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 81 tok/s/user: 10585 / 1471 tok/s/chip, $0.05 / $0.23 per million tokens. GB200 NVL72 is 372% cheaper per token; GB200 NVL72 delivers 620% more tok/s/chip.
Around the middle of the 46–187 tok/s/user interactivity band, at 117 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 9123 tok/s/chip at $0.06/M tokens, H200 runs 1151 at $0.30/M. GB200 NVL72 is 420% cheaper per token; GB200 NVL72 delivers 693% more tok/s/chip.
Setting 152 tok/s/user as the target on Qwen 3.5 397B-A17B, GB200 NVL72 produces 4254 tok/s/chip ($0.12 per million tokens) and H200 produces 930 ($0.36). GB200 NVL72 is 197% cheaper per token; GB200 NVL72 delivers 357% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:10584.7H200:1470.8 | GB200 NVL72:9122.6H200:1150.5 | GB200 NVL72:4254.1H200:929.9 |
| Cost ($/M tok) | GB200 NVL72:$0.049H200:$0.230 | GB200 NVL72:$0.057H200:$0.296 | GB200 NVL72:$0.122H200:$0.363 |
| tok/s/MW | GB200 NVL72:5660283H200:1073560 | GB200 NVL72:4878408H200:839798 | GB200 NVL72:2274930H200:678738 |
| Concurrency | GB200 NVL72:~1187H200:~17 | GB200 NVL72:~680H200:~9 | GB200 NVL72:~64H200:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.