MiniMax M2.5/M2.7 — GB300 NVL72 vs H200
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) 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.
Throughput at 49 tok/s/user on MiniMax M2.5/M2.7: GB300 NVL72 hits 4954 tok/s/chip, H200 hits 1218. Per-million costs land at $0.13 and $0.29 respectively. GB300 NVL72 is 123% cheaper per token; GB300 NVL72 delivers 307% more tok/s/chip.
GB300 NVL72 / H200 on MiniMax M2.5/M2.7 at 74 tok/s/user: 2117 / 612 tok/s/chip, $0.31 / $0.52 per million tokens. GB300 NVL72 is 68% cheaper per token; GB300 NVL72 delivers 246% more tok/s/chip.
Toward the upper edge of the 25–122 tok/s/user interactivity band, at 98 tok/s/user on MiniMax M2.5/M2.7: GB300 NVL72 runs 901 tok/s/chip at $0.74/M tokens, H200 runs 374 at $0.91/M. GB300 NVL72 is 23% cheaper per token; GB300 NVL72 delivers 141% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:4954.2H200:1218.4 | GB300 NVL72:2116.8H200:612.1 | GB300 NVL72:901.3H200:374.3 |
| Cost ($/M tok) | GB300 NVL72:$0.129H200:$0.289 | GB300 NVL72:$0.308H200:$0.518 | GB300 NVL72:$0.736H200:$0.906 |
| tok/s/MW | GB300 NVL72:2336885H200:889327 | GB300 NVL72:998497H200:446808 | GB300 NVL72:425139H200:273235 |
| Concurrency | GB300 NVL72:~512H200:~51 | GB300 NVL72:~128H200:~18 | GB300 NVL72:~32H200:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.