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 54 tok/s/user: 1874 / 469 tok/s/GPU, $0.34 / $0.83 per million tokens. GB200 NVL72 is 147% cheaper per token; GB200 NVL72 delivers 299% more tok/s/GPU.
Around the middle of the 28–132 tok/s/user interactivity band, at 80 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 736 tok/s/GPU at $0.82/M tokens, H200 runs 323 at $1.22/M. GB200 NVL72 is 48% cheaper per token; GB200 NVL72 delivers 128% more tok/s/GPU.
Setting 107 tok/s/user as the target on Qwen 3.5 397B-A17B, GB200 NVL72 produces 416 tok/s/GPU ($1.47 per million tokens) and H200 produces 271 ($1.45). H200 is 2% cheaper per token; GB200 NVL72 delivers 53% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:1874.3H200:469.3 | GB200 NVL72:735.7H200:322.8 | GB200 NVL72:415.8H200:271.4 |
| Cost ($/M tok) | GB200 NVL72:$0.338H200:$0.833 | GB200 NVL72:$0.823H200:$1.216 | GB200 NVL72:$1.470H200:$1.446 |
| tok/s/MW | GB200 NVL72:892542H200:271276 | GB200 NVL72:350356H200:186573 | GB200 NVL72:198012H200:156874 |
| Concurrency | GB200 NVL72:~376H200:~36 | GB200 NVL72:~40H200:~17 | GB200 NVL72:~17H200:~12 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.