DGX Spark/Guides/Recovery and reinstall
Guide · November 2026

DGX Spark recovery and DGX OS reinstall: repair first, reflash last (November 2026)

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

A DGX Spark Founders Edition has no installer ISO: the reinstall path is NVIDIA's System Recovery Image, version 1.155.49 (released 21 October 2026, a 5.9 GB tar.gz), written to a USB stick of 16 GB or more, which reflashes firmware and erases the internal SSD back to factory state. Try System Repair first, because it keeps your data: boot the previous kernel from the GRUB menu, or boot an Ubuntu ARM64 live USB to finish a broken package install or copy files off. The three kernel-update failures owners documented on NVIDIA's forum in 2026, listed below, were all fixed without reflashing. ASUS, Dell and the other OEMs supply their own images. A rented GPUwerk Spark needs none of this: the host is the operator's job, and a rebuild restores your container image.

MethodUse it whenErases the SSD?You needSource
Previous kernel from GRUBA new kernel will not boot, an older one is still installedNoKeyboard, displaySystem Repair
Recovery-mode root shellHalf-finished upgrade, missing initramfs, unsigned driverNoKeyboard, displaySystem Repair
Ubuntu ARM64 live USBGRUB unreachable; you need to copy data off; filesystem checkNoUSB stick, Ubuntu 26.04.1 ARM64 imageSystem Repair
NVIDIA recovery image 1.155.49Founders Edition beyond repair, or a clean startYes, OS and firmware to factory16 GB+ USB stick, wired USB keyboard, displaySystem Recovery
OEM imageASUS, Dell, HP, Lenovo, MSI, Gigabyte, Acer machinesYesThe maker's downloadDell KB 000382042
Field Diagnostic, then RMAHardware fault suspectedNodgx-spark-fieldiag, about 30 minutesSupport page

Read 7 November 2026. Recovery image version, date and size from NVIDIA's download page (the file is 6,372,258,034 bytes). NVIDIA says the System Repair procedure applies to Founders Edition and OEM systems; System Recovery applies to the Founders Edition only.

Repair or recover: which do you need?

Repair, unless the SSD's contents are worthless to you or repair has failed. NVIDIA's System Recovery page is blunt that recovery "erases the internal SSD" and points to System Repair "to repair the installed system or back up files without erasing the SSD". The failures that recur on the forum are software faults, and most left the disk readable:

Reported failureCause foundWhat fixed itSourceDate
Stuck in update or boot sequenceAn update rolled out between 5:30 AM and 7:30 PM PST left some units partially updated, NVIDIA saidReflash with the recovery imageNVIDIA staff, thread 3479632025-10-16
Kernel panic, "Unable to mount root fs"Kernel 6.17.0-1014 installed without its initrd after a dashboard update; GRUB timeout 0 hid the menuLive USB, chroot, finish the installowner, thread 3689392026-05-04
Same panic on new unitsKernel 7.0.0-1019 missing its initrd; DKMS conflictNVIDIA's package command (below)thread 3834502026-09-17
Reboots one to five minutes after bootKernel 7.0.0-1024Boot 1019 from GRUB, hold the kernel packagessoftware updates guide2026-10-30
Password prompt before the OSOwner set and forgot a UEFI passwordNothing the owner could do; RMA declined, he reportedowner, thread 3860272026-11-03

Owner reports are single accounts on NVIDIA's developer forum, not NVIDIA statements, except where the source says NVIDIA staff.

How do I boot an older kernel or repair a broken update?

Power off, then power on while holding Shift and Insert together; NVIDIA's System Repair page says that brings up GRUB, where "Advanced options for DGX OS GNU/Linux" lists each installed kernel with a normal and a recovery-mode entry. Boot the last kernel that worked. From a recovery-mode root shell, NVIDIA gives these for an interrupted upgrade:

dpkg --configure -a
apt update
apt -f install
apt dist-upgrade

and these for a missing or half-built initramfs:

dpkg --configure -a
update-initramfs -u -k all
update-grub

A third scenario on the same page reinstalls the signed NVIDIA driver packages for a kernel when Secure Boot blocks the modules. NVIDIA's warning there matters: do not install nvidia-dkms-* packages, because they are unsigned and Secure Boot will not load them. For the September 7.0.0-1019 panic, an NVIDIA moderator posted this on 17 September 2026 (thread 383450); the trailing hyphen asks apt to remove the DKMS package in the same transaction:

sudo apt update
sudo apt-get install linux-modules-nvidia-580-open-nvidia-hwe-24.04 nvidia-dkms-580-open-
sudo dpkg --audit
sudo apt-get check
test -s /boot/initrd.img-7.0.0-1019-nvidia && echo "7.0 initrd is present"

An owner confirmed it fixed his machine, after first reaching a shell by booting the older kernel from the GRUB command line. If no menu appears at all, the unit was probably set up with a zero GRUB timeout; the updates guide shows the drop-in file that brings the menu back, worth adding before you need it.

