MiniMax M3 428B — GB300 NVL72 vs H100
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and H100 (NVIDIA Hopper) 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.
Throughput at 75 tok/s/user on MiniMax M3 428B: GB300 NVL72 hits 15293 tok/s/chip, H100 hits 993. Per-million costs land at $0.04 and $0.33 respectively. GB300 NVL72 is 680% cheaper per token; GB300 NVL72 delivers 1440% more tok/s/chip.
GB300 NVL72 / H100 on MiniMax M3 428B at 129 tok/s/user: 4345 / 699 tok/s/chip, $0.16 / $0.46 per million tokens. GB300 NVL72 is 181% cheaper per token; GB300 NVL72 delivers 522% more tok/s/chip.
Toward the upper edge of the 22–235 tok/s/user interactivity band, at 182 tok/s/user on MiniMax M3 428B: GB300 NVL72 runs 2252 tok/s/chip at $0.29/M tokens, H100 runs 544 at $0.60/M. GB300 NVL72 is 104% cheaper per token; GB300 NVL72 delivers 314% more tok/s/chip. (Numbers reflect the default 8k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | GB300 NVL72:15293.2H100:992.7 | GB300 NVL72:4345.5H100:698.6 | GB300 NVL72:2252.3H100:544.3 |
| Cost ($/M tok) | GB300 NVL72:$0.042H100:$0.328 | GB300 NVL72:$0.164H100:$0.462 | GB300 NVL72:$0.295H100:$0.602 |
| tok/s/MW | GB300 NVL72:7213783H100:724633 | GB300 NVL72:2049745H100:509946 | GB300 NVL72:1062421H100:397317 |
| Concurrency | GB300 NVL72:~2048H100:~13 | GB300 NVL72:~55H100:~6 | GB300 NVL72:~11H100:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.