Qwen 3.5 397B-A17B — GB200 NVL72 vs H200
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) 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.
GB200 NVL72 / H200 on Qwen 3.5 397B-A17B at 74 tok/s/user: 10938 / 1540 tok/s/chip, $0.05 / $0.22 per million tokens. GB200 NVL72 is 366% cheaper per token; GB200 NVL72 delivers 610% more tok/s/chip.
Around the middle of the 36–187 tok/s/user interactivity band, at 112 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 6807 tok/s/chip at $0.08/M tokens, H200 runs 1190 at $0.29/M. GB200 NVL72 is 268% cheaper per token; GB200 NVL72 delivers 472% more tok/s/chip.
Setting 150 tok/s/user as the target on Qwen 3.5 397B-A17B, GB200 NVL72 produces 3508 tok/s/chip ($0.14 per million tokens) and H200 produces 940 ($0.36). GB200 NVL72 is 149% cheaper per token; GB200 NVL72 delivers 273% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:10938.5H200:1539.8 | GB200 NVL72:6806.9H200:1190.1 | GB200 NVL72:3508.4H200:939.8 |
| Cost ($/M tok) | GB200 NVL72:$0.047H200:$0.220 | GB200 NVL72:$0.078H200:$0.286 | GB200 NVL72:$0.144H200:$0.359 |
| tok/s/MW | GB200 NVL72:5849447H200:1123938 | GB200 NVL72:3640060H200:868695 | GB200 NVL72:1876167H200:685976 |
| Concurrency | GB200 NVL72:~1264H200:~19 | GB200 NVL72:~332H200:~10 | GB200 NVL72:~57H200:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.