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 51 tok/s/user on MiniMax M3 428B: GB300 NVL72 hits 2166 tok/s/GPU, H100 hits 848. Per-million costs land at $0.32 and $0.42 respectively. GB300 NVL72 is 30% cheaper per token; GB300 NVL72 delivers 155% more tok/s/GPU.
GB300 NVL72 / H100 on MiniMax M3 428B at 87 tok/s/user: 692 / 436 tok/s/GPU, $1.10 / $0.84 per million tokens. H100 is 32% cheaper per token; GB300 NVL72 delivers 59% more tok/s/GPU.
Toward the upper edge of the 15–158 tok/s/user interactivity band, at 122 tok/s/user on MiniMax M3 428B: GB300 NVL72 runs 257 tok/s/GPU at $2.82/M tokens, H100 runs 271 at $1.32/M. H100 is 113% cheaper per token; H100 delivers 5% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:2165.7H100:847.7 | GB300 NVL72:692.4H100:436.2 | GB300 NVL72:257.4H100:271.1 |
| Cost ($/M tok) | GB300 NVL72:$0.324H100:$0.423 | GB300 NVL72:$1.103H100:$0.837 | GB300 NVL72:$2.816H100:$1.322 |
| tok/s/MW | GB300 NVL72:1021543H100:618784 | GB300 NVL72:326620H100:318404 | GB300 NVL72:121434H100:197870 |
| Concurrency | GB300 NVL72:~484H100:~74 | GB300 NVL72:~78H100:~22 | GB300 NVL72:~22H100:~10 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.