MiniMax M3 428B — GB200 NVL72 vs H200
Head-to-head AI inference benchmark comparison of GB200 NVL72 (NVIDIA Blackwell) and H200 (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 52 tok/s/user on MiniMax M3 428B: GB200 NVL72 hits 1598 tok/s/GPU, H200 hits 1007. Per-million costs land at $0.38 and $0.39 respectively. GB200 NVL72 is 3% cheaper per token; GB200 NVL72 delivers 59% more tok/s/GPU.
GB200 NVL72 / H200 on MiniMax M3 428B at 87 tok/s/user: 638 / 634 tok/s/GPU, $1.00 / $0.62 per million tokens. H200 is 61% cheaper per token; throughput per GPU is essentially tied.
Toward the upper edge of the 17–157 tok/s/user interactivity band, at 122 tok/s/user on MiniMax M3 428B: GB200 NVL72 runs 211 tok/s/GPU at $2.86/M tokens, H200 runs 410 at $0.95/M. H200 is 201% cheaper per token; H200 delivers 94% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB200 NVL72:1598.2H200:1007.1 | GB200 NVL72:638.3H200:633.8 | GB200 NVL72:210.9H200:409.6 |
| Cost ($/M tok) | GB200 NVL72:$0.382H200:$0.394 | GB200 NVL72:$0.996H200:$0.618 | GB200 NVL72:$2.862H200:$0.951 |
| tok/s/MW | GB200 NVL72:761024H200:582133 | GB200 NVL72:303973H200:366335 | GB200 NVL72:100447H200:236761 |
| Concurrency | GB200 NVL72:~180H200:~42 | GB200 NVL72:~74H200:~16 | GB200 NVL72:~18H200:~7 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.