MiniMax M2.5/M2.7 — H200 vs MI325X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI325X (AMD CDNA 3) on MiniMax M2.5/M2.7. 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 46 tok/s/user as the target on MiniMax M2.5/M2.7, H200 produces 1405 tok/s/chip ($0.24 per million tokens) and MI325X produces 1643 ($0.19). MI325X is 31% cheaper per token; MI325X delivers 17% more tok/s/chip.
At 65 tok/s/user interactivity on MiniMax M2.5/M2.7, H200 delivers 764 tok/s/chip at $0.43 per million tokens; MI325X delivers 757 tok/s/chip at $0.41. MI325X is 5% cheaper per token; throughput per chip is essentially tied at this point.
H200 posts 492 tok/s/chip for $0.65 per million tokens at 85 tok/s/user on MiniMax M2.5/M2.7; MI325X posts 325 tok/s/chip for $0.94. H200 is 45% cheaper per token; H200 delivers 51% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H200:1405.1MI325X:1642.6 | H200:763.9MI325X:757.2 | H200:491.8MI325X:324.7 |
| Cost ($/M tok) | H200:$0.243MI325X:$0.186 | H200:$0.434MI325X:$0.412 | H200:$0.647MI325X:$0.940 |
| tok/s/MW | H200:1025592MI325X:971943 | H200:557576MI325X:448067 | H200:358982MI325X:192115 |
| Concurrency | H200:~62MI325X:~147 | H200:~25MI325X:~38 | H200:~12MI325X:~16 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.