Linux for GPU servers: Ubuntu 24.04 or 26.04, RHEL or Rocky Linux for the NVIDIA stack

The short list for a new GPU server

For a GPU server installed in late 2026, Ubuntu 24.04 LTS, RHEL 9 and RHEL 10 are the operating systems that every NVIDIA support list covers: CUDA, the GPU Operator, NVIDIA AI Enterprise and vGPU. Ubuntu 26.04 LTS is supported by CUDA, the GPU Operator and vGPU, but NVIDIA AI Enterprise 8.2 lists it only inside virtual machines. Rocky Linux is supported by CUDA and the GPU Operator and is absent from the AI Enterprise list.

The choice therefore depends less on taste than on which NVIDIA product needs support on the host. A server that only runs CUDA applications and containers can run any of the five. A server whose software is covered by an NVIDIA AI Enterprise licence, such as NIM, needs an operating system on the AI Enterprise matrix, and a vGPU host needs one on the vGPU matrix.

NVIDIA CUDA 13.4 installation guide for Linux (validated configurations), GPU Operator platform support (14 October 2026), NVIDIA AI Enterprise 8.2 support matrix (2 September 2026), Ubuntu release cycle page, Red Hat errata policy and Rocky Linux version page; all read on 24 October 2026. The RHEL releases in the bare-metal row are those NVIDIA lists for Kubernetes.

The table hides one trap: each list names minor releases, and they do not move together. CUDA 13.4 Update 1 is validated on Rocky Linux 9.8 and 10.2, while the GPU Operator page names Rocky 9.7 and 10.1. Before a minor upgrade, check the list of every NVIDIA component the server runs.

Ubuntu 24.04 LTS: what every NVIDIA list covers today

Ubuntu 24.04 LTS appears on all the lists above. Canonical's release-cycle page gives it standard security maintenance until May 2029 and Expanded Security Maintenance with Ubuntu Pro until May 2034. CUDA 13.4 Update 1 is validated on 24.04.4 with kernel 6.17.0-19, GCC 14.3.0 and glibc 2.39. NVIDIA AI Enterprise 8.2 lists it for bare-metal Kubernetes and standalone containers, as a guest and as a single-node hypervisor, and the GPU Operator supports it with containerd 2.0 to 2.3 and with CRI-O.

An Ubuntu LTS release does not keep one kernel line. For Ubuntu 22.04, the GPU Operator's platform page names kernels 5.15, 6.5 and 6.8, and CUDA 13.4 is validated on 22.04.5 with 6.5.0-45 and on 24.04.4 with 6.17.0-19. Which kernel a server runs depends on the installation and on hardware-enablement updates, so a driver built for one kernel needs checking after each kernel change.

The limit is time. A server commissioned in late 2026 runs past May 2029 in its third year. NVIDIA's CUDA guide states that its support for an operating system continues until that distribution's standard end of security support, so from mid-2029 a 24.04 server needs either Ubuntu Pro for security updates or a release upgrade. Our guide to the GPU server lifecycle over five years puts this next to the driver branch changes and the warranty.

Ubuntu 26.04 LTS: newer kernel, shorter lists

Ubuntu 26.04 LTS came out in April 2026 with standard security maintenance until May 2031, which covers most of a five-year server life. CUDA 13.4 Update 1 is validated on 26.04.1 with kernel 7.0.0-31, GCC 15.2.0 and glibc 2.43. NVIDIA's own systems moved to it: DGX OS 8.0.0, released on 31 August 2026, is based on Ubuntu 26.04 with kernel 7.0.0-1010-nvidia and the R595 driver 595.91.07, as our page on DGX OS 8 on the DGX B300 describes.

Four points on the October 2026 lists hold it back on some servers:

• NVIDIA AI Enterprise 8.2 lists Ubuntu 26.04 for virtualised guests and as a single-node Ubuntu hypervisor, not for bare-metal Kubernetes or standalone containers, where it stops at 24.04.
• The GPU Operator supports containerd on 26.04 but does not list CRI-O for it; precompiled drivers for 26.04 exist for kernel 7.0 with the R595 branch.
• The cryoSPARC guide warns, as of June 2026, that "CryoSPARC versions up to 5.0.6 are incompatible with the version 7 kernel that is the default for Ubuntu 26.04", and recommends Ubuntu 22.04 or 24.04.
• Kernel 7.0 is new, so out-of-tree modules from other vendors, such as storage or network drivers, need their own check.

vGPU is further ahead: NVIDIA's vGPU support matrix lists Ubuntu 26.04 LTS as a host since release 20.1. Our article on cryoSPARC and RELION servers shows how one application can decide the operating system.

Send us the applications the server will run, and we install the Ubuntu release, driver branch and CUDA version that all of them support.

RHEL 9 and RHEL 10: one kernel line per release

Red Hat's errata policy describes a ten-year life cycle for RHEL 8, 9 and 10: five years of Full Support, five of Maintenance Support, then an Extended Life Phase. Red Hat's release table shows RHEL 9.0 on 17 May 2022 and RHEL 10.0 on 20 May 2025, so by our arithmetic RHEL 9 is maintained to about mid-2032 and RHEL 10 to about mid-2035, both beyond a five-year server life.

For GPU drivers the kernel policy matters more. Every RHEL 9 minor release in Red Hat's table runs a 5.14.0 kernel, from 9.0 to 9.8 of 19 May 2026, and every RHEL 10 release a 6.12.0 kernel. Minor releases arrive every six months, in May and November in that table, and each brings a new kernel build in the same line. The NVIDIA lists follow the minors: the GPU Operator names RHEL 9.2, 9.4, 9.6, 9.7 and 9.8, and 10.0 to 10.2, and NVIDIA AI Enterprise 8.2 names 9.4, 9.6, 10.0 and 10.2 for bare-metal Kubernetes. Since GPU Operator 26.7.0, the driver container images for RHEL and Rocky Linux are tagged by major release, rhel9 or rocky10 for example, and the non-precompiled RHEL 9 driver containers exist for x86 only.

