Nvidia's RTX Spark laptops ship with Windows 11 on Arm and no official Linux support. Josephine Pfeiffer didn't want a Microsoft account and put Arch Linux ARM on an ASUS ProArt P16 with the N1X chip instead. CUDA sees 116.0 GiB of the 128 GB. A reviewer's Windows setup on the same model got 109.7 GiB.
If you want to install Linux on an RTX Spark laptop today, this is what I could piece together: which distro and kernel actually boot, the install steps, and a compatibility table pulled from the author's write-up, the build repo's docs and the Hacker News thread. I haven't run any of this myself, so every number here comes from the linked sources.
Why anyone is doing this
The write-up is Josephine Pfeiffer's Putting Linux on my RTX Spark laptop because I don't have a Microsoft account, posted 2026-10-19. The title tells you the motivation. The machine was a ProArt P16 H7607BA-SE033W with the N1X chip: 20 CPU cores, 6,144 CUDA cores, 128 GB of unified memory, a 2 TB SSD and a 3840 × 2400 120 Hz panel.
Officially there is nothing to install. ComputingForGeeks notes that Nvidia's spec sheet lists Windows 11 for every configuration, that Nvidia declined to comment on Linux when XDA asked, and that the public open kernel module 615.71.09 knows about GB10 (DGX Spark) but has no N1X device IDs.
So this is a weekend port by one maintainer, and the Hacker News thread (244 points, 121 comments) spent more time on the Microsoft account and the price of the laptop than on the port itself.
The distro and kernel that boot
| Piece | What the port uses |
|---|---|
| Distro | Arch Linux ARM, installed with Archinstall 4.5 |
| Kernel |
7.0.0-1022-nvidia (package linux-n1x-nvidia-bin 7.0.0.1022.23-3), built from Nvidia's NV-Kernels plus an ACPI override allocation patch |
| GPU driver |
nvidia-n1x-modules 615.78.08-4, nvidia-utils 615.78.08, open GPU kernel modules with a framebuffer patch |
| Bootloader | systemd-boot, mkinitcpio, an early ACPI archive built from your own machine's tables |
| Power |
power-profiles-daemon-n1x 0.30.nvidia3001c25-2, CPPC init (no asusd/asusctl) |
| Userspace | NetworkManager, PipeWire, rootless Podman with Nvidia CDI |
| Firmware | BIOS H7607BA.300 only |
The firmware row is the one that can bite you. The ACPI overrides match one firmware version, and the limits page says plainly: "do not update the BIOS after installing." The whole port has also been tested on exactly one machine.
Compatibility table: what works on Linux
This combines the blog post, the repo's hardware.md and limits.md. "Not documented" means the sources don't say either way, so don't read it as broken.
| Component | Status | Notes |
|---|---|---|
| GPU / CUDA | Works | 116.0 GiB reported as both system RAM and CUDA total |
| Internal display | Works | Tested at 3840x2400, 120 Hz, scale 2 |
| Keyboard + backlight | Works | Patched hid-asus, Fn-lock unchanged |
| Touchpad | Configured | libinput quirk for the PixArt pad |
| Speakers / headphone jack | Works, playback only | Needs three Cirrus CS35L56 tuning files from ASUS's driver package |
| Microphone | Excluded | No capture |
| Wi-Fi / Bluetooth | MT7925 firmware shipped | Installer tells you to set up Wi-Fi with nmtui, no explicit status |
| Suspend / resume | Disabled by policy | logind ignores lid and suspend-key events |
| Battery life on Linux | Not documented | Nothing in the post or repo |
| Camera, touchscreen, stylus, fingerprint, NPU | Excluded | Listed as out of scope |
For a machine whose whole pitch is local AI, the important rows hold up. CUDA works, and you get the full unified memory pool. A laptop you can't close the lid on is still a problem, though: the device policy has logind ignore lid and suspend-key events. I'd want that fixed before paying for one.
How to install it (the short version)
The repo is pfeifferj/archlinuxarm-proart-p16. Read the whole docs folder before you start.
1. Dump your ACPI tables
The overrides are built from your machine's own tables, so you can't skip this. Stage a FAT32 stick:
python3 01-platform/acpi/payload/scripts/stage-acpi-dump.py MOUNTPOINT
Boot the P16 from it with Secure Boot off. The EDK2 shell writes DSDT0000.bin, SSDT0000.bin and friends into H7607A-ACPI/. From any Linux live system you can copy /sys/firmware/acpi/tables/DSDT and the SSDT* files instead.
