Donating the NVIDIA DRA driver for GPUs to Kubernetes

[Kevin Klues]

Hi SIG-Node Folks,

We are writing to explore the possibility of donating the NVIDIA DRA Driver for GPUs repository to the Kubernetes project under SIG-Node stewardship.

What it is: This is a Dynamic Resource Allocation (DRA) driver that enables flexible GPU allocation and orchestration in Kubernetes.

It provides two primary capabilities:

• GPU device allocation — dynamic GPU management including support for static MIG partitioning (dynamic MIG partitioning is an alpha feature)
• ComputeDomains — an abstraction to enable secure isolation for high-bandwidth GPU-GPU memory sharing across multi-node workloads over Multi-Node NVLink (MNNVL)

The project is Apache 2.0 licensed, targets Kubernetes 1.32+, and is actively maintained.

Why SIG-Node: As DRA matures as a core Kubernetes feature, having a well-tested, full featured driver within the Kubernetes org would benefit the broader community by:

1. Providing a canonical example of a non-trivial DRA driver implementation
2. Enabling tighter collaboration between the team maintaining this component and upstream Kubernetes maintainers on the DRA API surface and future DRA directions
3. Broadening the maintainer base and ensuring long-term sustainability
4. Helping to ensure Kubernetes scheduling and node management fully support GPUs as first-class resources going forward
5. Serving as a reference implementation for GPU acceleration in the Kubernetes AI conformance program

Current state: The driver has 5 components (2 kubelet plugins, a controller, a dynamically provisioned set of compute-domain daemons, and a webhook), an active CI pipeline, and regular releases. We believe it is mature enough for community-driven development.

We are actively working on options to expand our community owned/driven prow infrastructure to allow for testing the features in this repo as well.

Looking forward to your thoughts.

Thanks,

Kevin Klues
Davanum Srinivas
(with our Nvidia Hats on)