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: TheBloke/phi-2-GGUF
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 410.12 tok/s | 96GB | ✓ Measured |
| NVIDIA B300 | 356.65 tok/s | 288GB | ✓ Measured |
| NVIDIA B200 | 348.72 tok/s | 192GB | ✓ Measured |
| NVIDIA H200 | 347.08 tok/s | 141GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 340.09 tok/s | 80GB | ✓ Measured |
| NVIDIA L40S | 283.26 tok/s | 48GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 235.55 tok/s | 80GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 228.1 tok/s | 40GB | Est. |
| NVIDIA A10G | 172.45 tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 114.99 tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 89.6 tok/s | 16GB | ✓ Measured |
NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the fastest card we've measured running phi-2, at 410.12 tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs phi-2, at 89.6 tok/s.