MiniMax M3 428B — H100 vs MI300X
Head-to-head AI inference benchmark comparison of H100 (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.
At 48 tok/s/user interactivity on MiniMax M3 428B, H100 delivers 882 tok/s/chip at $0.37 per million tokens; MI300X delivers 509 tok/s/chip at $0.51. H100 is 39% cheaper per token; H100 delivers 73% more tok/s/chip at this point.
H100 posts 479 tok/s/chip for $0.68 per million tokens at 82 tok/s/user on MiniMax M3 428B; MI300X posts 201 tok/s/chip for $1.31. H100 is 92% cheaper per token; H100 delivers 138% more tok/s/chip.
Throughput at 115 tok/s/user on MiniMax M3 428B: H100 hits 287 tok/s/chip, MI300X hits 142. Per-million costs land at $1.11 and $1.82 respectively. H100 is 64% cheaper per token; H100 delivers 102% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H100:882.2MI300X:508.7 | H100:478.7MI300X:200.9 | H100:287.3MI300X:142.4 |
| Cost ($/M tok) | H100:$0.365MI300X:$0.509 | H100:$0.682MI300X:$1.307 | H100:$1.111MI300X:$1.819 |
| tok/s/MW | H100:643974MI300X:366003 | H100:349405MI300X:144553 | H100:209742MI300X:102431 |
| Concurrency | H100:~83MI300X:~45 | H100:~26MI300X:~10 | H100:~11MI300X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.