DGX OS 8 on the NVIDIA DGX B300: what changes from DGX OS 7 and how a system gets there

What DGX OS 8.0.0 brings to the DGX B300

DGX OS 8 is NVIDIA's operating system for DGX servers on Ubuntu 26.04 LTS. Its first release, 8.0.0 of 31 August 2026, supports the NVIDIA DGX B300 and brings the R595 driver 595.91.07, a kernel based on Linux 7.0 and CUDA 13.3 Update 1 as a validated package that is installed after deployment. A DGX B300 on the latest DGX OS 7 release can be upgraded in place; NVIDIA advises upgrading a few systems first and checking the applications on them.

NVIDIA describes DGX OS 8 as "based on Ubuntu 26.04" and incorporating "the NVIDIA BaseOS operating environment", here BaseOS 9. The version scheme in its release guidance explains the jump: major releases track Ubuntu (DGX OS 6 on Ubuntu 22.04, DGX OS 7 on 24.04, DGX OS 8 on 26.04) and may include changes that are not backward compatible, while minor releases align with BaseOS releases, which define the NVIDIA drivers.

NVIDIA DGX OS 8 release notes, version 8.0.0 (release 31 August 2026; software table updated 17 September 2026; page updated 15 October 2026), read on 24 October 2026. NCCL 2.31.2 and GDRCopy 2.6 are listed as validated, installed after deployment.

The feature list adds several changes an administrator notices. Docker Engine now comes from Ubuntu's docker.io package, which NVIDIA says gives "faster security fixes through the Ubuntu repositories". Chrony is the default time daemon, and OpenSSH, SSSD and Samba are the Ubuntu 26.04 versions. DGX OS 8 supports Ubuntu's inbox OFED stack beside DOCA, and adds TuneD for profile-based system tuning. It also "includes an Ubuntu Pro Client Extended Security Maintenance (ESM) subscription for DGX customers".

The driver matters for planning. In NVIDIA's driver lifecycle table of 9 September 2026, R595 is the production branch, supported to March 2027, and 595.91.07 is the R595 version that NVIDIA's security bulletin of 30 September 2026 names as fixed. Our guide to NVIDIA driver branches and CUDA versions explains the branch types and how containers with a newer CUDA run on this driver.

Which DGX systems DGX OS 8 supports

The 8.0.0 release notes give each system a status, the earliest DGX OS ISO release NVIDIA lists for it and a firmware release.

NVIDIA DGX OS 8 release notes 8.0.0, supported systems table, read on 24 October 2026. Also supported: DGX H800, DGX A100 and DGX A800. Not supported: DGX Station A100 and A800, DGX-2, DGX-1 and DGX Station (V100).

All supported systems are x86. The Arm systems, among them the DGX GB300 and the DGX Spark, are marked "Future release", so a DGX Spark stays on DGX OS 7 for now; our DGX Spark day-one guide covers its own update path.

A new DGX B300 does not necessarily arrive on DGX OS 8. The feature summary in NVIDIA's DGX B300 user guide, updated on 21 October 2026, still lists "DGX OS 7 based on Ubuntu 24.04 LTS", with "Additional support for Ubuntu and Red Hat Enterprise Linux 8 and 9, Rocky Linux". Plan the commissioning of a new system as if it may need the upgrade below, and ask at delivery which release is installed.

DGX OS 8.1.0 is not a B300 release

Since 22 September 2026 NVIDIA also publishes DGX OS 8.1.0. Its release notes list one system, the preproduction DGX Vera Rubin NVL72, with BaseOS 10, the R615 driver 615.71.09 and CUDA 13.4.1, and they call it a technical preview that "is not intended for production use". Two of its statements are easy to misread. The 8.1.0 notes say that upgrades from Ubuntu 24.04 to Ubuntu 26.04 are not supported, and that to reach a future release you "reimage the system with a newer BaseOS release". Both apply to that preview on that system; the 8.0.0 guide documents the in-place upgrade from DGX OS 7 for the B300.

The online guide's "latest" pages now open on 8.1.0 and carry its preview notice. For a DGX B300, read the version-numbered 8.0.0 pages.

