Nvidia open-sources GPU orchestration
Nvidia donated its dynamic resource allocation (DRA) GPU driver to the Kubernetes community and rolled out tooling for confidential containers and large-scale AI orchestration, effectively putting GPU orchestration into open-source hands. The move pairs with industry focus on performance-per-watt and energy-aware AI deployments. (blogs.nvidia.com) (helpnetsecurity.com)
NVIDIA’s official post about its KubeCon announcements was published March 24, 2026 and is credited to Justin Boitano on the company blog. (blogs.nvidia.com) The k8s-dra-driver-gpu repository on GitHub lists compatibility with Kubernetes 1.32 or newer and shows roughly 1,478 commits alongside about 585 stars as of late March 2026. (github.com) Public participants named in coverage of the effort include Amazon Web Services, Broadcom, Canonical, Google Cloud, Microsoft, Nutanix, Red Hat and SUSE collaborating on community governance and integration work. (helpnetsecurity.com) Google also contributed a DRA implementation for Tensor Processing Units (TPUs) at KubeCon, signaling parallel vendor contributions for device-level DRA support beyond GPUs. (cloud.google.com) NVIDIA documented GPU support for Kata Containers via the CNCF Confidential Containers community and showcased a KubeCon session titled “Hybrid‑Confidential‑Cloud” demonstrating confidential GPU workloads with Kata. (blogs.nvidia.com) New open-source tooling announced alongside the driver includes NVIDIA’s aicr repository for validated GPU-accelerated runtimes and references to Grove being integrated with llm-d inference stacks for Kubernetes-native AI deployments. (github.com) The upstreamed codebase exposes primitives such as ComputeDomain and a ComputeDomainClique CRD for sharding Multi‑Node NVLink info, and the driver notes support for NVIDIA Multi‑Process Service (MPS), Multi‑Instance GPU (MIG) features and targeting of Grace Blackwell-class multi‑node training setups. (github.com)
