Chips that ran open weights, ranked by fastest quoted decode.
Host this hardware, earnEarn with a DGX Spark or DGX Station: estimate
| # | ||||
|---|---|---|---|---|
| 1 | 370 | 575W | $2,2802 used · from $2,195 | |
| 2 | 272.8 | 250W | — | |
| 3 | 264 | 1000W | — | |
| 4 | 238 | 600W | — | |
| 5 | 220 | 450W | $1,2851 used · from $1,290 | |
| 6 | 177.6 | 230W | — | |
| 7 | 158.3 | 700W | — | |
| 8 | 127 | 240W | $8,0501 used · from $7,990 | |
| 9 | 127 | 370W | — | |
| 10 | 126.5 | 70W | $3,3201 used · from $3,280 | |
| 11 | 120 | 600W | — | |
| 12 | 108.6 | 45W | — | |
| 13 | 100.6 | 350W | — | |
| 14 | 99.9 | 360W | — | |
| 15 | 92.8 | 600W | — | |
| 16 | 89.9 | 40W | — | |
| 17 | 88 | 160W | — | |
| 18 | 85.5 | 45W | $1,9101 used · from $1,890 | |
| 19 | 81.5 | 1400W | — | |
| 20 | 79 | 350W | — | |
| 21 | 66.7 | 250W | — | |
| 22 | 61.9 | 27W | — | |
| 23 | 60 | 1000W | — | |
| 24 | 57 | 250W | — | |
| 25 | 55 | 165W | — | |
| 26 | 54.7 | 180W | $3851 used · from $390 | |
| 27 | 53.8 | 140W | — | |
| 28 | 53.5 | 300W | — | |
| 29 | 52.7 | 320W | — | |
| 30 | 52 | 600W | — | |
| 31 | 50 | 170W | $1621 used · from $165 | |
| 32 | 50 | 360W | — | |
| 33 | 47 | 145W | — | |
| 34 | 40 | 125W | $781 used · from $75 | |
| 35 | 40 | 340W | — | |
| 36 | 37 | 60W | — | |
| 37 | 34.2 | 500W | — | |
| 38 | 32.4 | 150W | — | |
| 39 | 26.8 | 300W | — | |
| 40 | 11.3 | 70W | — | |
| 41 | 10 | 304W | — | |
| 42 | — | 500W | — | |
| 43 | — | 60W | — | |
| 44 | — | 27W | — | |
| 45 | — | 140W | — |
Hardware we are tracking. Not on the pool yet and no measured pairs.
30 Sep 2026: DeepSeek open-sourced DeepGEMM-Ascend, its matrix kernels for this chip, with the same API as the NVIDIA version. Not on the pool yet: no OpenAI-compatible server uses the kernels, no measured DeepSeek V4 speed exists, and the chip is only offered through Huawei Cloud in China.
