Kimi K2.5/K2.6/K2.7-Code 1T — B200 vs GB200 NVL72
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and GB200 NVL72 (NVIDIA Blackwell) on Kimi K2.5/K2.6/K2.7-Code 1T. 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.
Setting 64 tok/s/user as the target on Kimi K2.5/K2.6/K2.7-Code 1T, B200 produces 3754 tok/s/chip ($0.13 per million tokens) and GB200 NVL72 produces 8980 ($0.06). GB200 NVL72 is 123% cheaper per token; GB200 NVL72 delivers 139% more tok/s/chip.
At 103 tok/s/user interactivity on Kimi K2.5/K2.6/K2.7-Code 1T, B200 delivers 1594 tok/s/chip at $0.30 per million tokens; GB200 NVL72 delivers 1545 tok/s/chip at $0.33. B200 is 11% cheaper per token; B200 delivers 3% more tok/s/chip at this point.
B200 posts 623 tok/s/chip for $0.77 per million tokens at 141 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T; GB200 NVL72 posts 733 tok/s/chip for $0.69. GB200 NVL72 is 12% cheaper per token; GB200 NVL72 delivers 18% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:3753.8GB200 NVL72:8979.7 | B200:1594.1GB200 NVL72:1545.4 | B200:622.7GB200 NVL72:732.9 |
| Cost ($/M tok) | B200:$0.128GB200 NVL72:$0.057 | B200:$0.302GB200 NVL72:$0.334 | B200:$0.773GB200 NVL72:$0.688 |
| tok/s/MW | B200:2195224GB200 NVL72:4801979 | B200:932248GB200 NVL72:826425 | B200:364180GB200 NVL72:391932 |
| Concurrency | B200:~204GB200 NVL72:~1203 | B200:~45GB200 NVL72:~167 | B200:~2GB200 NVL72:~27 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.