MiniMax M3 428B — H200 vs MI300X
Head-to-head AI inference benchmark comparison of H200 (NVIDIA Hopper) and MI300X (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.
Setting 48 tok/s/user as the target on MiniMax M3 428B, H200 produces 1114 tok/s/chip ($0.31 per million tokens) and MI300X produces 509 ($0.51). H200 is 65% cheaper per token; H200 delivers 119% more tok/s/chip.
At 82 tok/s/user interactivity on MiniMax M3 428B, H200 delivers 666 tok/s/chip at $0.51 per million tokens; MI300X delivers 201 tok/s/chip at $1.31. H200 is 159% cheaper per token; H200 delivers 232% more tok/s/chip at this point.
H200 posts 444 tok/s/chip for $0.77 per million tokens at 115 tok/s/user on MiniMax M3 428B; MI300X posts 142 tok/s/chip for $1.82. H200 is 137% cheaper per token; H200 delivers 212% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H200:1114.1MI300X:508.7 | H200:666.5MI300X:200.9 | H200:444.2MI300X:142.4 |
| Cost ($/M tok) | H200:$0.307MI300X:$0.509 | H200:$0.505MI300X:$1.307 | H200:$0.769MI300X:$1.819 |
| tok/s/MW | H200:813204MI300X:366003 | H200:486476MI300X:144553 | H200:324222MI300X:102431 |
| Concurrency | H200:~50MI300X:~45 | H200:~18MI300X:~10 | H200:~9MI300X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.