Qwen 3.5 397B-A17B — GB300 NVL72 vs MI325X
Head-to-head AI inference benchmark comparison of GB300 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.
At 53 tok/s/user interactivity on Qwen 3.5 397B-A17B, GB300 NVL72 delivers 13149 tok/s/chip at $0.05 per million tokens; MI325X delivers 626 tok/s/chip at $0.48. GB300 NVL72 is 882% cheaper per token; GB300 NVL72 delivers 2000% more tok/s/chip at this point.
GB300 NVL72 posts 12795 tok/s/chip for $0.05 per million tokens at 58 tok/s/user on Qwen 3.5 397B-A17B; MI325X posts 503 tok/s/chip for $0.60. GB300 NVL72 is 1105% cheaper per token; GB300 NVL72 delivers 2442% more tok/s/chip.
Throughput at 63 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 hits 12458 tok/s/chip, MI325X hits 389. Per-million costs land at $0.05 and $0.82 respectively. GB300 NVL72 is 1485% cheaper per token; GB300 NVL72 delivers 3105% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:13149.0MI325X:626.0 | GB300 NVL72:12795.1MI325X:503.3 | GB300 NVL72:12457.6MI325X:388.7 |
| Cost ($/M tok) | GB300 NVL72:$0.049MI325X:$0.479 | GB300 NVL72:$0.050MI325X:$0.604 | GB300 NVL72:$0.051MI325X:$0.816 |
| tok/s/MW | GB300 NVL72:6202353MI325X:370435 | GB300 NVL72:6035411MI325X:297806 | GB300 NVL72:5876212MI325X:229993 |
| Concurrency | GB300 NVL72:~2304MI325X:~11 | GB300 NVL72:~1886MI325X:~8 | GB300 NVL72:~1516MI325X:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.