Memory-enforced GPU sharing in Kubernetes with NVIDIA CUDA 13.4
AINVIDIA's technical blog describes memory-enforced GPU sharing in Kubernetes using CUDA 13.4, letting platform teams run multiple AI workloads on a single GPU. The excerpt says that giving each workload a whole device can leave capacity idle, and that NVIDIA GPUs already support sharing approaches such as multi-instance GPU (MIG), which splits a GPU into hardware-isolated instances. The provided text does not give further details on the new mechanism.
Source details
NVIDIA Technical Blog · Elizabeth Goodman · developer.nvidia.comPublished · added here