Qwen 3.5 397B-A17B — B200 vs H100
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and H100 (NVIDIA Hopper) on Qwen 3.5 397B-A17B. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
At 56 tok/s/user interactivity on Qwen 3.5 397B-A17B, B200 delivers 9325 tok/s/chip at $0.05 per million tokens; H100 delivers 2115 tok/s/chip at $0.15. B200 is 198% cheaper per token; B200 delivers 341% more tok/s/chip at this point.
B200 posts 6764 tok/s/chip for $0.07 per million tokens at 102 tok/s/user on Qwen 3.5 397B-A17B; H100 posts 1541 tok/s/chip for $0.21. B200 is 197% cheaper per token; B200 delivers 339% more tok/s/chip.
Throughput at 147 tok/s/user on Qwen 3.5 397B-A17B: B200 hits 5101 tok/s/chip, H100 hits 1091. Per-million costs land at $0.09 and $0.30 respectively. B200 is 215% cheaper per token; B200 delivers 367% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:9325.3H100:2115.0 | B200:6763.7H100:1541.4 | B200:5100.8H100:1091.2 |
| Cost ($/M tok) | B200:$0.052H100:$0.154 | B200:$0.071H100:$0.211 | B200:$0.094H100:$0.297 |
| tok/s/MW | B200:5453369H100:1543820 | B200:3955389H100:1125109 | B200:2982902H100:796526 |
| Concurrency | B200:~79H100:~37 | B200:~33H100:~14 | B200:~17H100:~7 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.