Ubuntu 26.04 LTS (Kernel: 6.17.0) ARM64 on DGX Spark, anyone?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Ubuntu 26.04 LTS (Resolute Raccoon) Daily Build works on DGX Spark.. kinda

Downloaded the resolute-desktop-arm64.iso into a USB drive, and restarted as Live Ubuntu installation without any issue. Installed on a new partition:

Installation Issues:

  1. No wired network during installation even if appears to be available don’t select it.
  2. Wifi is enabled, and works. Well, not 1st time. If it seems hangs just go back and retry. Also it worked with Wifi 5 but not Wifi 6.
  3. Don’t hang around the installation screen too long. It crashes and reboot, and you get to see this cool Matrix like full of 00s scrolling by two times and need to start over.
  4. The monitor resolution during installation is 800 x 600.
  • If you haven’t figured out how to partition your 4TB storage, this is a easy way to do it. Use it as a live Ubuntu instance, to run Disk utility.

After the installation:

  1. After restart, you will be the grub menu. The 1st one is the new Ubuntu 26.04. The 2nd one is the original DGX OS. Both says Ubuntu :(
  2. The resolution is 800x600 and wasn’t able to change it yet. This is bugging me. Any help???
  3. You will see a notification that there are some 300 updates, but when you try to update it will says only partial update is possible. Still push it thru.
  4. After update, reboot, and the screen resolution is still 800x600.
  5. No wired ethernet, but if you reboot and start on DGX OS (remember , it’s 2nd one in the grub menu), then reboot, again on DGX OS, you will see your wired network live again, then reboot in Ubuntu 26.04, and you see the network.
  6. No NVidia driver installed or upgraded automatically.

Install the NVIDIA 580.95.05 AARCH64

  1. Set to text mode
    sudo systemctl set-default multi-user.target
  2. Reboot
  3. Login
  4. Install prerequsit:
    sudo apt install xorg xorg-dev build-essential pkg-config libglvnd-dev
  5. install Unix Archive Driver 580.95.05 AARCH64 from Driver Details | NVIDIA
  6. Reboot, still no higher resolution but hey Look! Got GB10.
  7. Who need display and wired network, anyway. Just SSH into it from a $50 on eBay HP ThinkCenter or any dumb terminal.
  8. Try to install conda but could get the installation working.

As I am writing this line, I decide to do another (3rd) installation from scratch and see if I can make thing little different. For time being, if anyone wants to try, please post what your finding is.

Especially if you figured out the higher screen resolution than 800x600. I can’t believe we used to program in 40x24 char screen.

1 Like
2

See this thread for more information: Has anyone tried an alternative Linux distro?

You can start from the end of the thread :)

TL;DR: need r8127 kernel module for ethernet, need nvidia open source drivers. May need to set nvidia-drm.modeset=0 kernel parameter. Best experience if you compile nvidia-specific kernel from the repo.

3

RealTek has the driver for “10G Ethernet LINUX driver r8127 for kernel up to 6.15”, But not ready to jump in yet. WiFi works great and ran out of Ethernet ports on my switch anyway.

Oh, and the issue about rebooting out of nowhere was the watchdog in UEFI. Just turn it off in Advanced and Advanced setting and suddenly it’s super stable.

After installing NVIDIA driver, Docker, NVIDIA Container Toolkit, It certainly works. NVIDIA Sync with VSCode works too.

I will try to run some non-docker playbooks.

4

Fixed the graphics resolution:

Use the Exact BusID “PCI:1:0:0” in xorg.conf which you can find it in “lspci”:

  1. Open the xorg.conf file for editing:
    sudo nano /etc/X11/xorg.conf
  • Modify the Section "Device" to include the correct BusID:
    Section "Device"
    Identifier     "Device0"
    Driver         "nvidia"
    VendorName     "NVIDIA Corporation"
    # Use this specific address from lspci output
    BusID          "PCI:1:0:0"
EndSection
  1. Edit the GRUB bootloader configuration file:
    sudo nano /etc/default/grub

  • Find the line that starts with GRUB_CMDLINE_LINUX_DEFAULT.

