DeepSeek V4 Pro 1.6T — B200 vs GB300 NVL72
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and GB300 NVL72 (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.
Throughput at 66 tok/s/user on DeepSeek V4 Pro 1.6T: B200 hits 6578 tok/s/chip, GB300 NVL72 hits 9839. Per-million costs land at $0.07 and $0.07 respectively. GB300 NVL72 is 12% cheaper per token; GB300 NVL72 delivers 50% more tok/s/chip.
B200 / GB300 NVL72 on DeepSeek V4 Pro 1.6T at 120 tok/s/user: 1526 / 3892 tok/s/chip, $0.32 / $0.17 per million tokens. GB300 NVL72 is 90% cheaper per token; GB300 NVL72 delivers 155% more tok/s/chip.
Toward the upper edge of the 13–226 tok/s/user interactivity band, at 173 tok/s/user on DeepSeek V4 Pro 1.6T: B200 runs 605 tok/s/chip at $0.79/M tokens, GB300 NVL72 runs 895 at $0.73/M. GB300 NVL72 is 8% cheaper per token; GB300 NVL72 delivers 48% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:6578.4GB300 NVL72:9839.2 | B200:1525.5GB300 NVL72:3891.7 | B200:604.6GB300 NVL72:894.8 |
| Cost ($/M tok) | B200:$0.073GB300 NVL72:$0.065 | B200:$0.316GB300 NVL72:$0.167 | B200:$0.795GB300 NVL72:$0.733 |
| tok/s/MW | B200:3847044GB300 NVL72:4641127 | B200:892127GB300 NVL72:1835702 | B200:353577GB300 NVL72:422058 |
| Concurrency | B200:~935GB300 NVL72:~1026 | B200:~133GB300 NVL72:~320 | B200:~32GB300 NVL72:~29 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.