens outnot statedtwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 125.6 | tok/s | gemma4:26b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 47 °C | core temp max | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 0 % | fan max | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 18.0 per cent | the exhibit's own margin over the pair, mixture | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M131,072256 tokens outnot statedtwo-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 7.8 per cent | the exhibit's own margin over the pair, one caller and four | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Ma 4,096 window256 tokens outnot statedtwo-used-3080s-priced · the-full-concurrency-ladder · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 136.1 | tok/s | gemma4:26b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 131,072 whole, 20,129 MiB used | largest window held whole, and the memory it used, the layer count named | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot stated256 tokens outnot statedtwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 131,072 whole | largest window held whole, the server's own placement | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot stated256 tokens outnot statedtwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 20,995 MiB, which is 515 more than two 10 GB cards have between them | what an unquantised cache needs at the full window | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_M · f16131,072-token windownot statednot statedthe same reading this page also prints in what-the-cache-type-coststwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the same page | |||||
| about 6 per cent | what compressing the cache to 8 bits costs | gemma4:26b | not stated | ollama · 0.32.13 | 2026-09-17 |
| gemma4:26bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_M · q8_0its full 131,072-token windownot statednot statedtwo-used-3080s-priced · what-the-cache-type-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 63 °C | core temp max | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 51 % | fan max | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 33.3 | tok/s | mistral-small3.2:24b | 250 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b250 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 57 °C | core temp max | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 0 % | fan max | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| the dense model by 7.0 | the exhibit's own margin over the pair, dense | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_M65,536256 tokens outnot statedthe page's own sentence gives the unit once for both halves: the mixture by 18.0 per cent and the dense model by 7.0two-used-3080s-priced · how-fast-does-each-one-write · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 47.8 | tok/s | mistral-small3.2:24b | 300 W | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24b300 Wollama · 0.32.132026-09-17 | |||||
| Q4_K_Meach rig's headline window256 tokens outnot statedthe same reading this page also prints in how-fast-does-each-one-writetwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 594 ms / 683 ms | a gate-shaped call, median / p95 | mistral-small3.2:24b | not stated | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot statedeight tokens out1 caller · 4 parallel slotstwo-used-3080s-priced · the-slot-trap-if-you-serve-other-people · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 1,179 ms / 6,011 ms | a gate-shaped call, median / p95 | mistral-small3.2:24b | not stated | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot statedeight tokens out4 callers · 4 parallel slotstwo-used-3080s-priced · the-slot-trap-if-you-serve-other-people · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 98,304 whole, 23,469 MiB used | largest window held whole, and the memory it used, the layer count named | mistral-small3.2:24b | not stated | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot stated256 tokens outnot statedtwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 56.2 % | residency under the server's own placement | mistral-small3.2:24b | not stated | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot statedeight tokens out4 parallel slotstwo-used-3080s-priced · the-slot-trap-if-you-serve-other-people · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 65,536 whole | residency under the server's own placement | mistral-small3.2:24b | not stated | ollama · 0.32.13 | 2026-09-17 |
| mistral-small3.2:24bnot statedollama · 0.32.132026-09-17 | |||||
| Q4_K_Mnot stated256 tokens outnot statedtwo-used-3080s-priced · how-much-can-each-one-hold · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 25.4 W | card board power, cards empty | not stated | 250 W | not stated | 2026-09-17 03:15 to 03:52 UTC |
| not stated250 Wnot stated2026-09-17 03:15 to 03:52 UTC | |||||
