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: HuggingFaceTB/SmolLM2-1.7B
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA B200 | 9672.9 serve tok/s | 192GB | ✓ Measured |
| NVIDIA H200 | 7107.8 serve tok/s | 141GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 6509.9 serve tok/s | 80GB | ✓ Measured |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 5642 serve tok/s | 96GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 4351.4 serve tok/s | 80GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 3865.2 serve tok/s | 40GB | Est. |
| NVIDIA L40S | 3525.9 serve tok/s | 48GB | ✓ Measured |
| NVIDIA A10G | 2242.7 serve tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 1524.9 serve tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 1222.5 serve tok/s | 16GB | ✓ Measured |
NVIDIA B200 is the fastest card we've measured running SmolLM2 1.7B served, at 9672.9 serve tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs SmolLM2 1.7B served, at 1222.5 serve tok/s.