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 46 tok/s/user interactivity on Qwen 3.5 397B-A17B, GB300 NVL72 delivers 5633 tok/s/GPU at $0.13 per million tokens; MI325X delivers 403 tok/s/GPU at $0.86. GB300 NVL72 is 571% cheaper per token; GB300 NVL72 delivers 1298% more tok/s/GPU at this point.
GB300 NVL72 posts 2534 tok/s/GPU for $0.29 per million tokens at 55 tok/s/user on Qwen 3.5 397B-A17B; MI325X posts 219 tok/s/GPU for $1.63. GB300 NVL72 is 462% cheaper per token; GB300 NVL72 delivers 1055% more tok/s/GPU.
Throughput at 64 tok/s/user on Qwen 3.5 397B-A17B: GB300 NVL72 hits 1595 tok/s/GPU, MI325X hits 134. Per-million costs land at $0.47 and $2.64 respectively. GB300 NVL72 is 460% cheaper per token; GB300 NVL72 delivers 1088% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:5632.9MI325X:402.9 | GB300 NVL72:2534.1MI325X:219.4 | GB300 NVL72:1595.1MI325X:134.3 |
| Cost ($/M tok) | GB300 NVL72:$0.129MI325X:$0.864 | GB300 NVL72:$0.290MI325X:$1.631 | GB300 NVL72:$0.472MI325X:$2.638 |
| tok/s/MW | GB300 NVL72:2657012MI325X:238420 | GB300 NVL72:1195328MI325X:129810 | GB300 NVL72:752417MI325X:79448 |
| Concurrency | GB300 NVL72:~1570MI325X:~37 | GB300 NVL72:~559MI325X:~16 | GB300 NVL72:~214MI325X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.