Qwen 3.5 397B-A17B — GB300 NVL72 vs H100
Head-to-head AI inference benchmark comparison of GB300 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.
GB300 NVL72 / H100 on Qwen 3.5 397B-A17B at 66 tok/s/user: 1478 / 422 tok/s/GPU, $0.51 / $0.87 per million tokens. GB300 NVL72 is 72% cheaper per token; GB300 NVL72 delivers 251% more tok/s/GPU.
Around the middle of the 32–171 tok/s/user interactivity band, at 101 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 runs 589 tok/s/GPU at $1.23/M tokens, H100 runs 308 at $1.16/M. H100 is 6% cheaper per token; GB300 NVL72 delivers 91% more tok/s/GPU.
Setting 136 tok/s/user as the target on Qwen 3.5 397B-A17B, GB300 NVL72 produces 334 tok/s/GPU ($2.19 per million tokens) and H100 produces 234 ($1.55). H100 is 41% cheaper per token; GB300 NVL72 delivers 43% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:1478.2H100:421.6 | GB300 NVL72:589.3H100:308.0 | GB300 NVL72:334.0H100:233.9 |
| Cost ($/M tok) | GB300 NVL72:$0.508H100:$0.873 | GB300 NVL72:$1.225H100:$1.159 | GB300 NVL72:$2.193H100:$1.553 |
| tok/s/MW | GB300 NVL72:697275H100:307714 | GB300 NVL72:277973H100:224844 | GB300 NVL72:157564H100:170740 |
| Concurrency | GB300 NVL72:~167H100:~26 | GB300 NVL72:~25H100:~12 | GB300 NVL72:~11H100:~7 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.