A model that doesn't fit doesn't run slowly, it doesn't run. Here's every card from our bench, sorted by measured speed.
Weights: unsloth/gpt-oss-20b-GGUF
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 375.09 tok/s | 96GB | ✓ Measured |
| NVIDIA B300 | 355.56 tok/s | 288GB | ✓ Measured |
| NVIDIA H200 | 354.23 tok/s | 141GB | ✓ Measured |
| NVIDIA B200 | 351.42 tok/s | 192GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 346.8 tok/s | 80GB | ✓ Measured |
| NVIDIA L40S | 233.82 tok/s | 48GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 212.63 tok/s | 80GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 205.1 tok/s | 40GB | Est. |
| NVIDIA A10G | 144.61 tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 91.78 tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 63.63 tok/s | 16GB | ✓ Measured |
NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the fastest card we've measured running gpt-oss-20b, at 375.09 tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs gpt-oss-20b, at 63.63 tok/s.