NVIDIA adds enforced GPU memory limits for Kubernetes sharing

NVIDIA says CUDA 13.4 adds enforced memory limits for workloads sharing a GPU, with Kubernetes support through the new open-source GPU Fractioning Operator. Workloads can request a fixed memory amount or a guaranteed minimum with room to grow when memory is free. The operator applies limits on each GPU node and measures each workload's usage. KAI Scheduler integrates it starting with version v0.18.0; other schedulers can use it by writing the required workload annotations. Workloads can take turns or run concurrently, but both modes cannot share one GPU. Concurrent workloads share an error domain, while time-sliced workloads have separate error isolation.