  • Add nvidia-drm.modeset=1 to the parameters inside the quotes. It should look something like this:

    GRUB_CMDLINE_LINUX_DEFAULT="quiet splash nvidia-drm.modeset=1"

  • Save the file (Ctrl+O, Enter, Ctrl+X).

  • Update GRUB to apply the new boot settings:

    sudo update-grub

  1. Reboot the system:

    sudo reboot
5

This gives me a new env with Kernel 6.17.0-5 with NVIDIA driver 580.95.05. Now what I am missing is CUDA 13.0.x.

$ inxi -Fz
...
System:
  Kernel: 6.17.0-5-generic arch: aarch64 bits: 64
  Desktop: GNOME v: 49.0 Distro: Ubuntu 26.04 (Resolute Raccoon)
Machine:
  Type: Server System: NVIDIA product: NVIDIA_DGX_Spark v: A.7
    serial: <filter>
  Mobo: NVIDIA model: P4242 v: A04 serial: <filter> UEFI: American
    Megatrends LLC. v: 5.36_0ACUM018 date: 08/06/2025
CPU:
  Info: 20-core model: N/A bits: 64 type: MCP cache: L2: 25 MiB
  Speed: N/A min/max: 338:1378/2808:3900 cores: No per core speed data
    found.
Graphics:
  Device-1: NVIDIA GB20B [GB10] driver: nvidia v: 580.95.05
  Display: unspecified server: X.Org v: 24.1.8 with: Xwayland v: 24.1.8
    driver: X: loaded: nvidia gpu: nvidia,nvidia-nvswitch resolution:
    1: 2560x1440~60Hz 2: 3440x1440~60Hz
  API: EGL v: 1.5 drivers: nvidia,swrast
    platforms: gbm,x11,surfaceless,device
  API: OpenGL v: 4.6.0 compat-v: 4.5 vendor: nvidia mesa v: 580.95.05
    renderer: NVIDIA Tegra NVIDIA GB10/PCIe
  Info: Tools: api: eglinfo,glxinfo gpu: nvidia-settings,nvidia-smi
    x11: xdriinfo, xdpyinfo, xprop, xrandr
...
6

You can download CUDA from NVidia: CUDA Toolkit 13.0 Update 2 Downloads | NVIDIA Developer

If deb version doesn’t work for some reason (it should), you can always use runfile one.

8

My guess would be newer packages. Otherwise one could just compile 6.17 from NV-Kernels on 24.04.

9

CUDA Toolkit 13.0 Update for Ubuntu 24.04 installs it no problem on Ubuntu 26.04. But that’s about as far as it goes. So far none of playbooks I tried worked, a docker container like vLLM works didn’t. SGLang container worked with about 65 T/s which was 70 T/s in 24.04 . But that’s as far as I went. If anyone else willing to try it, please post your result.

One thing though, its USB image is as good as DGX Spark Live USB image we never got. I already used it twice to save me from some dumb mistakes.

10

You probably need to set up paths and some other dev tools. Lots of development tools are pre-installed on Spark. As for speeds, as I mentioned before, you won’t achieve the same GPU performance as DGX OS without compiling a custom kernel from NVidia repository.

I’m able to run everything I can run on DGX OS under Fedora 43 with custom kernel at the same or faster speeds (mostly improved model loading, inferencing/prefill is the same).

But I guess, for most people it’s not worth it, and I hope NVidia releases 6.17 kernel at some point.

11

FYI I updated this morning and looks like kernel is now 6.17

nvidia@ai1:~$ uname -a
Linux ai1 6.17.0-1008-nvidia #8-Ubuntu SMP PREEMPT_DYNAMIC Wed Jan 21 17:56:56 UTC 2026 aarch64 aarch64 aarch64 GNU/Linux

1 Like