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: ZhengPeng7/BiRefNet
| GPU | Speed | VRAM | Basis |
|---|---|---|---|
| NVIDIA H200 | 1457.73 images/min | 141GB | ✓ Measured |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 1445.92 images/min | 96GB | ✓ Measured |
| NVIDIA H100 80GB HBM3 | 1310.97 images/min | 80GB | ✓ Measured |
| NVIDIA B200 | 1082.03 images/min | 192GB | ✓ Measured |
| NVIDIA B300 | 1064.73 images/min | 288GB | ✓ Measured |
| NVIDIA A100 80GB SXM4 | 916.67 images/min | 80GB | ✓ Measured |
| NVIDIA L40S | 863.81 images/min | 48GB | ✓ Measured |
| NVIDIA A100 40GB SXM4 | 611.56 images/min | 40GB | Est. |
| NVIDIA A10G | 432.33 images/min | 24GB | ✓ Measured |
| NVIDIA L4 | 305.81 images/min | 24GB | ✓ Measured |
| NVIDIA T4 | 183.48 images/min | 16GB | ✓ Measured |
NVIDIA H200 is the fastest card we've measured running BiRefNet, at 1457.73 images/min.
NVIDIA T4 ($2,299 MSRP) is the cheapest card on our bench that runs BiRefNet, at 183.48 images/min.