DeepSeek V4 Pro 1.6T — B200 vs B300
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and B300 (NVIDIA Blackwell) on DeepSeek V4 Pro 1.6T. 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 61 tok/s/user interactivity on DeepSeek V4 Pro 1.6T, B200 delivers 3770 tok/s/GPU at $0.14 per million tokens; B300 delivers 7530 tok/s/GPU at $0.09. B300 is 68% cheaper per token; B300 delivers 100% more tok/s/GPU at this point.
B200 posts 882 tok/s/GPU for $0.61 per million tokens at 116 tok/s/user on DeepSeek V4 Pro 1.6T; B300 posts 1850 tok/s/GPU for $0.34. B300 is 79% cheaper per token; B300 delivers 110% more tok/s/GPU.
Throughput at 171 tok/s/user on DeepSeek V4 Pro 1.6T: B200 hits 485 tok/s/GPU, B300 hits 722. Per-million costs land at $1.14 and $0.91 respectively. B300 is 26% cheaper per token; B300 delivers 49% more tok/s/GPU. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | B200:3770.5B300:7529.8 | B200:882.2B300:1850.0 | B200:485.0B300:722.0 |
| Cost ($/M tok) | B200:$0.144B300:$0.086 | B200:$0.613B300:$0.343 | B200:$1.144B300:$0.908 |
| tok/s/MW | B200:2204965B300:3963055 | B200:515922B300:973697 | B200:283620B300:380014 |
| Concurrency | B200:~76B300:~634 | B200:~8B300:~112 | B200:~3B300:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.