DGX Spark software updates: every release from launch to October 2026, known issues and how to update

The newest DGX Spark software is the October 2026 release, which began rolling out on 21 October 2026: it moves the kernel from 6.17 to Linux 7.0 and adds Dynamic Display Reserved Memory, DGX Spark Skills, native cloud-init and one software image for both 64 GB and 128 GB machines. NVIDIA's release notes list the current Founders Edition stack as DGX OS 7.6.0, driver 580.178.04, CUDA 13.0.3 and kernel 7.0.0-1019 (read 7 November 2026). Owners on NVIDIA's forum reported two problems around it: 7.2 GiB less RAM after the September kernel change, and a reboot loop on kernel 7.0.0-1024 from 30 October.

Read on 7 November 2026. NVIDIA says these versions apply to the DGX Spark Founders Edition only. The release notes give versions only for the current stack, not per release; the per-release versions below come from NVIDIA's DGX OS 7 release notes and from owners' forum posts, and each row says which.

What is the latest DGX Spark software version?

DGX OS 7.6.0 with the October 2026 release on top. NVIDIA ships DGX Spark software as dated releases (November 2025, January 2026 and so on) delivered over the air through the DGX Dashboard, and separately publishes DGX OS ISO images. The two numbering schemes do not line up neatly: the DGX OS 7 release notes date the 7.6.0 ISO to 14 September 2026 with kernel 7.0.0-1018-nvidia and driver 580.173.02, while the Spark release notes now list kernel 7.0.0-1019 and driver 580.178.04 as current.

Partner machines trail. The October announcement says partner GB10 systems get the same updates "in their next software release". In the forum thread "How safe is it to update right now?", an ASUS Ascent GX10 owner listed on 24 October the firmware his machine had against what the Founders Edition had: SoC 2.155.14 against 2.158.15 and EC 3.5.11 against 3.13.7. Another owner wrote that GX10 firmware arrives about three months after NVIDIA's. If you own a partner GB10 machine, read your maker's notes rather than NVIDIA's.

What changed in each DGX Spark release?

Ten entries in just over a year, most of them small. Feature lists are from the DGX Spark release notes unless marked; versions are from the DGX OS 7 notes or, where marked, from an owner's post.

Read 7 November 2026. "Owner's post" rows are versions reported by Spark owners on NVIDIA's forum, not NVIDIA statements. The September row is my reading of two sources: the DGX OS 7 ISO table and an owner's report that one over-the-air update took his machine from DGX OS 7.2.3 and kernel 6.11 to 7.5.0 and kernel 7.0.0-1019. On 23 October another owner called this "the early accidental kernel 7 push".

Two of these change what fits in memory. The July release made the display reserve adjustable, and the October release made it shrink to 512 MB when no display is attached; one owner reported about 1.5 GB more for CUDA on a headless machine afterwards. That matters most on the new 64 GB model (see 64 GB vs 128 GB). The January ConnectX-7 hot-plug is the change behind the lower idle draw in the power guide, and the June Cluster Assistant limits are the starting point for clusters beyond two Sparks.

What went wrong after recent updates?

Four problems recur on NVIDIA's forum. None is on NVIDIA's known issues page (updated 4 November 2026), which lists only general items such as the "Memory-Usage: Not Supported" line in nvidia-smi.

7.2 GiB less RAM after the kernel 7.0 update

On 14 September 2026 an owner posted boot logs showing available memory falling from 125,269,464 K to 117,723,484 K, with the reserved region growing from 8.4 to 16.1 million K, after the update to kernel 7.0.0-1019 and UEFI 5.36_0ACUM027 (thread 383222). A second owner compared two machines and found 117.7 million K available on kernel 7.0 against 127.3 million on 6.17. aniculescu, who posts NVIDIA's release announcements, replied on 17 September that the new kernel "enabled Kexec HandOver (KHO) which increases reserved memory" and that mitigations were rolling out. Owners in the thread suggested booting with kho=off; that is their workaround, not NVIDIA's.

Reboot loop on kernel 7.0.0-1024

From 30 October 2026, owners of Founders Edition, Gigabyte, Dell and ASUS units reported reboots about one to five minutes after boot (thread 387397). One owner on DGX OS 7.6.0 and driver 580.178.04 traced it to the kernel package moving from 7.0.0-1019.19 to 7.0.0-1024.25; others logged Xid 13 GPU faults and kernel oopses on 1024, and an ASUS owner a hard reset whenever HDMI was connected. Reverting to 1019 worked for several of them. NVIDIA asked for nvidia-bug-report logs the same day; by 7 November the thread had no fix or bug ID.

Dashboard stuck on the July update

From 15 July, owners reported the dashboard offering the July update again after each reboot (thread 376981). Two owners cleared it in August by installing the nvidia-firmware-580-580.159.03 package by hand. NVIDIA said on 14 August that engineering was working on it; an ASUS owner on 7.6.0 still saw the stale offer on 4 October.

GPU clock capped after the March update

On 20 March 2026 an owner on kernel 6.17.0-1014 and driver 580.142 reported the GPU running at 750 MHz instead of about 2,400 MHz at 96% load, roughly halving his training speed (thread 364166). Unplugging the power supply at the wall for 30 seconds to 5 minutes fixed it for him and another owner; a third blamed a USB Power Delivery firmware change.

How do I update a DGX Spark?

