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 46 tok/s/user: 3800 / 403 tok/s/GPU, $0.16 / $0.86 per million tokens. GB200 NVL72 is 431% cheaper per token; GB200 NVL72 delivers 843% more tok/s/GPU.
Around the middle of the 37–72 tok/s/user interactivity band, at 55 tok/s/user on Qwen 3.5 397B-A17B: GB200 NVL72 runs 1781 tok/s/GPU at $0.36/M tokens, MI325X runs 219 at $1.63/M. GB200 NVL72 is 358% cheaper per token; GB200 NVL72 delivers 712% more tok/s/GPU.
Setting 64 tok/s/user as the target on Qwen 3.5 397B-A17B, GB200 NVL72 produces 1239 tok/s/GPU ($0.51 per million tokens) and MI325X produces 134 ($2.64). GB200 NVL72 is 422% cheaper per token; GB200 NVL72 delivers 823% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:3800.3MI325X:402.9 | GB200 NVL72:1781.3MI325X:219.4 | GB200 NVL72:1239.3MI325X:134.3 |
| Cost ($/M tok) | GB200 NVL72:$0.163MI325X:$0.864 | GB200 NVL72:$0.356MI325X:$1.631 | GB200 NVL72:$0.506MI325X:$2.638 |
| tok/s/MW | GB200 NVL72:1809650MI325X:184830 | GB200 NVL72:848237MI325X:100633 | GB200 NVL72:590129MI325X:61591 |
| Concurrency | GB200 NVL72:~962MI325X:~37 | GB200 NVL72:~338MI325X:~16 | GB200 NVL72:~113MI325X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.