Kimi K2.5/K2.6 1T — B300 vs MI325X
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and MI325X (AMD CDNA 3) on Kimi K2.5/K2.6 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.
B300 / MI325X on Kimi K2.5/K2.6 1T at 33 tok/s/user: 795 / 127 tok/s/GPU, $0.84 / $2.79 per million tokens. B300 is 234% cheaper per token; B300 delivers 524% more tok/s/GPU.
Around the middle of the 28–51 tok/s/user interactivity band, at 39 tok/s/user on Kimi K2.5/K2.6 1T: B300 runs 644 tok/s/GPU at $1.02/M tokens, MI325X runs 94 at $3.80/M. B300 is 272% cheaper per token; B300 delivers 583% more tok/s/GPU.
Setting 45 tok/s/user as the target on Kimi K2.5/K2.6 1T, B300 produces 520 tok/s/GPU ($1.28 per million tokens) and MI325X produces 68 ($5.25). B300 is 310% cheaper per token; B300 delivers 665% more tok/s/GPU. (Numbers reflect the default 1k/1k · int4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/gpu) | B300:795.0MI325X:127.3 | B300:643.6MI325X:94.2 | B300:520.4MI325X:68.0 |
| Cost ($/M tok) | B300:$0.836MI325X:$2.789 | B300:$1.020MI325X:$3.800 | B300:$1.281MI325X:$5.251 |
| tok/s/MW | B300:366374MI325X:58415 | B300:296599MI325X:43231 | B300:239817MI325X:31189 |
| Concurrency | B300:~64MI325X:~16 | B300:~64MI325X:~10 | B300:~32MI325X:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.