MiniMax M3 428B — H200 vs MI355X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) 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.
H200 posts 678 tok/s/GPU for $0.57 per million tokens at 80 tok/s/user on MiniMax M3 428B; MI355X posts 1837 tok/s/GPU for $0.22. MI355X is 156% cheaper per token; MI355X delivers 171% more tok/s/GPU.
Throughput at 152 tok/s/user on MiniMax M3 428B: H200 hits 333 tok/s/GPU, MI355X hits 750. Per-million costs land at $1.16 and $0.55 respectively. MI355X is 110% cheaper per token; MI355X delivers 126% more tok/s/GPU.
H200 / MI355X on MiniMax M3 428B at 223 tok/s/user: 166 / 437 tok/s/GPU, $2.43 / $0.95 per million tokens. MI355X is 156% cheaper per token; MI355X delivers 163% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | H200:677.7MI355X:1836.5 | H200:332.6MI355X:750.1 | H200:166.1MI355X:437.4 |
| Cost ($/M tok) | H200:$0.572MI355X:$0.224 | H200:$1.161MI355X:$0.551 | H200:$2.429MI355X:$0.949 |
| tok/s/MW | H200:494684MI355X:878714 | H200:242783MI355X:358913 | H200:121223MI355X:209305 |
| Concurrency | H200:~19MI355X:~49 | H200:~4MI355X:~11 | H200:~3MI355X:~4 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.