Qwen 3.5 397B-A17B — B200 vs MI325X
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) 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.
B200 / MI325X on Qwen 3.5 397B-A17B at 36 tok/s/user: 10970 / 1221 tok/s/chip, $0.04 / $0.25 per million tokens. B200 is 471% cheaper per token; B200 delivers 799% more tok/s/chip.
Around the middle of the 26–68 tok/s/user interactivity band, at 47 tok/s/user on Qwen 3.5 397B-A17B: B200 runs 10007 tok/s/chip at $0.05/M tokens, MI325X runs 799 at $0.38/M. B200 is 696% cheaper per token; B200 delivers 1153% more tok/s/chip.
Setting 57 tok/s/user as the target on Qwen 3.5 397B-A17B, B200 produces 9253 tok/s/chip ($0.05 per million tokens) and MI325X produces 527 ($0.57). B200 is 1002% cheaper per token; B200 delivers 1655% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:10970.1MI325X:1220.7 | B200:10007.4MI325X:798.9 | B200:9253.0MI325X:527.3 |
| Cost ($/M tok) | B200:$0.044MI325X:$0.250 | B200:$0.048MI325X:$0.382 | B200:$0.052MI325X:$0.573 |
| tok/s/MW | B200:6415251MI325X:722280 | B200:5852282MI325X:472744 | B200:5411121MI325X:312008 |
| Concurrency | B200:~145MI325X:~32 | B200:~99MI325X:~16 | B200:~77MI325X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.