MiniMax M2.5/M2.7 — H100 vs H200
Head-to-head AI inference benchmark comparison of H100 (NVIDIA Hopper) and H200 (NVIDIA Hopper) 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.
At 60 tok/s/user interactivity on MiniMax M2.5/M2.7, H100 delivers 599 tok/s/chip at $0.53 per million tokens; H200 delivers 880 tok/s/chip at $0.38. H200 is 39% cheaper per token; H200 delivers 47% more tok/s/chip at this point.
H100 posts 353 tok/s/chip for $0.92 per million tokens at 79 tok/s/user on MiniMax M2.5/M2.7; H200 posts 551 tok/s/chip for $0.57. H200 is 61% cheaper per token; H200 delivers 56% more tok/s/chip.
Throughput at 98 tok/s/user on MiniMax M2.5/M2.7: H100 hits 207 tok/s/chip, H200 hits 374. Per-million costs land at $1.56 and $0.91 respectively. H200 is 72% cheaper per token; H200 delivers 81% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | H100:599.4H200:880.4 | H100:353.0H200:551.3 | H100:207.1H200:374.3 |
| Cost ($/M tok) | H100:$0.534H200:$0.384 | H100:$0.920H200:$0.570 | H100:$1.563H200:$0.906 |
| tok/s/MW | H100:437502H200:642651 | H100:257637H200:402419 | H100:151188H200:273235 |
| Concurrency | H100:~41H200:~30 | H100:~18H200:~15 | H100:~9H200:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.