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: bartowski/Phi-3.5-mini-instruct-GGUF
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 362.79 tok/s | 96GB | ✓ Measured |
| NVIDIA B300 | 352.25 tok/s | 288GB | ✓ Measured |
| NVIDIA H200 | 348.14 tok/s | 141GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 345.07 tok/s | 80GB | ✓ Measured |
| NVIDIA B200 | 325 tok/s | 192GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 229.46 tok/s | 80GB | ✓ Measured |
| NVIDIA L40S | 229.06 tok/s | 48GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 222.53 tok/s | 40GB | Est. |
| NVIDIA A10G | 145.97 tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 90.56 tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 72.07 tok/s | 16GB | ✓ Measured |
NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the fastest card we've measured running Phi-3.5-mini, at 362.79 tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs Phi-3.5-mini, at 72.07 tok/s.