AMD ATOM serves Kimi-K3 AgentX on MI355X
AMD details how its ROCm-native inference engine, ATOM, serves agentic coding workloads like Kimi-K3 on eight AMD Instinct MI355X GPUs. The approach combines memory optimizations including LMCache and prefix caching, scheduling adjustments via PrefillDelayer, and compute techniques such as Mono decode to maintain high token throughput and interactivity. Open article →