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 B300 | 26114.3 train tok/s | 288GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 18910.3 train tok/s | 80GB | ✓ Measured |
| NVIDIA B200 | 18064.7 train tok/s | 192GB | ✓ Measured |
| NVIDIA H200 | 17650.4 train tok/s | 141GB | ✓ Measured |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 16255.4 train tok/s | 96GB | ✓ Measured |
| NVIDIA L40S | 12052.1 train tok/s | 48GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 9750.8 train tok/s | 80GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 6938.8 train tok/s | 40GB | Est. |
| NVIDIA A10G | 3668.3 train tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 3240.2 train tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 1561.9 train tok/s | 16GB | ✓ Measured |
NVIDIA B300 is the fastest card we've measured running SmolLM2 1.7B LoRA, at 26114.3 train tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs SmolLM2 1.7B LoRA, at 1561.9 train tok/s.