Raspberry Pi support (arm64)?

[geerlingguy]

I've been trying to get an RTX 3080 Ti working on my Raspberry Pi 5 — see issue here: geerlingguy/raspberry-pi-pcie-devices#549

I've been able to get the amdgpu kernel driver to work with full desktop environment and hardware acceleration support, though we still run into occasional bugs that we're squashing.

The Nvidia drivers seem to run better on supported arm64 platforms, like Ampere Altra and AmpereOne, and many have reported success on other arm64 platforms (like Honeycomb or NXP dev boards).

Judging by my latest attempt with the newest open kernel driver (installed with sudo ./NVIDIA-Linux-aarch64-550.127.05.run -m=kernel-open), there may be a few memory mapping-related errors, in my testing on a Pi 5 with 8 GB of RAM:

[    6.496707] NVRM: clCheckUpstreamLtrSupport_IMPL: PCIE config space is inaccessible!
[    6.595440] NVRM: kgspExecuteFwsec_TU102: failed to execute FWSEC for FRTS: no initialized WPR2 found
[    6.595447] NVRM: kgspExecuteFwsec_TU102: (note: VBIOS version 94.02.71.80.74)
[    6.595451] NVRM: nvAssertOkFailedNoLog: Assertion failed: Failure: Generic Error [NV_ERR_GENERIC] (0x0000FFFF) returned from status @ kernel_gsp_tu102.c:445
[    6.595463] NVRM: RmInitAdapter: Cannot initialize GSP firmware RM
[    6.596427] NVRM: GPU 0000:01:00.0: RmInitAdapter failed! (0x62:0xffff:1792)
[    6.596676] NVRM: GPU 0000:01:00.0: rm_init_adapter failed, device minor number 0

It would be great if I could get any level of working support for the card (or newer generation cards, 40 series or workstation cards) on the Pi 5.

Here is a picture showing an overview of my hardware setup for testing (in this case, with an A400 plugged in):

[geerlingguy]

Same error happens with an A400, A4000, and 4090 on the Pi, also tested with the latest driver (both proprietary and open), version 565.77. See: geerlingguy/raspberry-pi-pcie-devices#677

[aritger]

Hi @geerlingguy . I'm sorry I missed your earlier posts here. I've filed NVIDIA bug 5053788 internally for this.
Can you please retest with the latest release (565.77, as of this writing), and then generate and attach an nvidia-bug-report.log.gz? Thanks.

[geerlingguy]

Will do! For my own reference, looks like I should su to root and then run nvidia-bug-report.sh.

[geerlingguy]

I have run the following to install the 565.77 driver, on a brand new / clean install of Pi OS 12/Bookworm:

sudo apt update
sudo apt install -y raspberrypi-kernel-headers
sudo apt dist-upgrade -y
sudo reboot

...

cd Downloads
wget https://us.download.nvidia.com/XFree86/aarch64/565.77/NVIDIA-Linux-aarch64-565.77.run
chmod +x NVIDIA-Linux-aarch64-565.77.run
sudo ./NVIDIA-Linux-aarch64-565.77.run

I installed the MIT/GPL version, and rebooted...

After the reboot, I got the same errors in dmesg:

[   16.186734] NVRM: _issueRpcAndWait: rpcRecvPoll timedout for fn 47!
[   16.186736] NVRM: kgspInitRm_IMPL: SET_GUEST_SYSTEM_INFO failed: 0x65
[   16.186828] NVRM: RmInitAdapter: Cannot initialize GSP firmware RM
[   16.188354] NVRM: GPU 0000:01:00.0: RmInitAdapter failed! (0x62:0x65:1863)
[   16.188740] NVRM: GPU 0000:01:00.0: rm_init_adapter failed, device minor number 0

I ran nvidia-bug-report.sh, and the vulkaninfo process seemed to use 100% CPU for a while, while many messages went by in dmesg like the following:

[  120.167666] NVRM: gpuHandleSanityCheckRegReadError_GM107: Possible bad register read: addr: 0x110100,  regvalue: 0xbadf5720,  error code: Unknown SYS_PRI_ERROR_CODE

The full nvidia-bug-report.log.gz is attached below:

nvidia-bug-report.log.gz

[zackify]

Dang jeff, i thought this already worked so I bought the parts and I am receiving them tomorrow lol! My fault for reading your amd article and assuming it would.

