Kimi K2.5/K2.6/K2.7-Code 1T — B200 vs MI325X
Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) 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.
B200 posts 520 tok/s/chip for $0.94 per million tokens at 40 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T; MI325X posts 89 tok/s/chip for $3.46. B200 is 270% cheaper per token; B200 delivers 486% more tok/s/chip.
Throughput at 44 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T: B200 hits 465 tok/s/chip, MI325X hits 71. Per-million costs land at $1.05 and $4.28 respectively. B200 is 310% cheaper per token; B200 delivers 551% more tok/s/chip.
B200 / MI325X on Kimi K2.5/K2.6/K2.7-Code 1T at 48 tok/s/user: 415 / 59 tok/s/chip, $1.16 / $5.24 per million tokens. B200 is 351% cheaper per token; B200 delivers 605% more tok/s/chip. (Numbers reflect the default 1k/1k · int4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:519.9MI325X:88.7 | B200:465.2MI325X:71.4 | B200:415.2MI325X:58.9 |
| Cost ($/M tok) | B200:$0.936MI325X:$3.462 | B200:$1.045MI325X:$4.284 | B200:$1.162MI325X:$5.244 |
| tok/s/MW | B200:304059MI325X:52470 | B200:272076MI325X:42268 | B200:242782MI325X:34866 |
| Concurrency | B200:~55MI325X:~9 | B200:~45MI325X:~7 | B200:~36MI325X:~5 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.