DRA Driver for NVIDIA GPUs Overview

The DRA Driver for NVIDIA GPUs is a Kubernetes Dynamic Resource Allocation (DRA) driver that enables flexible GPU allocation and provisioning of multi-node NVLink fabrics for Kubernetes workloads. It requires Kubernetes 1.32 or later. Starting in Kubernetes 1.34, DRA is enabled by default. On Kubernetes 1.32 and 1.33, the DynamicResourceAllocation feature gate must be enabled.

The driver manages two types of resources:

• GPUs (gpu.nvidia.com) — allocates full GPUs, Multi-Instance GPU (MIG) slices, and Virtual Function I/O (VFIO) passthrough devices with fine-grained sharing and configuration control.
• ComputeDomains (compute-domain.nvidia.com) — provisions ephemeral multi-node NVLink fabrics using IMEX, enabling pods on different nodes to share GPU memory at full NVLink bandwidth.

The Kubernetes device plugin framework treats hardware resources as opaque countable integers. A workload either gets a whole unit or it doesn’t and there is no way to express sharing, per-workload configuration, capability constraints, or topology requirements within that model.

DRA is a replacement for the device plugin architecture itself, not just a different driver for the same interface. It brings hardware resource management closer to the Persistent Volume model: a workload declares what it needs in a ResourceClaim, and the driver fulfills it. This separation of declaration from consumption enables capabilities that are fundamentally out of scope for the device plugin framework:

• Share a single GPU across multiple containers or pods (time-slicing or MPS)
• Request GPUs by capability — memory size, compute capacity, architecture — rather than just count
• Attach per-workload configuration (sharing strategy, MIG profile, MPS thread limits) directly to the claim
• Allocate MIG devices dynamically without pre-configuring profiles on the node
• Provision scoped, ephemeral NVLink fabrics that exist only for the lifetime of the workload

Use cases

GPU allocation

Request GPUs, MIG slices, or VFIO passthrough devices in a ResourceClaim, with an optional sharing strategy (time-slicing or Multi-Process Service (MPS)). Refer to Architecture for how requests are fulfilled.

Multi-node NVLink with ComputeDomains

Create a ComputeDomain resource to provision an ephemeral NVLink fabric across nodes. Workload pods claim a channel from the domain and receive the IMEX device mounts needed for direct cross-node GPU memory access. The fabric is torn down automatically when the workload finishes.

Get started

• Install and validate

Learn more

• Architecture — components, request flows, and reference links.
• API reference — opaque config types for ResourceClaim specs.
• Feature gates — available feature flags and their defaults.