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: Qwen/Qwen2.5-1.5B
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA B300 | 22351.1 train tok/s | 288GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 16034.9 train tok/s | 80GB | ✓ Measured |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 15753.9 train tok/s | 96GB | ✓ Measured |
| NVIDIA B200 | 15322 train tok/s | 192GB | ✓ Measured |
| NVIDIA H200 | 14931.7 train tok/s | 141GB | ✓ Measured |
| NVIDIA L40S | 12120.1 train tok/s | 48GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 8252.8 train tok/s | 80GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 6349.7 train tok/s | 40GB | Est. |
| NVIDIA A10G | 4214.4 train tok/s | 24GB | ✓ Measured |
| NVIDIA L4 | 3720.9 train tok/s | 24GB | ✓ Measured |
| NVIDIA T4 | 1434.4 train tok/s | 16GB | ✓ Measured |
NVIDIA B300 is the fastest card we've measured running Qwen2.5 1.5B LoRA, at 22351.1 train tok/s.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs Qwen2.5 1.5B LoRA, at 1434.4 train tok/s.