Qwen 3.5 397B-A17B — GB300 NVL72 vs H200
Head-to-head AI inference benchmark comparison of GB300 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.
Near the low end of the 40–187 tok/s/user interactivity band, at 77 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 runs 11644 tok/s/chip at $0.06/M tokens, H200 runs 1510 at $0.22/M. GB300 NVL72 is 307% cheaper per token; GB300 NVL72 delivers 671% more tok/s/chip.
Setting 114 tok/s/user as the target on Qwen 3.5 397B-A17B, GB300 NVL72 produces 7392 tok/s/chip ($0.09 per million tokens) and H200 produces 1174 ($0.29). GB300 NVL72 is 230% cheaper per token; GB300 NVL72 delivers 530% more tok/s/chip.
At 151 tok/s/user interactivity on Qwen 3.5 397B-A17B, GB300 NVL72 delivers 3829 tok/s/chip at $0.16 per million tokens; H200 delivers 935 tok/s/chip at $0.36. GB300 NVL72 is 120% cheaper per token; GB300 NVL72 delivers 310% more tok/s/chip at this point. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:11644.2H200:1510.0 | GB300 NVL72:7392.3H200:1174.1 | GB300 NVL72:3829.5H200:934.8 |
| Cost ($/M tok) | GB300 NVL72:$0.055H200:$0.224 | GB300 NVL72:$0.088H200:$0.290 | GB300 NVL72:$0.164H200:$0.361 |
| tok/s/MW | GB300 NVL72:5492531H200:1102201 | GB300 NVL72:3486939H200:856983 | GB300 NVL72:1806354H200:682333 |
| Concurrency | GB300 NVL72:~1299H200:~18 | GB300 NVL72:~364H200:~10 | GB300 NVL72:~61H200:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.