How do I copy my data off a Spark that will not boot?

Boot an Ubuntu ARM64 live USB, choose "Try Ubuntu", and mount the internal NVMe drive. NVIDIA's System Repair page names Ubuntu 26.04.1 LTS ARM64 as the current image; an x86 image will not boot on the Spark's Arm cores, and an owner notes the Spark has only USB-C ports, so bring a USB-C stick or an adapter. NVIDIA's commands:

lsblk
sudo mkdir -p /mnt/nvme_mount_dir
sudo mount /dev/nvme0n1p2 /mnt/nvme_mount_dir

User files are under /mnt/nvme_mount_dir/home/. The hard part is size. One owner in May 2026 had about 750 GB in his Hugging Face cache alone and said rsync of terabytes of small files "can easily take 6+ hours" (thread 370981). By my calculation, at line rate:

Copy targetRate assumed750 GB takes4 TB takes
NAS over 1 GbE0.125 GB/s (line rate)100 min8.9 h
NAS over the 10 GbE port1.25 GB/s (line rate)10 min53 min
USB-C NVMe enclosure2 GB/s (owner-measured best case)6.3 min33 min

Time = size / rate, for example 750 GB / 1.25 GB/s = 600 s. Line rates are theoretical; the 2 GB/s USB figure is an owner's best result with an ASM2464-based enclosure on a port that lsusb reports as 20 Gb/s (thread 349121), and the same owner saw it fall to 40 MB/s when the enclosure heated up. Many small files are slower than one large one.

The practical answer: do not back up the model cache at all, since it can be downloaded again, and copy only what you made. The backup guide covers schedules; the model storage guide covers keeping the cache on a NAS or external drive so a reflash does not touch it.

How do I create recovery media and reinstall DGX OS?

Download the recovery image, run NVIDIA's USB script, reset UEFI, boot the stick and start recovery. NVIDIA's download page needs no login. Its own DGX OS page says not to use the Enterprise Download Center's DGX OS ISO for a Spark, which is where one owner looking for an ISO got stuck in August 2026 (thread 379524). From the extracted archive, run one of these with administrator rights; each erases the stick:

# Linux
CreateUSBKey.sh
# macOS
CreateUSBKeyMacOS.sh
# Windows, from an administrator prompt
CreateUSBKey.cmd

Then, per the recovery steps: unplug other USB storage, use a wired keyboard in a Spark USB port (NVIDIA says some Bluetooth keyboards fail here), hold Esc or Del at power-on, go to Save & Exit, Restore Defaults, Save Changes and Reset, enter UEFI again, and pick the stick under Boot Override. If the Spark boots into the old OS anyway, disable Secure Boot, retry, and let recovery re-enable it. On the recovery screen choose START RECOVERY, watch the progress output, and confirm the factory-state message at the end. The unit then boots into first-time setup, including the hotspot for a headless setup; one owner used exactly that to set a Spark up again without a monitor (thread 348457).

What state are UEFI and firmware in afterwards?

Factory defaults, which means your own UEFI changes are gone and your firmware may be older than it was. NVIDIA says recovery restores "the operating system and firmware to their original factory state", and the Restore Defaults step loads optimized defaults before that. Anything you set yourself (display reserve, PXE, Wi-Fi or Bluetooth off, from the UEFI settings page) needs setting again. Run the dashboard or fwupdmgr updates after setup to get back to current firmware; the current Founders Edition versions in NVIDIA's release notes are UEFI 1.111.17, EC 3.13.7 and SoC 2.158.15. A UEFI password is the one thing recovery cannot get you past, because the prompt appears at power-on, before Boot Override. The owner in thread 386027 got no further and reported that NVIDIA would not replace the unit once he said he had set the password himself.

The 1.155.49 image was released on 21 October 2026, the day the October 2026 release began rolling out, so it should also be the "October 2026 Release or later" image that NVIDIA says cloud-init provisioning needs.

What is different on ASUS, Dell and other GB10 machines?

