NUMA Non-Uniform Memory Access
A multi-socket system where each CPU has "its own" RAM — accessing local memory is faster than cross-socket access over the interconnect.
Dual-socket Xeon/EPYC servers are the typical case. On Linux, numactl manages NUMA allocations. The AMD Zen architecture is NUMA within a single socket (chiplet CCDs have their own "nearby" memory). Database servers and high-performance apps need to pin threads to sockets and allocate local memory — otherwise performance drops dramatically. perf stat -e mem_load_uops_retired:l3_miss shows NUMA traffic. Kubernetes has a topologyManager to manage this in containers.