I am installing the latest version now, 575.64, and let you all know if it happens to work yet, using a 2070 super

[aritger]

I could also use a description of your Raspberry Pi hardware configuration: what is the combination of hardware modules that you're using to connect a GPU to the Raspberry Pi?

[geerlingguy]

For most of my testing, this is the hardware setup (pictured above):

• Raspberry Pi 5
• Pineboards HatDrive! Bottom
• JMT Oculink to PCIe Dock
• chenyang Oculink cable
• Lian-Li 750W SFX Power supply, with 24p header connected to JMT dock

For quick testing with smaller cards, I use a smaller HAT that puts a PCIe slot directly on top of the Pi:

• Raspberry Pi 5 model B (8 GB, and also testing 16 GB now that it is out)
• Pineboards HAT uPCIty Lite
• Lian-Li 750W SFX Power supply, connected to the card and the uPCIty Lite board via 4-pin CPU header (for slot power)

[mickkael]

@aritger could you share if there is any update from the internal ticket ?

[aritger]

I'm sorry, I'm not aware of any progress so far.

[jstockdale]

Any luck with progress here @aritger? I also would be excited to run NVIDIA hardware on the rapsberry pi 5 / cm5 platform. Thanks!

[aritger]

I'm sorry, there hasn't been any interesting progress to report, yet.

[aritger]

I'm sorry; I don't have any news to report. Though, it sounds like @RSC-Games is making great progress.

[RSC-Games]

I haven't gotten anywhere on Nvidia so far due to a pretty stressful school workload- I think @yanghaku is making the most progress for the RPi boards and like Jeff said @mariobalanica has made progress with the RK3588 boards. How well it'll work on those I can't say since it has the bus quirks of the pi 4 but the pi 5 should work after patching. I'll won't say it'll never work but it won't be ideal

[mariobalanica]

Just a heads-up that my repo isn't meant to be RK3588 specific. I can only test on boards with this particular SoC, but these exact same issues affect most common Arm SoCs, including the Pi. So if anybody wants to try this on a Pi, I've already posted nouveau patches in a branch there and I'm also preparing patches for NVIDIA's driver. The goal is to get them upstream and extensive testing on different platforms will certainly help.

How well it'll work on those I can't say since it has the bus quirks of the pi 4 but the pi 5 should work after patching. I'll won't say it'll never work but it won't be ideal

It works surprisingly well, I've posted some screenshots in that repo.

IIRC the Pi 4 was much, much worse and I doubt the open NVIDIA driver has a chance there. The Pi 5 has larger MMIO space than RK3588 - this will probably matter for Intel cards, but it's not a showstopper.

[RSC-Games]

I can start doing extensive testing with nouveau in a couple weeks but I have a pascal card so it won't reclock properly...

[mariobalanica]

See: https://github.com/mariobalanica/arm-pcie-gpu-patches/tree/nvidia-wip/open-gpu-kernel-modules

I could send a pull request if there's interest from NVIDIA, though judging by the other closed PRs, it appears most get picked up & applied manually. This patchset is quite extensive and also touches some auto-generated code, which from my understanding is built from a larger shared codebase and isn't meant to be modified directly.

[geerlingguy]

Just wanted to post an update from over in the Pi PCIe repo, @RSC-Games is exploring IOMMU support on the Pi, possibly helping both Intel and Nvidia drivers work better on the Pi 5's PCIe bus, but it's very experimental currently, definitely with no success guaranteed: geerlingguy/raspberry-pi-pcie-devices#754

[RSC-Games]

I have yet to do testing yet on amdgpu with forced cache flushes between the test ring read and write (with TTM set to cached) but if that works, and we can somehow get the IOMMU to map all memory as uncached, regardless of what any driver requests, then we're pretty much solid on almost all PCIe device drivers in the kernel, barring x64-specific code.

As Jeff said above, check the issue for more updates and a live view of what's going on with the testing and development of these patches. I'll update this comment every now and then (when/if we make any breakthroughs) so stay posted!