Use the maker's image; NVIDIA's recovery note says OEM images "are only available from the respective OEMs" and recovery procedures may differ. Dell's knowledge base article (version 6, updated 22 October 2026, KB 000382042) has you download an ISO from Dell's site by Service Tag, check its MD5, write it with dd, press F7 at the Dell logo, and choose "DGX Spark Installation Options" in GRUB; installation takes 25 to 30 minutes. If it keeps failing, Dell's fix is the BIOS Media Sanitization tool on the drive. For ASUS, an owner linked an ASUS image named "NVIDIA DGX OS for ASUS Ascent GX10 7.4.0-3" in March 2026, and another said his eight GX10s run fine on NVIDIA's recovery image (thread 363185). That second route is unsupported, and it may change who supplies your firmware updates; I would use the maker's image on any machine still under warranty. The GB10 machines comparison lists all eight makers; expect their images and firmware to trail NVIDIA's releases.

When should I contact support or start an RMA?

When Field Diagnostic fails. NVIDIA's support page calls it a hardware health check and "a pre-check for RMA qualification". It installs from apt, needs Secure Boot off while it runs, takes about 30 minutes and ends with a PASS or FAIL banner:

sudo apt-get update
sudo apt-get install dgx-spark-fieldiag
sudo init 3
cd /opt/nvidia/dgx-spark-fieldiag
sudo ./partnerdiag --field

The ConnectX-7 stress test needs DOCA OFED and NVIDIA's firmware tools installed first. With a FAIL, open a hardware ticket through NVIDIA's support portal for a Founders Edition, or with the maker for a partner unit. NVIDIA's staff asked for exactly this output in the August 2026 thread about fans that stopped spinning (thread 378945). A boot loop, a missing initrd or a broken package is not an RMA case. A dead USB controller, a drive that will not reflash, or shutdowns that persist after a factory recovery are the cases I would take to support.

What about a rented Spark?

You never reinstall DGX OS on a rental, because you never touch the host. Per the tenant environment page, you get root inside a container; a rebuild reinstalls the container image and keeps /workspace, and terminate deletes /workspace for good. The operator keeps no backups, so the copy-your-data advice above still applies to /workspace; the storage and persistence guide lists what survives each action.

My order on an owned Spark: GRUB and the previous kernel, then a live USB and chroot, then copy your work off, and only then the recovery stick. Keep a recovery stick made from 1.155.49 next to the machine; at 5.9 GB it fits on the 16 GB stick NVIDIA asks for with room to spare.

FAQ

How do I reinstall DGX OS on a DGX Spark?

On a Founders Edition, download NVIDIA's DGX Spark System Recovery Image (version 1.155.49, released 21 October 2026, 5.9 GB), run CreateUSBKey on a 16 GB or larger USB stick, restore UEFI defaults, boot the stick with Boot Override and select START RECOVERY. It erases the internal SSD and returns the OS and firmware to factory state, so back up first.

Does DGX Spark recovery delete my data?

System Recovery does: NVIDIA's guide says it erases the internal SSD. System Repair does not. Booting a previous kernel from the GRUB menu, using recovery mode, or booting an Ubuntu ARM64 live USB to repair packages or copy files leaves the SSD intact, and NVIDIA says that procedure applies to Founders Edition and OEM systems.

Is there a DGX OS ISO for the DGX Spark?

Not for the Founders Edition. NVIDIA's DGX OS page says not to use the Enterprise Download Center or the DGX OS ISO, and the recovery guide says the Founders Edition OS ships as a tar.gz recovery image. OEM machines are different: Dell publishes an ISO per Service Tag for the Pro Max with GB10, and ASUS publishes its own GX10 image.

My DGX Spark shows a kernel panic after an update. Do I have to reflash?

Usually not. Owners in May and September 2026 traced 'Unable to mount root fs' panics to a new kernel installed without its initrd. Booting the older kernel, or a live USB and chroot, then completing the package install fixed it. NVIDIA posted a package command for the September 7.0.0-1019 case on its forum on 17 September 2026.

When should I RMA my DGX Spark?

When the hardware fails a test, not when the OS will not boot. NVIDIA's support page describes Field Diagnostic, the dgx-spark-fieldiag package, as a roughly 30-minute check of CPU, SSD, memory and ConnectX-7 and a pre-check for RMA qualification. Run it, then open a ticket with NVIDIA, or with your OEM for a partner machine.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

Buy one that is tested before it ships.

Every GB10 machine in the GPUwerk store is checked before it leaves Czechia, and you can have it hosted instead of keeping the recovery stick in a drawer.

GB10 machines shipped by DHL Express from Czechia, or hosted in GPUwerk's EU datacenter and rented out for you.

Browse machines Plan the backup first