Head-to-head AI inference benchmark comparison of B200 (NVIDIA Blackwell) and MI355X (AMD CDNA 4) 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
Throughput at 48 tok/s/user on Kimi K3 2.8T: B200 hits 6774 tok/s/chip, MI355X hits 10299. Per-million costs land at $0.07 and $0.04 respectively. MI355X is 75% cheaper per token; MI355X delivers 52% more tok/s/chip.
B200 / MI355X on Kimi K3 2.8T at 87 tok/s/user: 3320 / 6880 tok/s/chip, $0.14 / $0.06 per million tokens. MI355X is 139% cheaper per token; MI355X delivers 107% more tok/s/chip.
Toward the upper edge of the 10β163 tok/s/user interactivity band, at 125 tok/s/user on Kimi K3 2.8T: B200 runs 2217 tok/s/chip at $0.22/M tokens, MI355X runs 3213 at $0.13/M. MI355X is 67% cheaper per token; MI355X delivers 45% more tok/s/chip. (Numbers reflect the default agentic-traces Β· fp4 selection for this URL β table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B200:6774.2MI355X:10299.0 | B200:3319.8MI355X:6879.8 | B200:2216.9MI355X:3213.0 |
| Cost ($/M tok) | B200:$0.071MI355X:$0.040 | B200:$0.145MI355X:$0.061 | B200:$0.217MI355X:$0.130 |
| tok/s/MW | B200:3961513MI355X:4927748 | B200:1941387MI355X:3291749 | B200:1296406MI355X:1537300 |
| Concurrency | B200:~22MI355X:~42 | B200:~10MI355X:~35 | B200:~6MI355X:~5 |
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.
