Red Hat OpenShift Setup Guide
${product_name} is compatible with Red Hat OpenShift Container Platform (RHOCP), enabling GPU-accelerated workloads to be deployed and managed within enterprise Kubernetes environments.
Red Hat OpenShift is an enterprise Kubernetes platform that provides developer and operational tooling on top of Kubernetes. Use the official Red Hat OpenShift installation guide to install and prepare the OpenShift cluster itself. This guide covers the next step: enabling supported GPUs on that cluster by using the available Kubernetes integration components.
Dynamic Resource Allocation (DRA) is the strategic direction for Intel GPU enablement on Kubernetes. It expresses GPUs as structured, selectable devices rather than opaque counters, which is what makes more complex workload requirements possible, such as aligning a GPU allocation with a particular NIC or PCIe root complex. Prefer DRA for new deployments. The device plugin remains fully supported and is the right choice when you need the longest-established, OpenShift-certified path.
Choose one of the following enablement approaches based on your cluster requirements and the maturity level you are comfortable operating in:
Enablement approach |
Maturity |
Recommended use case |
Summary |
|---|---|---|---|
Beta |
New deployments, and any workload that needs fine-grained device selection or topology-aware placement |
Uses the Kubernetes DRA framework, exposing per-device attributes and capacity that workloads select with CEL expressions. |
|
Beta |
Deploying the DRA stack together with XPU Manager telemetry, Kueue, and firmware management from one operator |
Provides a unified operator that deploys either the DRA-based stack or the device plugin, along with its supporting components. |
|
General Availability, certified on OpenShift |
Deployments that need the most established and OpenShift-validated path |
Uses the traditional Kubernetes device plugin model and is managed through OpenShift Operator Lifecycle Manager (OLM). |
Validated configuration
The procedures in this guide were validated on a Single-Node OpenShift (SNO) system with eight Intel® Arc™ Pro B70 GPUs:
Component |
Version |
|---|---|
OpenShift Container Platform |
4.22.14 |
Kubernetes |
1.35.6 |
RHEL CoreOS |
9.8.20260908-0, kernel 5.14.0-687.46.1.el9_8.x86_64 |
CRI-O |
1.35.7 |
Helm |
4.0.2 |
Node Feature Discovery Operator |
4.22.0 |
cert-manager Operator for Red Hat OpenShift |
1.20.0 |
Intel Device Plugins Operator |
0.36.0 ( |
Intel GPU Base Operator |
0.4.0 |
Intel GPU Resource Driver |
0.12.0 |
GPUs |
Eight Intel® Arc™ Pro B70 (Battlemage G31, |
The Integration with Red Hat OpenShift AI chapter lists the additional versions validated for the RHOAI stack.
Prerequisites
Before you deploy any Intel GPU enablement approach, ensure that you have the following:
Red Hat OpenShift Container Platform 4.14 or later. We recommend bare-metal multi-node clusters.
Supported Intel Xe-based GPUs installed on worker nodes.
A functioning xe kernel driver on the host. The xe driver is included in-tree starting with RHEL 9.x kernels (kernel 5.14.0-570+) shipped with OpenShift 4.21+. For older OpenShift releases, verify that the xe module is available in your kernel version by using
lsmod | grep xe.An Intel compute runtime in the workload’s container image that supports the GPU. For Intel® Arc™ Pro B70 Graphics (PCI ID
0xe223),libze-intel-gpuversions earlier than 26.x do not expose a GPU platform, even when the device nodes are correctly injected into the container. See Verifying GPU Access for images known to work.Node Feature Discovery (NFD) Operator installed.
Helm 4.0 or later if you deploy the Intel GPU Base Operator or the Intel GPU Resource Driver from their Helm charts. The Intel Device Plugins Operator is installed through OLM and does not need Helm.
Warning
Intel® Arc™ Pro B70 Graphics requires OpenShift 4.22 or later. In OpenShift 4.21, the RHEL 9.6 kernel does not include support for the B70 GPU in the xe driver, so the device may appear in lspci but does not bind to the driver and no /dev/dri/ nodes are created. This issue is resolved in the RHCOS 9.8 kernel shipped with OpenShift 4.22.
The following chapters cover node discovery, Intel GPU enablement options, validation, troubleshooting, and integration with Red Hat OpenShift AI:
References
For more information about OpenShift, Intel GPU enablement, and RHOAI, see the following resources: