GPU workloads on DRA Driver for NVIDIA GPUs

View available GPU resources

Mon, 01 Jan 0001 00:00:00 +0000

This guide shows how to inspect the ResourceSlice objects and pools the DRA Driver for NVIDIA GPUs publishes, so you can see which GPU, MIG, and VFIO devices are allocatable on your cluster and confirm each node is advertising the devices you expect.

For what each device attribute and capacity field means and the exact names to use in CEL selectors, refer to ResourceSlice device attributes. For details on how the driver publishes ResourceSlices, refer to Publishing GPUs in ResourceSlices. For the generic DRA model (what ResourceSlices, DeviceClasses, and ResourceClaims are and how the scheduler uses them), refer to the Kubernetes DRA documentation.

Request full GPUs

Mon, 01 Jan 0001 00:00:00 +0000

This guide covers how to request full GPUs (gpu.nvidia.com) for workloads, from requesting any available GPU to targeting a specific model or memory size with CEL selectors.

A ResourceClaimTemplate defines the GPU request. You reference it from a pod in spec.resourceClaims. Kubernetes creates one ResourceClaim per pod when it is scheduled. Refer to the Kubernetes DRA documentation for more details on the ResourceClaim and ResourceClaimTemplate.

For an overview of GPU allocation, refer to the GPU allocation concept documentation.

Time-slicing

Mon, 01 Jan 0001 00:00:00 +0000

Time-slicing lets multiple containers share one physical GPU by scheduling their CUDA contexts in turns. CUDA preempts each context at a configurable interval and switches to the next one, similar to CPU time-sharing on a single-core machine.

Use time-slicing when you want to increase GPU utilization across workloads that do not run continuously — for example, multiple batch jobs or development containers that would otherwise sit idle waiting for their own dedicated GPU. For how time-slicing compares to MIG and the other sharing modes, refer to Choosing a resource type.

MIG (Multi-Instance GPU)

Mon, 01 Jan 0001 00:00:00 +0000

MIG lets you partition a single NVIDIA GPU into multiple isolated GPU instances, each with a fixed allocation of compute resources, memory, and memory bandwidth. Unlike time-slicing, MIG instances have hardware-level isolation. Each instance’s memory and compute are fully separate.

Use MIG when you need predictable, isolated GPU resources for multiple concurrent workloads on the same physical GPU. For how MIG compares to time-slicing and the other sharing modes, refer to Choosing a resource type.

KubeVirt VFIO GPU passthrough

Mon, 01 Jan 0001 00:00:00 +0000

KubeVirt can attach VFIO GPU devices to virtual machines through Kubernetes Dynamic Resource Allocation (DRA). This guide covers the DRA Driver for NVIDIA GPUs configuration needed for KubeVirt passthrough workloads.

For KubeVirt feature gates, VM fields, and VM examples, refer to the KubeVirt user guide.

Feature status

This guide uses two Alpha feature gates, both disabled by default:

Feature gate	Default	Stage
`PassthroughSupport`	`false`	Alpha
`DeviceMetadata`	`false`	Alpha

Refer to the feature gates reference and constraints for details.