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 60 tok/s/user interactivity on DeepSeek V4 Pro 1.6T, B200 delivers 7374 tok/s/chip at $0.07 per million tokens; B300 delivers 3480 tok/s/chip at $0.19. B200 is 188% cheaper per token; B200 delivers 112% more tok/s/chip at this point.
B200 posts 1668 tok/s/chip for $0.29 per million tokens at 116 tok/s/user on DeepSeek V4 Pro 1.6T; B300 posts 1432 tok/s/chip for $0.44. B200 is 52% cheaper per token; B200 delivers 16% more tok/s/chip.
Throughput at 171 tok/s/user on DeepSeek V4 Pro 1.6T: B200 hits 608 tok/s/chip, B300 hits 707. Per-million costs land at $0.79 and $0.89 respectively. B200 is 12% cheaper per token; B300 delivers 16% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:7374.4B300:3479.7 | B200:1667.8B300:1431.9 | B200:607.9B300:707.3 |
| Cost ($/M tok) | B200:$0.065B300:$0.188 | B200:$0.289B300:$0.438 | B200:$0.791B300:$0.886 |
| tok/s/MW | B200:4312520B300:1831422 | B200:975311B300:753617 | B200:355473B300:372278 |
| Concurrency | B200:~2668B300:~94 | B200:~146B300:~7 | B200:~31B300:~2 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.