To stay on one minor release longer, Red Hat now offers ELCP, which replaces Extended Update Support: six years from general availability for the even minor releases .2, .4, .6 and .8 and nine years for .10. A GPU server pinned to an even minor that NVIDIA lists, such as 9.6, can then wait for a planned change instead of following every release.

RHEL is also the base of Red Hat's AI stack: Red Hat AI Inference Server needs RHEL 9.2 or later, and OpenShift worker nodes run Red Hat Enterprise Linux CoreOS, as our guide to OpenShift AI on GPU servers shows. A RHEL server needs a Red Hat subscription.

Rocky Linux: same kernels, no NVIDIA enterprise support

Rocky Linux rebuilds RHEL and tracks its kernels: the CUDA 13.4 table lists Rocky 9.8 with kernel 5.14.0-687.15 and Rocky 10.2 with 6.12.0-211, the same as RHEL 9.8 and 10.2. CUDA and the GPU Operator support it, cryoSPARC recommends Rocky Linux 8, 9 and 10 alongside Ubuntu, and Rocky Linux gives Rocky 8 an end of life on 31 May 2029 and Rocky 9 on 31 May 2032. The Rocky version page we read has no row for Rocky 10.

Two points separate it from RHEL on a GPU server. NVIDIA AI Enterprise 8.2 does not list Rocky Linux, neither on bare metal nor as a hypervisor, so a server whose NVIDIA software is under an AI Enterprise licence belongs on RHEL or Ubuntu. And Rocky minor releases are short-lived: its version page gives Rocky 9.5, released on 19 November 2024, an end of life in May 2025. A Rocky GPU server therefore follows the minor upgrades, and with it the NVIDIA list for that minor. Rocky suits research and compute servers whose software is supported by CUDA alone and whose team already runs RHEL-like systems.

vGPU hosts: RHEL KVM or Ubuntu

A server that hosts vGPU virtual machines needs a hypervisor on NVIDIA's vGPU matrix. Among Linux hosts, the matrix of 29 September 2026 lists RHEL KVM 8.10, 9.4, 9.6, 9.8 and 10.0 to 10.2, with 9.8 and 10.2 added in vGPU 20.2 and 9.7 listed for 20.0 and 20.1 only, and Ubuntu 20.04 to 26.04 LTS. The RTX PRO 6000 and RTX PRO 4500 Blackwell Server Editions, the L40S and the L4 are on both pages. The RTX PRO 5000 Blackwell Workstation Edition 72 GB appears only on RHEL KVM 9.6, since release 20.2, and the H200 NVL on neither page.

NVIDIA AI Enterprise 8.2 supports the RHEL KVM, Ubuntu and Nutanix AHV hypervisors as single-node only. VMware ESXi 8.0 and 9.0 are listed without that restriction; our article on GPUs in vSphere, passthrough or vGPU covers that route.

Kubernetes distributions and container runtimes

On Kubernetes, the GPU Operator's platform page narrows the choice per distribution. Upstream Kubernetes 1.33 to 1.37 is supported on all the systems above. Rancher RKE2 covers Ubuntu and RHEL 9 and 10, not RHEL 8; K3s, Mirantis k0s and Canonical MicroK8s are listed for Ubuntu only. For the container runtime, containerd 2.0 to 2.3 is listed for Ubuntu and RHEL 8 and 9, and CRI-O for Ubuntu 22.04 and 24.04 and RHEL 8 and 9, not for Ubuntu 26.04. NVIDIA AI Enterprise 8.2 lists upstream Kubernetes 1.32 to 1.36, one release behind the Operator; the serving stack on such a cluster is the subject of our private LLM platform on Kubernetes.

Kernel updates need a rule on every distribution. For Ubuntu 22.04, NVIDIA's platform page recommends disabling kernel updates through unattended-upgrades so that a node does not move to a kernel the driver containers do not support. The same logic applies to a driver installed on the host: pin the kernel and driver together and update both in a maintenance window, as our guide to NVIDIA driver branches and CUDA versions explains. Secure Boot and signed kernel modules are covered in our guide to the NVIDIA driver with Secure Boot on Linux servers.

Which Linux for which GPU server

• Inference or training server with NVIDIA AI Enterprise software such as NIM: Ubuntu 24.04 LTS or RHEL 9 or 10 on a minor NVIDIA lists. Plan the move off 24.04 before May 2029, or Ubuntu Pro.
• Kubernetes GPU nodes without AI Enterprise: Ubuntu 24.04 or 26.04 with containerd, or RHEL; Ubuntu 26.04 once every operator and module on the node lists it. Our Kubernetes DRA guide adds the driver and runtime requirements for DRA.
• Company standardised on Red Hat or OpenShift: RHEL 9 or 10, pinned to an even minor release; OpenShift nodes run CoreOS.
• Research server with CUDA applications only: Ubuntu 24.04 or Rocky Linux 9; check each application, as the cryoSPARC kernel note shows.
• vGPU host without vSphere: RHEL KVM or Ubuntu on the vGPU matrix, single node.

What we supply

We build AI servers to order with RTX PRO 6000 Server Edition, H200 NVL, L40S or L4 cards, assembled and tested before delivery and delivered across the EU on one contract and invoice. On request they arrive with the operating system, driver, CUDA and a container runtime installed, on the distribution and release your team runs. Tell us your distribution, release and the NVIDIA software you license, and our team replies within one business day with a configuration whose versions are on each NVIDIA list.

FAQ