Qwen 3.5 397B-A17B — H100 vs H200
Head-to-head AI inference benchmark comparison of H100 (NVIDIA Hopper) and H200 (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 64 tok/s/user interactivity on Qwen 3.5 397B-A17B, H100 delivers 433 tok/s/chip at $0.76 per million tokens; H200 delivers 510 tok/s/chip at $0.66. H200 is 15% cheaper per token; H200 delivers 18% more tok/s/chip at this point.
H100 posts 311 tok/s/chip for $1.04 per million tokens at 100 tok/s/user on Qwen 3.5 397B-A17B; H200 posts 366 tok/s/chip for $0.92. H200 is 12% cheaper per token; H200 delivers 18% more tok/s/chip.
Throughput at 135 tok/s/user on Qwen 3.5 397B-A17B: H100 hits 236 tok/s/chip, H200 hits 302. Per-million costs land at $1.38 and $1.10 respectively. H200 is 26% cheaper per token; H200 delivers 28% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H100:433.3H200:510.5 | H100:310.5H200:366.2 | H100:236.1H200:302.5 |
| Cost ($/M tok) | H100:$0.763H200:$0.664 | H100:$1.036H200:$0.922 | H100:$1.383H200:$1.102 |
| tok/s/MW | H100:316265H200:372593 | H100:226676H200:267292 | H100:172330H200:220775 |
| Concurrency | H100:~28H200:~32 | H100:~13H200:~15 | H100:~7H200:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.