Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and GB200 NVL72 (NVIDIA Blackwell) on Kimi K3 2.8T. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics β same interactions as the main inference chart.
AgentX replays real coding-agent sessions rather than fixed-length prompts, so context grows turn over turn and most of each request is served from cache instead of being recomputed. That turns the comparison into a systems question: KV transfer between nodes, prefix-aware routing, and cache capacity all move the curve alongside raw chip throughput. Fixed-sequence workloads stay the clean baseline for kernel and silicon performance, so the two scenarios answer different questions about the same hardware. Learn more about AgentX β
B200 posts 5835 tok/s/chip for $0.08 per million tokens at 55 tok/s/user on Kimi K3 2.8T; GB200 NVL72 posts 4157 tok/s/chip for $0.12. B200 is 51% cheaper per token; B200 delivers 40% more tok/s/chip.
Throughput at 101 tok/s/user on Kimi K3 2.8T: B200 hits 2926 tok/s/chip, GB200 NVL72 hits 2364. Per-million costs land at $0.16 and $0.22 respectively. B200 is 33% cheaper per token; B200 delivers 24% more tok/s/chip.
B200 / GB200 NVL72 on Kimi K3 2.8T at 147 tok/s/user: 1724 / 1138 tok/s/chip, $0.28 / $0.45 per million tokens. B200 is 63% cheaper per token; B200 delivers 51% more tok/s/chip. (Numbers reflect the default agentic-traces Β· fp4 selection for this URL β table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:5835.0GB200 NVL72:4157.4 | B200:2926.2GB200 NVL72:2363.9 | B200:1723.8GB200 NVL72:1138.2 |
| Cost ($/M tok) | B200:$0.082GB200 NVL72:$0.124 | B200:$0.164GB200 NVL72:$0.219 | B200:$0.279GB200 NVL72:$0.454 |
| tok/s/MW | B200:3412305GB200 NVL72:2223219 | B200:1711226GB200 NVL72:1264133 | B200:1008090GB200 NVL72:608688 |
| Concurrency | B200:~19GB200 NVL72:~16 | B200:~8GB200 NVL72:~16 | B200:~4GB200 NVL72:~3 |
Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.
No measurements to plot for this selection. Review the benchmark controls above or adjust quick filters.
