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 84 tok/s/user: 11097 / 1809 tok/s/chip, $0.06 / $0.18 per million tokens. GB300 NVL72 is 213% cheaper per token; GB300 NVL72 delivers 513% more tok/s/chip.
Around the middle of the 48–192 tok/s/user interactivity band, at 120 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 runs 6591 tok/s/chip at $0.10/M tokens, H100 runs 1340 at $0.24/M. GB300 NVL72 is 144% cheaper per token; GB300 NVL72 delivers 392% more tok/s/chip.
Setting 156 tok/s/user as the target on Qwen 3.5 397B-A17B, GB300 NVL72 produces 3646 tok/s/chip ($0.17 per million tokens) and H100 produces 1025 ($0.32). GB300 NVL72 is 83% cheaper per token; GB300 NVL72 delivers 256% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:11096.6H100:1808.9 | GB300 NVL72:6591.4H100:1339.8 | GB300 NVL72:3646.3H100:1024.8 |
| Cost ($/M tok) | GB300 NVL72:$0.058H100:$0.181 | GB300 NVL72:$0.100H100:$0.243 | GB300 NVL72:$0.174H100:$0.317 |
| tok/s/MW | GB300 NVL72:5234249H100:1320349 | GB300 NVL72:3109171H100:977955 | GB300 NVL72:1719957H100:748023 |
| Concurrency | GB300 NVL72:~1184H100:~19 | GB300 NVL72:~257H100:~10 | GB300 NVL72:~56H100:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.