Upgrading a DGX B300 from DGX OS 7 to DGX OS 8

NVIDIA's "Upgrading the OS" page for 8.0.0 supports an in-place release upgrade "only from the latest DGX OS 7 release"; "A direct upgrade from DGX OS 6 to DGX OS 8 is not supported". Before starting, it asks for four things:

• Read the release notes, and upgrade a subset of systems first.
• Prefer the console, directly or through the BMC. SSH is possible, but a DHCP address can change during the reboot and a VPN can drop the session, and the upgrade tool itself advises against SSH because a failure is harder to recover.
• If the system is connected to a 172.17.x.x subnet, use the console directly: Docker, which DGX OS installs, uses that range by default, and NVIDIA's page moves Docker to another subnet after the upgrade.
• Confirm with wget that the Ubuntu, BaseOS, CUDA and DOCA repositories are reachable; NVIDIA's check should create six non-empty files.

The sequence itself is short:

1. Bring DGX OS 7 to its latest release with sudo apt update and sudo apt -y full-upgrade.
2. Install the upgrade tool, apt install -y nvidia-release-upgrade.
3. Remove XPMEM with sudo apt purge -y xpmem-dkms: "The XPMEM module shipped in DGX OS 7.6.0 is not compatible with Linux kernel 7.0", and a newer version is installed later.
4. Start the upgrade with sudo nvidia-release-upgrade --baseos 9, or with sudo -E behind a proxy; 9 is the only accepted BaseOS value.
5. Answer the prompts and reboot when asked.

During the upgrade, NVIDIA says to keep the local versions of /etc/ssh/sshd_config and two GDM files. The upgrader starts a second sshd on port 1022 as a fallback and disables third-party package sources, keeping Canonical's and the DGX repositories; re-enable the others afterwards. The removal of obsolete packages can take hours, and NVIDIA asks you to review the list before answering "y".

Afterwards the kernel should be "at least 7.0.0-1010-nvidia" and nvidia-smi should report a 595 driver; NVIDIA's example shows 595.91.07 with CUDA 13.3.1. If the upgrade is interrupted, the page gives a recovery sequence: dpkg --configure -a, apt -f install, a restart of Ubuntu's release upgrader from its temporary directory and then nvidia-post-release-upgrade, with the reboot only after that last step. One known issue concerns the package step: DOCA DKMS packages can pull in the headers of the 7.0.0-1005-gke kernel and make the configuration fail, so NVIDIA's known-issues page says to run sudo apt install linux-nvidia -y before apt full-upgrade.

We supply the DGX B300 per project, on one EU contract and invoice; send us the company, the project and the intended use, and we check price and availability for your project.

Firmware before the operating system

The release notes list firmware release 26.07.1 next to the DGX B300. By our reading, a B300 should run that firmware release or a later one before it moves to DGX OS 8. NVIDIA's firmware update guide for the B300, updated on 22 September 2026, lists eight releases from 25.11.1 to 26.09.1, and both 26.07.1 and 26.09.1 name "NVIDIA DGX OS 7.5.0 and later" as the supported operating system, so the firmware can be brought up to date while the system still runs DGX OS 7.

NVIDIA DGX B300 System Firmware Update Guide, version pages 26.07.1 and 26.09.1 (updated 22 September 2026), read on 24 October 2026. 26.09.1 also updates the HMC, ConnectX-8 and NVSwitch firmware on the GPU tray.

The tool is nvfwupd, downloaded from the NVIDIA Enterprise Support Portal. It talks to the Redfish API of the BMC, so it can run from an admin host. NVIDIA's sequence updates the BMC first, then the motherboard tray components, then the GPU tray, which holds the VBIOS, NVSwitch, ERoT and FPGA firmware, and ends with an AC power cycle. All GPU activity must stop first, nvidia-smi included, because it "can block the VBIOS update", and "A high workload can disrupt the firmware update process and result in an unusable component." For an update from 26.08.1, NVIDIA gives 34 minutes 43 seconds for the motherboard tray in parallel mode and 10 minutes 47 seconds for the GPU tray, which sets the maintenance window. Our guide to GPU server firmware updates explains the steps and where other server makers publish theirs.

