DeepSeek R1 — GB300 NVL72 vs H100
Head-to-head AI inference benchmark comparison of GB300 NVL72 (NVIDIA Blackwell) and H100 (NVIDIA Hopper) on DeepSeek R1. 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.
At 40 tok/s/user interactivity on DeepSeek R1, GB300 NVL72 delivers 7720 tok/s/GPU at $0.09 per million tokens; H100 delivers 272 tok/s/GPU at $1.34. GB300 NVL72 is 1314% cheaper per token; GB300 NVL72 delivers 2738% more tok/s/GPU at this point.
GB300 NVL72 posts 5403 tok/s/GPU for $0.14 per million tokens at 70 tok/s/user on DeepSeek R1; H100 posts 77 tok/s/GPU for $4.65. GB300 NVL72 is 3312% cheaper per token; GB300 NVL72 delivers 6885% more tok/s/GPU.
Throughput at 99 tok/s/user on DeepSeek R1: GB300 NVL72 hits 2481 tok/s/GPU, H100 hits 23. Per-million costs land at $0.30 and $15.6 respectively. GB300 NVL72 is 5167% cheaper per token; GB300 NVL72 delivers 10529% more tok/s/GPU. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | GB300 NVL72:7719.7H100:272.0 | GB300 NVL72:5403.3H100:77.4 | GB300 NVL72:2481.4H100:23.3 |
| Cost ($/M tok) | GB300 NVL72:$0.095H100:$1.339 | GB300 NVL72:$0.136H100:$4.655 | GB300 NVL72:$0.297H100:$15.632 |
| tok/s/MW | GB300 NVL72:3676036H100:157238 | GB300 NVL72:2573015H100:44718 | GB300 NVL72:1181606H100:13494 |
| Concurrency | GB300 NVL72:~2869H100:~605 | GB300 NVL72:~959H100:~57 | GB300 NVL72:~571H100:~8 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.