NVIDIA donates GPU DRA to CNCF
NVIDIA donated its GPU Dynamic Resource Allocation (DRA) driver to the CNCF to standardize GPU resource management for Kubernetes AI workloads, aiming for better efficiency and isolation in multi‑tenant clusters. The move could simplify GPU orchestration and adoption of Kubernetes-native AI patterns across MLOps teams. (ejscomputers.com)
The news was announced at KubeCon + CloudNativeCon Europe in Amsterdam during the conference running March 23–26, with coverage dated March 24, 2026. (helpnetsecurity.com) The codebase is published on GitHub as NVIDIA/k8s-dra-driver-gpu and shows roughly 1,478 commits and about 585 stars on the main repository page. (github.com) NVIDIA’s upstream documentation and the repository README describe the driver as two largely independent kubelet plugins — gpu-kubelet-plugin and compute-domain-kubelet-plugin — and the driver appears in NVIDIA’s NGC catalog at version v25.12.0. (github.com) (catalog.ngc.nvidia.com) Core capabilities listed in the project and NVIDIA docs include controlled sharing and dynamic reconfiguration of GPUs, dynamic MIG management, ComputeDomains for Multi‑Node NVLink, and internode memory exchange (IMEX) primitives. (docs.nvidia.com) (github.com) Upstream Kubernetes work treats DRA as part of the device allocation roadmap — the repository specifies operation on Kubernetes 1.32 or newer, and CNCF notes DRA as a building block retained into the Kubernetes v1.35 release stream. (pkg.go.dev) (cncf.io) Cloud integrations and operator support are already appearing in docs and vendor blogs: Azure posted a walkthrough for DRA with vGPUs on AKS on March 6, 2026, and NVIDIA’s NIM and GPU Operator documentation include DRA installation and usage guidance. (blog.aks.azure.com) (docs.nvidia.com) The repository and NGC catalog state the project is licensed under Apache‑2.0 and accepts contributions with a Developer Certificate of Origin (DCO), indicating maintainers expect community pull requests and upstream collaboration. (catalog.ngc.nvidia.com) (github.com)