| not statednot statednothing generatingnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 58 % | fan, maximum | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedthe page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as onetwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 26.87 a minute | images a minute, sustained over ten minutes | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedtwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the row itself | |||||
| 2.156 | seconds an image, median | not stated | 250 W | ComfyUI · 0.21.1 | 2026-09-17 |
| not stated250 WComfyUI · 0.21.12026-09-17 | |||||
| not statednot statedTen minutes of continuous drawingnot statedthe page's own caveat: the two cards did different amounts of work and the single card began the run 11 °C hotter, so this is not a cooler comparison and should not be read as onetwo-used-3080s-priced · renders-one-3080-against-one-3090-seconds-per-image · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| 27.7 W | what the model itself costs to hold | not stated | 250 W | not stated | 2026-09-17 03:15 to 03:52 UTC |
| not stated250 Wnot stated2026-09-17 03:15 to 03:52 UTC | |||||
| not statednot statednothing generatingnot statedtwo-used-3080s-priced · power-heat-and-the-fifty-watts · carried by hand, proved against that page’s own row · identity from the table’s own column head | |||||
| about 24 W each by the second card's own reading over the same day | one card idle, by its own reading | not stated | 280 W | not stated | 9 September 2026 |
| not stated280 Wnot stated9 September 2026 | |||||
| not statednot statednot statednot statedhome-inference-and-diffusion-rig-cost · what-the-rig-costs-to-run · carried by hand, proved against that page’s own row · identity from the same page | |||||
| five batches spread 34.9% around their median where the gate allowed 15% | the embedding arm's spread, against the gate it was pre-registered under | not stated | not stated | one ollama build · version not stated | 2026-09-16 (UTC) |
| not statednot statedone ollama build · version not stated2026-09-16 (UTC) | |||||
| not statednot statednot statednot statedthe arm voided itself against its own pre-registered gate, and the page's own rule is that a rung rejected by its own gate gets no ratio — so this shelf carries the spread and no ratio from ita-short-history-of-mistral · what-a-power-cap-actually-costs · carried by hand, proved against that page’s own row · identity from the same page | |||||
| 2.51 times | the exhibit's own ratio, one card against the two | not stated | not stated | not stated | read 17 September 2026 in the same half hour |
| not statednot statednot statedread 17 September 2026 in the same half hour | |||||
| not statednot statednot statednot statedthe record behind this figure is public — the listing its page prints in full, read in the window that page namestwo-used-3080s-priced · what-it-costs-priced-17-september-2026 · carried by hand, proved against that page’s own row · identity from the same section | |||||
Where it appears
Every released exhibit that names this identity, in the shelf’s own order. A count beside an exhibit is how many of its figures are carried above — and it links them, because a number with no way back to what it counts is the thing this shelf exists against.
- A Rig Your Friend Already Owns — 15 figures on this page
- A Short History of Mistral — 29 figures on this page
- Teaching a Music Model Chopin in Five Minutes — 2 figures on this page
- Half an Hour with Dead Composers — named, and no figure carried from it — yet
- The cost to purchase and run a very capable home AI rig — 26 figures on this page
- Ten Minutes with Living Artists — named, and no figure carried from it — yet
- The Ceiling Is Not the Corpus — 2 figures on this page
- The Instrument Travels — named, and no figure carried from it — yet
- Two 3080s against one 3090, and the cap decides — 80 figures on this page
- What 150 Watts Buys — named, and no figure carried from it — yet
How this one is named here
The canonical name above is the full product name. Everything below is a spelling a published page or kit has actually used for it, kept so a reader who searches an older page’s words lands on this row. The set is closed and historical: no page cut after 2026-09-17 writes a memory class where it can write a card.
- A five-year-old RTX 3090 — 24 gigabytes of memory, power-capped at 250 watts
- EVGA 24G-P5-3975-KR GeForce RTX 3090 XC3 Ultra Gaming, 24GB GDDR6X, iCX3 Cooling, ARGB LED, Metal Backplate (Renewed)
- EVGA's own new FTW3 Ultra, one seller, one left
- GeForce RTX 3090 24 GB
- GeForce RTX 3090 XC3 Ultra
- NVIDIA GeForce RTX 3090 (23.6 GiB)
- RTX 3090
- RTX 3090 (24 GB)
- RTX 3090 (24 GB) — a bench instance
- RTX 3090 24G (same box)
- RTX 3090 24G, single GPU (consumer desktop)
- Two EVGA GeForce RTX 3090 XC3 Ultras, renewed, 24 GB of graphics memory each, 48 GB between them
- a ZOTAC Trinity OC
- a several-year-old consumer RTX 3090
- a used RTX 3090 (24 GB)
- a used RTX 3090
- an RTX 3090 purchased over five years ago
- an EVGA K|NGP|N Hybrid
- an EVGA XC3 Ultra
- one 3090
- the 936 GB/s NVIDIA publishes for a GeForce RTX 3090
- the RTX 3090 a friend bought for gaming back in 2021
- the cheapest new RTX 3090 of any brand