Platform Support

Use this page to confirm that your hardware, operating system, and Kubernetes platform are supported by NVIDIA Network Operator before deployment.

Support terms used on this page:

• Supported — A configuration that NVIDIA maintains and backs for this release.

• Validated — A hardware or software configuration that NVIDIA has tested with this release.

• GA (Generally Available) — Production-ready support tier.

• Tech Preview — Limited testing; not recommended for production deployments.

NVIDIA Network Operator uses calendar versioning in the form YY.MM.PP (for example, 26.4.0). The first two fields identify the major version and release timeframe; the third field identifies the patch version, used for critical bug and CVE fixes.

When a new major version is released, the previous major version enters a Deprecated state and receives only patch updates for critical fixes. Earlier major versions reach End of Support and no longer receive updates.

Upgrades are supported within a major release, or to the next major release. For upgrade procedures and operational guidance, see Life Cycle Management.

Review the following limitations before deployment:

• Firmware reset on BMC-controlled platforms — NVIDIA DGX/HGX GB200 NVL72, B200, and B300 systems require additional configuration so that firmware updates apply without a reboot loop. See Platforms with external BMC (DGX, HGX GB200/B200/B300, Vera Rubin GB).

• No GPUDirect RDMA on Arm-based validated configurations — NVIDIA IGX Orin and NVIDIA Grace ARM Server are validated for RoCE only, without GPUDirect RDMA. See Support for GPUDirect RDMA.

• Precompiled DOCA-OFED driver containers are currently unsigned.

• Precompiled DOCA-OFED kernel-flavor support — Only the Ubuntu generic kernel flavor is GA. The nvidia, aws, azure, and oracle flavors are Tech Preview.

• For per-release known issues, see the Release Notes.

• RDMA-capable NVIDIA network adapters

  • NVIDIA ConnectX NICs and SuperNICs

  • NVIDIA BlueField Networking Platforms

• NVIDIA GPU Operator v25.3.x or newer – required for workloads that use NVIDIA GPUs and GPUDirect RDMA. See Support for GPUDirect RDMA.

The following adapters have been tested and validated with NVIDIA Network Operator:

The following NVIDIA Data Center systems have been tested and validated with NVIDIA Network Operator:

NVIDIA Network Operator has been validated on the following OS and platform combinations for Ethernet (RoCE) and InfiniBand (IB RDMA):

NVIDIA Network Operator has been validated in the following scenarios:

NVIDIA Network Operator has been validated with the following NVIDIA Spectrum-X Reference Architecture (RA) versions:

NVIDIA Network Operator supports attaching SR-IOV Virtual Functions (VFs) to KubeVirt virtual machines via VFIO PCI passthrough. For deployment instructions, see KubeVirt SR-IOV Integration.

Key limitations:

• VFIO passthrough requires IOMMU-capable hardware (Intel VT-d or AMD-Vi) with IOMMU enabled at boot.

• Live migration is not supported for VMs that use SR-IOV VFIO passthrough.

• Host-side RDMA is not available for VFIO-passed VFs; RDMA inside the guest is provided by the mlx5_core driver shipped with the guest image.

NVIDIA Network Operator enables GPUDirect RDMA between NVIDIA GPUs and supported NICs/SuperNICs when deployed together with NVIDIA GPU Operator. For deployment instructions, see Network Operator Deployment for GPUDirect Workloads. For GPU Operator-side support details, see NVIDIA GPU Operator Platform Support — Support for GPUDirect RDMA.

Requirements:

• NVIDIA GPU Operator v25.3.x or newer

• NVIDIA DOCA-OFED Driver v5.5-1.0.3.2 or newer

• nvidia_peermem kernel module (auto-loaded by recent NVIDIA GPU drivers)

• Supported NVIDIA GPU (Ampere, Hopper, or Blackwell)

• Supported NVIDIA NICs (NVIDIA ConnectX or BlueField)

For background on the underlying technology, see the NVIDIA GPUDirect RDMA documentation.

NVIDIA Network Operator provides the RDMA networking fabric (RoCE and InfiniBand) required by GPUDirect Storage (GDS). GDS-compatible storage clients operate over the RDMA fabric provisioned by Network Operator.

For configuration details (including the ENABLE_NFSRDMA setting and the NVMe inbox kernel module workaround), see DOCA-OFED Driver Container. For background on the underlying technology and the list of GDS-compatible storage clients and version requirements, see the NVIDIA GPUDirect Storage documentation.

Overview

To save startup time and operational effort, precompiled DOCA-OFED driver container images are available for common OS/flavor/kernel/architecture variants.

The container image tag pattern used for common variants is: driver_ver-container_ver-kernel_ver-flavor-os-arch. For example: 24.07-0.6.1.0-0-6.8.0-49-generic-ubuntu24.04-amd64

Supported Operating Systems

Currently precompiled DOCA-OFED driver container images are provided for the following operating systems:

• Ubuntu 24.04 (amd64/arm64)

• Ubuntu 22.04 (amd64/arm64)

Limitations

• NVIDIA supports precompiled driver containers for the most recently released DOCA-OFED GA drivers.

• NVIDIA builds precompiled driver containers for generic, nvidia, aws, azure, and oracle kernel flavors.

• Precompiled driver containers are currently unsigned.

• If your hosts use a different kernel variant, you can create a custom precompiled driver container and host it in your own container registry. Please refer to the Precompiled Container Build Instructions for NVIDIA DOCA-OFED Driver Container section.

The following component versions are deployed by NVIDIA Network Operator:

Container management tools:

• Helm v3

• Red Hat Operator Lifecycle Manager (OLM)

Orchestration & resource scheduling:

• Run:ai