MiniMax M3 428B — GB300 NVL72 vs MI325X
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) on MiniMax M3 428B. 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.
GB300 NVL72 posts 16377 tok/s/chip for $0.04 per million tokens at 56 tok/s/user on MiniMax M3 428B; MI325X posts 813 tok/s/chip for $0.38. GB300 NVL72 is 858% cheaper per token; GB300 NVL72 delivers 1915% more tok/s/chip.
Throughput at 90 tok/s/user on MiniMax M3 428B: GB300 NVL72 hits 13339 tok/s/chip, MI325X hits 560. Per-million costs land at $0.05 and $0.55 respectively. GB300 NVL72 is 1038% cheaper per token; GB300 NVL72 delivers 2282% more tok/s/chip.
GB300 NVL72 / MI325X on MiniMax M3 428B at 125 tok/s/user: 4610 / 374 tok/s/chip, $0.15 / $0.82 per million tokens. GB300 NVL72 is 460% cheaper per token; GB300 NVL72 delivers 1134% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:16377.2MI325X:812.7 | GB300 NVL72:13339.1MI325X:560.0 | GB300 NVL72:4610.3MI325X:373.7 |
| Cost ($/M tok) | GB300 NVL72:$0.039MI325X:$0.376 | GB300 NVL72:$0.048MI325X:$0.547 | GB300 NVL72:$0.147MI325X:$0.824 |
| tok/s/MW | GB300 NVL72:7725089MI325X:480863 | GB300 NVL72:6292042MI325X:331354 | GB300 NVL72:2174679MI325X:221150 |
| Concurrency | GB300 NVL72:~2048MI325X:~8 | GB300 NVL72:~1019MI325X:~4 | GB300 NVL72:~60MI325X:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.