MiniMax M2.5/M2.7 — GB200 NVL72 vs H100
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and H100 (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, GB200 NVL72 delivers 3263 tok/s/chip at $0.16 per million tokens; H100 delivers 599 tok/s/chip at $0.53. GB200 NVL72 is 239% cheaper per token; GB200 NVL72 delivers 444% more tok/s/chip at this point.
GB200 NVL72 posts 1297 tok/s/chip for $0.41 per million tokens at 79 tok/s/user on MiniMax M2.5/M2.7; H100 posts 353 tok/s/chip for $0.92. GB200 NVL72 is 123% cheaper per token; GB200 NVL72 delivers 267% more tok/s/chip.
Throughput at 98 tok/s/user on MiniMax M2.5/M2.7: GB200 NVL72 hits 761 tok/s/chip, H100 hits 207. Per-million costs land at $0.69 and $1.56 respectively. GB200 NVL72 is 125% cheaper per token; GB200 NVL72 delivers 267% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB200 NVL72:3263.0H100:599.4 | GB200 NVL72:1296.5H100:353.0 | GB200 NVL72:760.8H100:207.1 |
| Cost ($/M tok) | GB200 NVL72:$0.157H100:$0.534 | GB200 NVL72:$0.412H100:$0.920 | GB200 NVL72:$0.694H100:$1.563 |
| tok/s/MW | GB200 NVL72:1744903H100:437502 | GB200 NVL72:693331H100:257637 | GB200 NVL72:406865H100:151188 |
| Concurrency | GB200 NVL72:~318H100:~41 | GB200 NVL72:~64H100:~18 | GB200 NVL72:~32H100:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.