GPU workloads often need hardware-aware placement to avoid silent latency regressions. This post explains what NUMA locality means for Kubernetes GPU nodes, how CPU Manager, Topology Manager, and Memory Manager work together, and the operational gotchas to watch for.