Reimaging does not replace this. NVIDIA's reimaging page says the installer updates ConnectX firmware, while other system firmware is not updated. After installing DGX OS 8, NVIDIA also asks you to reset the default password of the BlueField-3 BMC, with a bf.cfg file and the bfb-install tool.

Reimaging with the DGX OS 8 ISO

A clean install is the other route. The ISO for 8.0.0 is DGXOS-8.0.0-2026-08-24-16-17-42-amd64.iso; it comes from NVIDIA Enterprise Support, which needs an active support account, and its MD5 hash is checked against the release announcement. It boots through the BMC's virtual media or from a UEFI-bootable USB drive or DVD.

"Reimaging the system erases all data stored on the OS drives", /home included. The data drives form a RAID array mounted at /raid, which NVIDIA describes as a cache; the installer can rebuild it or leave it intact. Root filesystem encryption is chosen in the boot menu, needs a passphrase at the console on every boot, and "cannot be enabled or disabled after the installation". Persistent hardware settings such as the MIG configuration and data drive encryption are not changed. NVIDIA also documents installing the BaseOS software on a standard Ubuntu 26.04, for custom partition schemes and automated cluster deployment.

Updates within DGX OS 8

NVIDIA's release guidance says "System upgrades are cumulative": the repositories carry fixes between numbered releases, and by default a system cannot be pinned to one DGX OS version. "NVIDIA recommends that you do not update packages individually", and the upgrade page adds "Hold packages only in extremely rare cases." The routine is sudo apt update, a dry run with sudo apt full-upgrade -s, then the upgrade and a reboot. Without the reboot after a driver update, nvidia-smi fails with "Driver/library version mismatch".

The GPU driver, the CUDA Toolkit and DOCA change only with a supported BaseOS release. On an R595 system the driver therefore follows the branch, and a change to another branch arrives with a new BaseOS. The Ubuntu Pro ESM subscription is attached with sudo pro attach and a token from NVIDIA Enterprise Support. For security news, NVIDIA points to Ubuntu's security notices and its own Product Security page; our guide to securing a GPU server lists the NVIDIA bulletins of autumn 2026.

The 8.0.0 known-issues page lists a few points for operators. NvKmsKapiDevice allocation errors from the GPU driver "can be safely disregarded". General-purpose hypervisors such as VMware ESXi and KVM "are not intended for use on DGX systems and have not been tested", outside supported confidential computing setups. And nv-disk-encrypt can fail on pre-owned NVMe drives until ownership is taken with sedutil-cli, which can erase the drive.

Planning the change on a B300 in production

For a DGX B300 that already serves models, the order follows from NVIDIA's documents:

1. Record the installed firmware with show_version and bring it to 26.07.1 or later, still on DGX OS 7.
2. Bring DGX OS 7 to its latest release, as the upgrade page requires.
3. Upgrade one system, or the less critical one of a pair, and run your own containers against driver 595.91.07, NCCL 2.31.2 and DOCA-OFED 3.4.1; NVIDIA warns that its ConnectX and DOCA-OFED updates can affect applications that depend on them.
4. Install the CUDA Toolkit 13.3 Update 1 and the other validated packages that your workloads use on the host.
5. Record the new versions, then upgrade the remaining systems.

Containers bring their own CUDA runtime, so most framework images keep working on the new driver; images built for a newer CUDA than 13.3 rely on the compatibility rules in our driver guide. For two B300 systems serving one model, our page on high-availability LLM inference on two nodes shows how to keep the service up while one is in maintenance. The DGX B300 itself, its power and its software lines are in our DGX B300 specifications.

What we supply

We supply the NVIDIA DGX B300 per project in the EU, on one EU contract and invoice. Before we can quote, the distribution asks us for the company, the project and the intended use; tell us about your project, and we check price and availability for it. The ordering steps are in our guide to buying a DGX B300 in the EU. Where an eight-card PCIe server fits the workload better, we build AI servers to order, assembled and tested before delivery, and install the operating system, drivers, CUDA and a container runtime on request; our comparison of an 8× H200 NVL server and the DGX B300 helps with that choice.

FAQ