Kimi K2.5/K2.6/K2.7-Code 1T — B300 vs MI355X
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) 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.
B300 / MI355X on Kimi K2.5/K2.6/K2.7-Code 1T at 38 tok/s/user: 4317 / 4004 tok/s/chip, $0.15 / $0.10 per million tokens. MI355X is 38% cheaper per token; B300 delivers 8% more tok/s/chip.
Around the middle of the 11–123 tok/s/user interactivity band, at 67 tok/s/user on Kimi K2.5/K2.6/K2.7-Code 1T: B300 runs 2747 tok/s/chip at $0.23/M tokens, MI355X runs 2493 at $0.17/M. MI355X is 36% cheaper per token; B300 delivers 10% more tok/s/chip.
Setting 95 tok/s/user as the target on Kimi K2.5/K2.6/K2.7-Code 1T, B300 produces 1848 tok/s/chip ($0.34 per million tokens) and MI355X produces 1739 ($0.24). MI355X is 41% cheaper per token; B300 delivers 6% more tok/s/chip. (Numbers reflect the default 8k/1k · fp4 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B300:4316.9MI355X:4003.6 | B300:2746.6MI355X:2492.7 | B300:1847.7MI355X:1738.7 |
| Cost ($/M tok) | B300:$0.145MI355X:$0.105 | B300:$0.228MI355X:$0.167 | B300:$0.336MI355X:$0.238 |
| tok/s/MW | B300:2272062MI355X:1915595 | B300:1445598MI355X:1192676 | B300:972471MI355X:831906 |
| Concurrency | B300:~51MI355X:~102 | B300:~19MI355X:~17 | B300:~9MI355X:~9 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.