Best Elixir runs per hardware.
This board groups model benchmarking runs by hardware, i.e. the exact GPUs, PCIe lanes and RAM they run on, so you can pick the model that best matches the hardware you have. Split mode (tensor or layer) depends on the PCIe topology of the motherboard, if more than one GPU is involved.
PCIe 3.0 x16
| Model | Quant | Wall | PP | TG | Needle | Quality |
|---|---|---|---|---|---|---|
| Ornith-1.5-35B-A3B MiniPlus V2.1 | APEX-I MiniPlus V2.1 | 1227.33 s | 819.65 | 57.92 | hit | 2/5 |
PCIe 4.0 x16 (CPU) + 4.0 x4 (PCH)
| Model | Quant | Wall | PP | TG | Needle | Quality |
|---|---|---|---|---|---|---|
| Groot-19B-A3B-261023 MTP | Q5_K_M | 227.7 s | 2496 | 80.2 | hit | 2/5 |
PCIe 5.0 x8/x8
| Model | Quant | Wall | PP | TG | Needle | Quality |
|---|---|---|---|---|---|---|
| Nemotron-3.5-Bolt-74B-A9B IQ2 QAT | IQ2 QAT | 284.8 s | 1280 | 107.9 | hit | 4/5 |
| Muse-Glimmer NVFP4 | NVFP4 | 113.4 s | 1224 | 74.9 | hit | 4/5 |
| Ornith 1.5 35B-A3B NVFP4 | NVFP4 | 178 s | 2340 | 112 | hit | 4/5 |
| flint NVFP4 | NVFP4 | 371.5 s | 815 | 84.9 | cliff | 2/5 |
| LFM2.5-24B-A2B Q8_0 | Q8_0 | 135.2 s | 1403 | 151 | hit | 2/5 |
A paragraph about what you are working on is enough to start.