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 63 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 hits 1282 tok/s/GPU, H100 hits 517. Per-million costs land at $0.49 and $0.73 respectively. GB200 NVL72 is 48% cheaper per token; GB200 NVL72 delivers 148% more tok/s/GPU.
GB200 NVL72 / H100 on Qwen 3.5 397B-A17B at 97 tok/s/user: 498 / 220 tok/s/GPU, $1.20 / $1.66 per million tokens. GB200 NVL72 is 39% cheaper per token; GB200 NVL72 delivers 127% more tok/s/GPU.
Toward the upper edge of the 29–165 tok/s/user interactivity band, at 131 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 240 tok/s/GPU at $2.56/M tokens, H100 runs 86 at $4.06/M. GB200 NVL72 is 59% cheaper per token; GB200 NVL72 delivers 178% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:1282.2H100:517.2 | GB200 NVL72:498.4H100:220.0 | GB200 NVL72:239.9H100:86.2 |
| Cost ($/M tok) | GB200 NVL72:$0.489H100:$0.726 | GB200 NVL72:$1.201H100:$1.664 | GB200 NVL72:$2.557H100:$4.057 |
| tok/s/MW | GB200 NVL72:610561H100:298955 | GB200 NVL72:237351H100:127176 | GB200 NVL72:114230H100:49829 |
| Concurrency | GB200 NVL72:~129H100:~39 | GB200 NVL72:~22H100:~10 | GB200 NVL72:~8H100:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.