Qwen 3.5 397B-A17B — GB200 NVL72 vs MI325X
Head-to-head AI inference benchmark comparison of GB200 NVL72 (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.
GB200 NVL72 / MI325X on Qwen 3.5 397B-A17B at 51 tok/s/user: 12378 / 679 tok/s/chip, $0.04 / $0.45 per million tokens. GB200 NVL72 is 966% cheaper per token; GB200 NVL72 delivers 1723% more tok/s/chip.
Around the middle of the 46–68 tok/s/user interactivity band, at 57 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 12241 tok/s/chip at $0.04/M tokens, MI325X runs 527 at $0.57/M. GB200 NVL72 is 1257% cheaper per token; GB200 NVL72 delivers 2221% more tok/s/chip.
Setting 62 tok/s/user as the target on Qwen 3.5 397B-A17B, GB200 NVL72 produces 12060 tok/s/chip ($0.04 per million tokens) and MI325X produces 411 ($0.77). GB200 NVL72 is 1691% cheaper per token; GB200 NVL72 delivers 2837% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:12378.1MI325X:679.2 | GB200 NVL72:12241.0MI325X:527.3 | GB200 NVL72:12059.6MI325X:410.6 |
| Cost ($/M tok) | GB200 NVL72:$0.042MI325X:$0.445 | GB200 NVL72:$0.042MI325X:$0.573 | GB200 NVL72:$0.043MI325X:$0.768 |
| tok/s/MW | GB200 NVL72:6619279MI325X:401869 | GB200 NVL72:6546010MI325X:312008 | GB200 NVL72:6449001MI325X:242971 |
| Concurrency | GB200 NVL72:~1899MI325X:~13 | GB200 NVL72:~1568MI325X:~9 | GB200 NVL72:~1420MI325X:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.