Qwen 3.5 397B-A17B — B300 vs H100
Head-to-head AI inference benchmark comparison of B300 (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.
B300 posts 3055 tok/s/chip for $0.20 per million tokens at 65 tok/s/user on Qwen 3.5 397B-A17B; H100 posts 427 tok/s/chip for $0.77. B300 is 280% cheaper per token; B300 delivers 615% more tok/s/chip.
Throughput at 100 tok/s/user on Qwen 3.5 397B-A17B: B300 hits 1723 tok/s/chip, H100 hits 311. Per-million costs land at $0.37 and $1.04 respectively. B300 is 182% cheaper per token; B300 delivers 455% more tok/s/chip.
B300 / H100 on Qwen 3.5 397B-A17B at 136 tok/s/user: 1122 / 234 tok/s/chip, $0.56 / $1.40 per million tokens. B300 is 149% cheaper per token; B300 delivers 380% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B300:3055.1H100:427.2 | B300:1723.4H100:310.5 | B300:1121.7H100:233.9 |
| Cost ($/M tok) | B300:$0.204H100:$0.775 | B300:$0.368H100:$1.036 | B300:$0.561H100:$1.397 |
| tok/s/MW | B300:1607939H100:311848 | B300:907067H100:226676 | B300:590361H100:170740 |
| Concurrency | B300:~96H100:~27 | B300:~36H100:~13 | B300:~18H100:~7 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.