EDIT: Progress was made by someone outside of my existing efforts- see the below comment.

[RSC-Games]

Nvidia-open was tested, patched, and is currently working on the RPi 5. Success is not guaranteed and support is in the early stages, but it is working.

See geerlingguy/raspberry-pi-pcie-devices#756 for more information.

[aritger]

Are the pineboards still what you're using to connect an NVIDIA dGPU to the Raspberry Pi 5? I was going to pick up a hardware configuration to test in a different geography than where our current setup is, but it looks like pineboards.io is out of business? Are there other PCIe options for Raspberry Pi 5?

[pepijndevos]

Yea he links them here: https://www.jeffgeerling.com/blog/2025/using-amd-gpus-on-raspberry-pi-without-recompiling-linux

[l4z41]

I'm using a ComputeBlade with Raspberry Pi Compute Module 5 and M.2 Oculink OCuP4 where the Nvidia GPU is recognized with lspci but the driver fails and waiting for some resolution. Glad to share the setup!

[RSC-Games]

I used the JMT oculink dock from Amazon
It's like 50 bucks and it's pretty good. Then you'll just need an atx PSU and an oculink cable
Jeff uses the same one too

[geerlingguy]

Ditto what was written above — the most stable configuration for now seems to be either Compute Module 5 + IO Board + M.2 to Oculink, or Raspberry Pi 5 with an M.2 HAT with 2280 length slot, and then that to M.2 to Oculink.

I've successfully connected most cards using the JMT dock as mentioned above (https://www.amazon.com/dp/B0CRR7HMX5), which includes an M.2 to Oculink adapter. You will need a separate Oculink cable (I bought this chenyang: https://www.amazon.com/dp/B07TG9DK4W), and then for the Pi 5 specifically, this M.2 HAT seems to be reliable: https://www.amazon.com/GeeekPi-NVMe-Adapter-Raspberry-Support/dp/B0CRK4YB4C/ (though there are many other options).

Right now my main setup uses a Minisforum DEG1 (https://www.amazon.com/MINISFORUM-DEG1-Docking-Station-PCIE4-0x4/dp/B0DNJPGTHZ/) and this Micro SATA Cables adapter, which has a built-in redriver: https://www.amazon.com/dp/B09XY3F2GC

[pepijndevos]

the most stable configuration for now seems to be either Compute Module 5 + IO Board + M.2 to Oculink, or Raspberry Pi 5 with an M.2 HAT with 2280 length slot, and then that to M.2 to Oculink.

Have you done much testing with your sentinel core prototype? If there are any stability problems with it I'd love to know of course.

[mariobalanica]

I've opened a pull request for this: #972

It should work on the Pi as well.

[RSC-Games]

Heck yeah! I'll test with my rtx in a couple weeks once I get the whole pi set up back together again

[geerlingguy]

Just to provide an update in this issue, I've been testing an RTX 3080 Ti and RTX A4000, and both work great on the Raspberry Pi 5/CM5—just without any display output.

The use of the patch also required me switching to the kernel8.img kernel with 4K page size (instead of the Pi's default 16K kernel). So I'm going to post a separate discussion about adding 16K page size support, as @mariobalanica found the driver code only supports 4K, 64K, 128K, 2MB, 512MB, and 256GB.

Edit: Opened separate discussion here for 16K page size: #977

[RSC-Games]

Oddly specific page sizes. I see huge pages in there but weird they skipped straight from 4k to 64k. I have a sneaking suspicion these are the page sizes supported by modern x86

[geerlingguy]

I know I've seen 64K on many server platforms. 16K is used mainly by Apple, Android, and Raspberry Pi, in my experience. It seems less popular on x86 platforms, where they kinda skipped straight from 4K to 64K.

[KWAC-EE]

I can confirm that low-end Turing cards (T400, T1000, GTX1650) function for nvenc on the CM5 with these patches in Ubuntu 25.10 and RaspiOS 12-04 release, both with 4K page size instead of default, though I don't have the performance metrics. attempts to encode with mainline open modules are still met with RmInitAdapter failed in both 580 and 590 branches.

[dude365635]

Any progress on this guys? Testing a 2070 right now