NVIDIA DGX Spark — hardware support, downloads and testing

[jdvmi00]

I've been working on Omarchy support for the NVIDIA DGX Spark, building on the
shared ARM/Dragon installer and runtime work. This thread tracks the Spark
results, preview downloads and remaining issues.

Preview ISO

The September 23 preview installed successfully on my DGX Spark alongside DGX OS
and booted to the desktop. This is a community build, not an official Omarchy release.

• Download ISO — 8.48 GB
• Download SHA256SUMS

Download both files into the same directory and verify before writing the ISO:

sha256sum -c SHA256SUMS

The image uses the shared installer flow, the current Omarchy runtime, kernel
7.2.6-1.5, stock systemd 261.3-1 and NVIDIA 615.71.09 with the pending
DisplayPort detach fix. It includes CUDA, Ollama, Nsight Compute/Systems, AI
Workbench, DGX Dashboard and RDMA tools. When installing alongside DGX OS, the
installer uses free space and creates its own EFI partition. Check the target
carefully before installing.

Hardware results

These results apply to my DGX Spark and do not establish support for other GB10
machines. After installing the September 23 ISO I checked USB-C DisplayPort, CUDA,
GPU containers and RDMA link state; the other rows come from the September 21 ISO.

Feature	Result
USB-C DisplayPort	Desktop at 6K, 120 Hz. With the NVIDIA fix, reconnecting in both cable orientations, input switching and display off/on recovered the picture.
HDMI	Accelerated desktop at 6K, 120 Hz with a suitable HDMI cable.
Ethernet	Stable during the test at 2.5 Gb/s.
Audio and Bluetooth	HDMI playback and AirPods pairing, reconnection and audio passed.
GPU compute	Native CUDA, GPU containers and Ollama GPU inference passed.
Nsight	Both profilers captured a CUDA workload and opened reports.
Workbench and Dashboard	GPU notebook workflows and service restarts passed, using temporary compatibility integrations.
ConnectX/RDMA	Single-Spark cable-loop tests using both internal PCIe paths achieved 183.52 Gb/s RDMA payload left-to-right and 183.48 Gb/s right-to-left, with approximately 194.6 Gb/s physical-port traffic each way. Two-machine and GPU communication tests remain pending.

Remaining work

• Sleep/wake failed; further investigation is deferred.
• The first system update and snapshot creation worked on this image; restoring from a snapshot has not yet been tested.
• AI Workbench and DGX Dashboard work in our tests using compatibility changes for Omarchy. NVIDIA does not officially support this configuration.

Upstream work

The Spark package selections for RDMA tools, Nsight, Workbench and Dashboard are
merged in omarchy-iso. The preview carries local builds while these changes go
through review:

• Arch Linux ARM: audio, RDMA and Bluetooth.
• NVIDIA: DisplayPort detach fix, carried for aarch64 in omarchy-pkgs#601.
• Omarchy packages: CUDA/Ollama, ARM boot settings, GPU-container setup, Workbench, Dashboard and perftest.
• Omarchy runtime and installer: recovery packages, NVIDIA app setup and boot entry preservation.
• A black screen for about 20 seconds at every boot, where the firmware appears to overwrite all the memory Limine hands back before it starts a UKI. Reported to Limine and to NVIDIA, with no reply from either by 11 October. omarchy #13426, merged into n1x on 11 October, boots the Spark without a UKI until one of them fixes it.

The separate PKCS#8 kernel fix
for private-key handling is not included in this image. Enterprise Wi-Fi
authentication remains untested. Local test results do not mean the fixes are
available through normal updates.

If you try it, please include your exact machine, firmware version, ISO date,
kernel version and the steps that worked or failed. Distinguish live-USB results
from an installed system, and include relevant error messages with private
details removed.

[zytek]

Hey. Just used the linked ISO to install Omarchy on my Asus GX10. So far so good, very happy about it.

• smooth installation from USB-C drive onto unallocated space on main disk alongside stock ubuntu/dgx os
• boots fine, display works via thunderbolt
• wifi works, bluetooth works

Will test 10Gbe LAN + local inference later this week.

[jdvmi00]

Awesome, glad to hear. I've been running Omarchy on my two sparks and it's working great. Let me know if you find anything. There are some small changes we've made that will go into the final ISO, but this image should be stable for you. Enjoy, I've been loving it so much more over the DGX OS.

[zytek]

Thanks for this work, I moved back to Linux from macos after >10 years thanks to this.

I added the UKI=0 to make boot faster (plus disabling serial console) and am booting in under 30 seconds (as neither sleep nor hibernation seem to work, not a big deal for a minipc though).

Local inference works almost as good as on dgx os, I moved my vllm setup with orith 1.5 model but while doing that I re-created the venv based on python 3.14 and not 3.12 as previously. My pi dev harness reports slower tokens/s so this is something I should tackle next, but it does work which is great.

Where do you guys hand out (discord?), is there any way I could help/contribute and also be up to speed with Omarchy ARM/spark development?

[zytek]

the slower inference seem to be gpu clock issue, pinned to 900MHz, power cycling did help

 ┌────────────────────────┬────────────────────┬───────────────────┐
 │ Metric                 │ Before power cycle │ After power cycle │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ BF16 matrix throughput │ 37.53 TFLOPS       │ 73.07 TFLOPS      │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ Memory-copy bandwidth  │ 228.09 GB/s        │ 236.24 GB/s       │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ Loaded SM clock        │ ~877–903 MHz       │ 2450–2502 MHz     │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ GPU utilization        │ ~96%               │ 96%               │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ Loaded power draw      │ ~20 W              │ ~100–102 W        │
 ├────────────────────────┼────────────────────┼───────────────────┤
 │ Temperature            │ ~45°C              │ 56–70°C           │
 └────────────────────────┴────────────────────┴───────────────────┘

[benwgarton]

I would love to replace the DGX OS and run Omarchy on the DGX Spark 2 node cluster. Keep me posted if I can be of assistance or you have a build to test.

[YF36]

Two closed cases from a DGX Spark (GB10) on Omarchy 4.5 ARM64 — evidence below in case it saves someone the same multi-day hunt.

Random hard resets → fixed by linux-omarchy 7.3 (mtk_pwrap, omarchy-pkgs #1191). Ours reset with zero kernel output — even with ignore_loglevel, per-second journald and initcall tracing — and empty pstore. On 7.3-2, verified at four levels: 30 min bare idle, 30 min with screensaver cycles, the virtual-cursor trigger that killed ≤7.2.8-21 machines in one shot, and physical mouse. Post-fix log signature: mtk_pwrap NVDA6210:00: Not find device config data. If your Spark resets silently, upgrade to 7.3 before debugging anything else.

"Random multi-minute boot hang" → invisible LUKS prompt. On LUKS-encrypted installs with an NVIDIA-only display, the omarchy nvidia mkinitcpio drop-in strips the kms hook, so plymouth has no graphics device and the passphrase prompt is never drawn. The screen just freezes while the machine waits — it looks exactly like an early-kernel stall. Our five "stalls" (3–70 min) all matched human arrival time; initcall_debug put the gap precisely between NVMe probe and BTRFS scan of /dev/mapper/root (22.7 s → 150.5 s). Pressing Enter makes the prompt appear. Fixes: keyfile auto-unlock (our choice — reboot is now zero-touch, SSH back in ~50 s) or dropping plymouth so the prompt renders.