Qwen 3.5 397B-A17B — B200 vs MI300X
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and MI300X (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.
Near the low end of the 23–66 tok/s/user interactivity band, at 34 tok/s/user on Qwen 3.5 397B-A17B: B200 runs 11201 tok/s/chip at $0.04/M tokens, MI300X runs 1042 at $0.25/M. B200 is 488% cheaper per token; B200 delivers 975% more tok/s/chip.
Setting 45 tok/s/user as the target on Qwen 3.5 397B-A17B, B200 produces 10167 tok/s/chip ($0.05 per million tokens) and MI300X produces 710 ($0.37). B200 is 687% cheaper per token; B200 delivers 1332% more tok/s/chip.
At 56 tok/s/user interactivity on Qwen 3.5 397B-A17B, B200 delivers 9325 tok/s/chip at $0.05 per million tokens; MI300X delivers 451 tok/s/chip at $0.59. B200 is 1036% cheaper per token; B200 delivers 1968% more tok/s/chip at this point. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:11200.9MI300X:1042.3 | B200:10166.6MI300X:710.2 | B200:9325.3MI300X:450.9 |
| Cost ($/M tok) | B200:$0.043MI300X:$0.253 | B200:$0.047MI300X:$0.372 | B200:$0.052MI300X:$0.586 |
| tok/s/MW | B200:6550241MI300X:749891 | B200:5945409MI300X:510939 | B200:5453369MI300X:324354 |
| Concurrency | B200:~155MI300X:~29 | B200:~105MI300X:~15 | B200:~79MI300X:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.