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 47 tok/s/user on MiniMax M3 428B: GB300 NVL72 hits 1302 tok/s/GPU, H100 hits 894. Per-million costs land at $0.56 and $0.40 respectively. H100 is 39% cheaper per token; GB300 NVL72 delivers 46% more tok/s/GPU.
GB300 NVL72 / H100 on MiniMax M3 428B at 79 tok/s/user: 433 / 507 tok/s/GPU, $1.69 / $0.71 per million tokens. H100 is 136% cheaper per token; H100 delivers 17% more tok/s/GPU.
Toward the upper edge of the 15–142 tok/s/user interactivity band, at 110 tok/s/user on MiniMax M3 428B: GB300 NVL72 runs 181 tok/s/GPU at $3.90/M tokens, H100 runs 302 at $1.17/M. H100 is 233% cheaper per token; H100 delivers 67% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:1302.0H100:893.6 | GB300 NVL72:433.2H100:506.7 | GB300 NVL72:180.6H100:302.3 |
| Cost ($/M tok) | GB300 NVL72:$0.556H100:$0.401 | GB300 NVL72:$1.686H100:$0.714 | GB300 NVL72:$3.895H100:$1.170 |
| tok/s/MW | GB300 NVL72:620016H100:516540 | GB300 NVL72:206282H100:292878 | GB300 NVL72:85998H100:174737 |
| Concurrency | GB300 NVL72:~150H100:~86 | GB300 NVL72:~18H100:~28 | GB300 NVL72:~5H100:~12 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.