Memory-Enforced GPU Sharing in Kubernetes with NVIDIA CUDA 13.4
CUDA 13.4 introduces memory-enforced GPU sharing for Kubernetes workloads
NVIDIA's CUDA 13.4 adds memory-enforced GPU sharing in Kubernetes, allowing multiple AI workloads to safely share a single GPU with hard memory boundaries. This extends beyond existing MIG support to offer more flexible, software-enforced partitioning. For platform teams running many small AI workloads, this could meaningfully improve GPU utilization and reduce infrastructure costs.