Qwen 3.5 397B-A17B — GB200 NVL72 vs H100
Head-to-head AI inference benchmark comparison of GB200 NVL72 (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.
Throughput at 82 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 hits 10540 tok/s/chip, H100 hits 1846. Per-million costs land at $0.05 and $0.18 respectively. GB200 NVL72 is 262% cheaper per token; GB200 NVL72 delivers 471% more tok/s/chip.
GB200 NVL72 / H100 on Qwen 3.5 397B-A17B at 119 tok/s/user: 9005 / 1350 tok/s/chip, $0.06 / $0.24 per million tokens. GB200 NVL72 is 319% cheaper per token; GB200 NVL72 delivers 567% more tok/s/chip.
Toward the upper edge of the 46–192 tok/s/user interactivity band, at 156 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 4153 tok/s/chip at $0.12/M tokens, H100 runs 1025 at $0.32/M. GB200 NVL72 is 154% cheaper per token; GB200 NVL72 delivers 305% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:10540.5H100:1846.4 | GB200 NVL72:9005.0H100:1350.2 | GB200 NVL72:4153.4H100:1024.8 |
| Cost ($/M tok) | GB200 NVL72:$0.049H100:$0.177 | GB200 NVL72:$0.058H100:$0.241 | GB200 NVL72:$0.125H100:$0.317 |
| tok/s/MW | GB200 NVL72:5636619H100:1347765 | GB200 NVL72:4815490H100:985574 | GB200 NVL72:2221079H100:748023 |
| Concurrency | GB200 NVL72:~1174H100:~20 | GB200 NVL72:~653H100:~10 | GB200 NVL72:~61H100:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.