RAM Coffers: NUMA-Aware LLM Inference — Why Hardware Topology Still Matters

RustChain’s RAM Coffers targets NUMA-aware LLM inference on multi-socket servers.

It detects memory topology, places weights in local memory, and pins threads to matching CPU cores to cut cross-node access.

The system also records hardware attestation so older enterprise servers can be used more efficiently for decentralized inference and earn network rewards.

The approach highlights how memory layout can materially affect LLM throughput on surplus hardware.