Throughput with 32 requests in flight at once, which is the question when a model sits behind an API rather than on your desk.
Weights: Qwen/Qwen2.5-1.5B
Measured in serve tok/s. Longer is faster.
| # | GPU | serve tok/s |
|---|---|---|
| 1 | NVIDIA B200 | 9186.4 |
| 2 | NVIDIA H200 | 7541.6 |
| 3 | NVIDIA H100 80GB HBM3 | 6794.8 |
| 4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 5786.7 |
| 5 | NVIDIA L40S | 4686.8 |
| 6 | NVIDIA A100 80GB SXM4 | 4564.7 |
| 7 | NVIDIA A100 40GB SXM4 | 4111.1 |
| 8 | NVIDIA A10G | 2702.5 |
| 9 | NVIDIA L4 | 1827 |
| 10 | NVIDIA T4 | 1337.1 |
Weights: HuggingFaceTB/SmolLM2-1.7B
Measured in serve tok/s. Longer is faster.
| # | GPU | serve tok/s |
|---|---|---|
| 1 | NVIDIA B200 | 9672.9 |
| 2 | NVIDIA H200 | 7107.8 |
| 3 | NVIDIA H100 80GB HBM3 | 6509.9 |
| 4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 5642 |
| 5 | NVIDIA A100 80GB SXM4 | 4351.4 |
| 6 | NVIDIA A100 40GB SXM4 | 3865.2 |
| 7 | NVIDIA L40S | 3525.9 |
| 8 | NVIDIA A10G | 2242.7 |
| 9 | NVIDIA L4 | 1524.9 |
| 10 | NVIDIA T4 | 1222.5 |
Weights: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Measured in serve tok/s. Longer is faster.
| # | GPU | serve tok/s |
|---|---|---|
| 1 | NVIDIA B200 | 11562.2 |
| 2 | NVIDIA H200 | 9137.7 |
| 3 | NVIDIA H100 80GB HBM3 | 8336.7 |
| 4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 7678.7 |
| 5 | NVIDIA L40S | 6278.4 |
| 6 | NVIDIA A100 80GB SXM4 | 5701.5 |
| 7 | NVIDIA A10G | 3558.2 |
| 8 | NVIDIA A100 40GB SXM4 | 3221.3 |
| 9 | NVIDIA L4 | 2584.3 |
| 10 | NVIDIA T4 | 1897.8 |
Weights: Qwen/Qwen2.5-7B
Measured in serve tok/s. Longer is faster.
| # | GPU | serve tok/s |
|---|---|---|
| 1 | NVIDIA B200 | 5799.1 |
| 2 | NVIDIA H200 | 4394.8 |
| 3 | NVIDIA H100 80GB HBM3 | 3601.1 |
| 4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 2282.8 |
| 5 | NVIDIA A100 80GB SXM4 | 2187.2 |
| 6 | NVIDIA A100 40GB SXM4 | 1769.6 |
| 7 | NVIDIA L40S | 1406.8 |
| 8 | NVIDIA A10G | 858.7 |
| 9 | NVIDIA L4 | 469.9 |
Every number is a first-party run on our own bench. Cards missing from a board have not been measured on that model yet, or cannot fit it.