Use the DGX Dashboard, or five commands if it is unavailable. NVIDIA's OS and Component Update Guide calls the dashboard the "primary and recommended" method and gives this manual sequence:

sudo apt update
sudo apt dist-upgrade
sudo fwupdmgr refresh
sudo fwupdmgr upgrade
sudo reboot

The guide asks for stable power, closed applications, a recovery plan and a maintenance window, and says it covers the Founders Edition only. It describes no rollback; it points to the System Repair and System Recovery pages instead. Before you start, write down what you have now, so you can tell afterwards whether anything changed:

uname -r                       # running kernel, e.g. 7.0.0-1019-nvidia
nvidia-smi | head -n 4         # driver and CUDA version
fwupdmgr get-devices           # UEFI, EC, SoC and USB PD firmware
sudo dmesg | grep -m1 "Memory:"  # memory available at boot
dpkg -l | grep 7.0.0-1024      # is the kernel from the reboot reports installed?

The dmesg line is the one the owner in thread 383222 used to spot the 7.2 GiB loss; compare it before and after. Back up anything you cannot recreate first, as in the backup guide, because a failed boot can mean a factory recovery.

What should I check after an update?

That the machine boots twice cleanly, reports the versions you expected, and serves your model at the same speed. Concretely:

• Versions: rerun the commands above and compare with the release notes table at the top of this page.
• Memory: the boot Memory: line and free -g should be within a few hundred MB of before, or better on a headless machine after the October release. A drop of several GiB is the KHO pattern.
• Stability: leave it running for at least 15 minutes, with and without a monitor if you use one; the 1024 reboots came within five minutes, and other owners reported that sudo poweroff rebooted instead.
• Speed: run one known prompt through your engine and compare tok/s with your notes, which would have caught the March clock cap. The benchmark page has reference figures.
• Cluster: on linked Sparks, check that both cabled interfaces show (Up) in ibdev2netdev and rerun an NCCL test, as in the two-node guide. The February release existed to fix a multi-Spark regression.

If the kernel is the problem, the verified workaround in thread 387397 is to boot the previous kernel from GRUB ("Advanced options for DGX OS"), then hold the packages so the next update does not reinstall it:

sudo apt-mark hold linux-nvidia-hwe-24.04 \
  linux-image-nvidia-hwe-24.04 linux-headers-nvidia-hwe-24.04 \
  linux-tools-nvidia-hwe-24.04 \
  linux-modules-nvidia-580-open-nvidia-hwe-24.04 \
  linux-modules-nvidia-fs-nvidia-hwe-24.04

The same owner found the GRUB menu hidden by /etc/default/grub.d/no-grubmenu.cfg and added a file named zz-timeout.cfg in that directory with GRUB_TIMEOUT=10 and GRUB_TIMEOUT_STYLE=menu to get it back. Remember to apt-mark unhold once NVIDIA ships a fixed kernel.

Should I update now or wait?

If the Spark is a server that has to stay up, wait a week after each release and read the forum first. NVIDIA's announcer wrote on 23 October that "all update issues have been addressed"; the reboot thread opened seven days later. On the same forum, an owner who runs Sparks around the clock as LLM servers spent two days repairing one after a firmware update and summed it up as "If it's not broken, don't fix it." I would follow a middle path: install firmware and kernel updates in a planned window when the release adds something you need (the headless memory gain, a cluster fix, a security patch), keep the kernel packages on hold the rest of the time, and keep a recovery USB ready. A Spark that runs unattended, as in running a Spark 24/7, should never take updates automatically.

What does a rented Spark get?

On a rented Spark you do not run these updates yourself. The host layer (DGX OS, the kernel, the NVIDIA driver and firmware) is the operator's job, and you get root inside a container whose CUDA userspace you pin and upgrade yourself; the driver and firmware guide draws that line, and the tenant environment page explains what survives a rebuild. If you are deciding whether to buy, an hour on a GPUwerk Spark at $0.79 lets you test your stack without owning the update risk (how renting works).

FAQ

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.

• DGX Spark Release Notes, NVIDIA
• DGX OS 7 Release Notes, NVIDIA, 2026-10-27
• DGX Spark Software Updates - October 2026 Release, NVIDIA Developer Forums, 2026-10-21
• DGX Spark Software Updates - July 2026 Release, NVIDIA Developer Forums, 2026-07-14
• Improved unified memory handling with driver 580.159.03, NVIDIA Developer Forums, 2026-06-11
• DGX Spark Software Updates 02/2026, NVIDIA Developer Forums, 2026-02-12
• DGX Spark Software Updates, NVIDIA Developer Forums, 2025-11-19
• OS and Component Update Guide, NVIDIA
• DGX Spark Known Issues, NVIDIA, 2026-11-04
• DGX OS 7.5.0 OTA (UEFI 5.36_0ACUM027): ~7.2 GiB less RAM available at boot, NVIDIA Developer Forums, 2026-09-14
• DGX Spark repeatedly rebooting after most recent update, NVIDIA Developer Forums, 2026-10-30
• How safe is it to update right now?, NVIDIA Developer Forums, 2026-10-23
• DGX Spark July 2026 software update issue, NVIDIA Developer Forums, 2026-07-15
• Latest Update (20Mar 2026) on Nvidia Spark FE caps GPU performance, NVIDIA Developer Forums, 2026-03-20