2. Build the overrides
Build ACPICA iasl 20260408 from the release tarball, then run:
python3 01-platform/acpi/payload/provenance/hardware/power-archive-builder.py \
--dump /mnt/stick/H7607A-ACPI --iasl /path/to/iasl
It refuses to continue unless the tables match the pinned H7607BA.300 hashes, which is how the BIOS lock gets enforced in code. Nothing gets flashed to firmware.
3. Assemble and build the installer images
Assembly needs Bash, jq and Python. The image build has to run on native AArch64 Linux as root. The documented USB writer runs from macOS with a Lima VM, so if you're on something else, expect to adapt it.
4. Install
Disable Secure Boot, plug in AC power, hold Esc at startup and pick the n1x p16 installer entry. Connect Ethernet or run nmtui for Wi-Fi, then:
archinstall
The installer checks the board name and BIOS version, then needs an ALARM mirror and NTP. If the clock doesn't sync within 90 seconds it stops before touching the disk. Use a FAT32 ESP at /boot (1 GiB recommended, 512 MiB minimum) and ext4 or Btrfs root. Plain, LUKS, LVM and LVM-on-LUKS all work. TPM/FIDO unlock, UKIs and Plymouth are rejected.
5. Live with the holds
The kernel, Nvidia, h7607a-hardware, power-profiles-daemon-n1x and the MediaTek/Cirrus firmware sit in IgnorePkg. Everything else updates from Arch Linux ARM as usual. The held packages get no updates, security fixes included, until a new build ships, and you're signing up for that when you install.
Is Linux slower than Windows on N1X?
From the numbers in the post, no. Pfeiffer compared 7-Zip 26.04 results against StorageReview's HP OmniBook Ultra 16 on Windows. Single-thread was within about 1% (-1.2% compression, -0.8% decompression). On 20 threads Linux came out ahead, +5.0% compression and +2.7% decompression. Those are different laptops, but ComputerBase's Windows score for a 64 GB ProArt P16 (155.774 GIPS) sits inside the author's own three-run range.
For local inference, Nvidia's Qwen3.8-Flash-Next-FP4-Local checkpoint on vLLM ran a code-review task at 30.9 tokens/s median. Frandroid measured 29.4 and 31.4 tokens/s on Windows with llama.cpp and a different quant. The author says that comparison doesn't isolate the OS, and I agree. The soak test is more convincing: 150 requests over ten minutes, all passing, and response times drifted less than 1.6%. Loading and compiling took about 16 minutes before the first request, which is worth knowing before you plan a demo.
Where Linux clearly wins is memory. Frandroid's Windows CUDA total was 109.7 GiB with a 91 GiB dedicated reservation. Linux gives you 116.0 GiB with no reservation to set up.
The skeptics have a point
The HN thread was harsh about the hardware. chem83 doubted "there's a strong use case for one of these $6K laptops yet". Melatonic said the chips are "severely hampered by memory bandwidth". Nvidia quotes 300 GB/s, and ComputingForGeeks, citing XDA, points out that GB10 ships at 273 GB/s. chem83's alternative was a mini PC you reach remotely from a lighter laptop, and wmedrano reported about 60 tokens/s on Strix Halo.
I mostly agree on price. Even so, if you want CUDA, 128 GB of unified memory and a battery in one bag, I don't know of anything besides the N1X laptops that sells that combination right now.
If you have to stay on Windows
Plenty of people will. One HN commenter pointed out that these machines ship with Windows 11 Home, so the Pro-only "Domain join instead" route for a local account won't help. Pureinfotech reports that Microsoft is removing OOBE\BYPASSNRO, starting with build 26120.3653 (Beta) and 26200.5516 (Dev), and that start ms-cxh:localonly stopped working in 26220.6772 and 26120.6772. The same article lists the registry and HideOnlineAccountScreens methods as still working, and Rufus or Ventoy can strip the requirement from install media.
So: buy now or wait?
My take: if you need CUDA on a laptop and can live without suspend, a mic and BIOS updates, this port is usable today and the memory numbers favor Linux. If you need any of those, wait for Nvidia to ship N1X device IDs in the open module, or for a second person to reproduce this on a different model.
Would you buy an ARM laptop for local AI while Linux support is still this half-baked, or wait a year? And if you already own an N1X machine, tell me in the comments whether you've got suspend working.
Sources
- Josephine Pfeiffer, Putting Linux on my RTX Spark laptop because I don't have a Microsoft account
- pfeifferj/archlinuxarm-proart-p16 (installing.md, limits.md, hardware.md, acpi-overrides.md)
- Hacker News discussion
- ComputingForGeeks: NVIDIA RTX Spark Laptops, N1X Specs, CUDA and Linux
- Pureinfotech: Windows 11 without Microsoft account
Top comments (0)