MiniMax M3 428B — GB200 NVL72 vs MI355X
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) 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.
Setting 62 tok/s/user as the target on MiniMax M3 428B, GB200 NVL72 produces 4174 tok/s/GPU ($0.14 per million tokens) and MI355X produces 5145 ($0.08). MI355X is 81% cheaper per token; MI355X delivers 23% more tok/s/GPU.
At 94 tok/s/user interactivity on MiniMax M3 428B, GB200 NVL72 delivers 1871 tok/s/GPU at $0.33 per million tokens; MI355X delivers 3884 tok/s/GPU at $0.10. MI355X is 216% cheaper per token; MI355X delivers 108% more tok/s/GPU at this point.
GB200 NVL72 posts 793 tok/s/GPU for $0.78 per million tokens at 125 tok/s/user on MiniMax M3 428B; MI355X posts 3053 tok/s/GPU for $0.14. MI355X is 476% cheaper per token; MI355X delivers 285% more tok/s/GPU. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:4173.7MI355X:5145.0 | GB200 NVL72:1871.0MI355X:3884.0 | GB200 NVL72:792.8MI355X:3053.3 |
| Cost ($/M tok) | GB200 NVL72:$0.145MI355X:$0.080 | GB200 NVL72:$0.331MI355X:$0.105 | GB200 NVL72:$0.780MI355X:$0.135 |
| tok/s/MW | GB200 NVL72:2231943MI355X:2461699 | GB200 NVL72:1000508MI355X:1858371 | GB200 NVL72:423951MI355X:1460890 |
| Concurrency | GB200 NVL72:~302MI355X:~44 | GB200 NVL72:~49MI355X:~22 | GB200 NVL72:~15MI355X:~13 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.