Guides on DRA Driver for NVIDIA GPUs

GPU sharing

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

Coming soon — this page will cover the three GPU-sharing modes the driver supports:

Time-slicing — round-robin scheduling of full GPUs across pods.
MPS — NVIDIA Multi-Process Service for concurrent kernel execution.
MIG — hardware-partitioned Multi-Instance GPU slices.

It will explain when to pick each, the trade-offs, and how to request them through ResourceClaims.

VFIO passthrough

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

Coming soon — this page will document the VFIO passthrough path for VM workloads.

ComputeDomain workloads

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

For background on what a ComputeDomain is and how it fits together, see ComputeDomains.

Prerequisites

See Prerequisites for hardware and software requirements, including the ComputeDomain-specific requirements for Multi-Node NVLink hardware, GPU Feature Discovery, and nvidia-imex service configuration.

Create a ComputeDomain

The minimal ComputeDomain spec requires only the name of the ResourceClaimTemplate the controller will create for channel allocation:

apiVersion: resource.nvidia.com/v1beta1
kind: ComputeDomain
metadata:
 name: my-compute-domain
spec:
 numNodes: 0
 channel:
 resourceClaimTemplate:
 name: imex-channel-0

numNodes is deprecated. Set it to 0 (the recommended value when IMEXDaemonsWithDNSNames is enabled — its default state).

Observability

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

Coming soon — this page will list driver metrics, log fields, and the events you can rely on for monitoring.

Admission webhook

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

Coming soon — this page will cover the webhook’s validation rules for opaque ResourceClaim config and how to interpret rejection messages.