HAMi enforces NVIDIA T4 GPU memory limits in Kubernetes

itnext.io7 min read

The author examines how HAMi enforces hard memory limits and soft compute throttling for shared workloads on NVIDIA T4 GPUs in Kubernetes. Testing showed that HAMi successfully stopped greedy pods at memory limits while compute limits operated as soft runtime throttles that behaved inconsistently depending on concurrent workloads. The evaluation also highlights configuration caveats such as potential conflicts between HAMi and NVIDIA device plugins. Open article →