Requesting help deciding to switch from a ROG Z13 to a Framework

Framework Desktop Linux
1

Hi all,

I have an ASUS ROG Flow Z13 (Strix Halo / AI Max+ 395). My TB4 eGPU works perfectly on Windows but is completely broken on Linux — amdgpu can’t initialize the GPU through the Thunderbolt tunnel. AMD’s Mario Limonciello is aware of the bug, but it doesn’t look like a quick fix.

I’m thinking about switching to a Framework system to escape this. Two honest questions before I spend the money:

1. Are Framework AMD users hitting the same TB4 eGPU issue on Linux?

I’ve already seen bug 221319 on kernel bugzilla where a Framework 16 owner has the same family of issues, and issue 4047 where a Framework 13 hangs at boot with an AMD eGPU attached. So it sounds like at least some of you are dealing with this too.

2. If you switched to OCuLink, did it actually fix it?

Specifically for AMD-host + AMD-GPU + Linux setups. I keep seeing OCuLink recommended, but I want to verify it works for the specific combo I’d be running, not just NVIDIA or Intel hosts.

If OCuLink reliably solves this on Framework hardware, I’m leaning toward grabbing a Framework Desktop and putting my 7900 XTX or buying a 5090 to drop in. If you’ve made that switch, I’d love to hear how it went.

Thanks.

2

There are similar bug reports for the FW16 AMD. E.g. the DEG2 over usb4 / thunderbolt works on windows, but not on linux.
All the AMD pcie low level programming docs are under NDA, so impossible to make progress on by you or me. Progress is only possible if AMD spend the time on the problem.

In short, it appears that usb4 / thunderbolt pcie link training and redriver control works differently in linux than windows.

Regarding usb4 / thunderbolt vs oculink.
Oculink has lower latency. So usb4 vs oculink 4i apparently sees noticable improvements for games. Going to oculink 8i vs 4i is apparently hardly noticable.

3

With my DEG2 (that I’ll be returning) I cannot get Occulink or USB4 to work for my 5060ti

That being said, I an using all-ways-egpu with my UT3G and performance has been an uplift from the dGPU.

I have an occulink adapter still and will look to try a different occulink dock eventually. But for now, the UTEG over USB4 works well.

4

Thank you for the replies, they are much appreciated. I’m generally doing AI work so it doesn’t really both.

5

Hey that’s interesting are you running it on Fedora or Windows?

6

Fedora - I have a windows partition soley for testing - but anything I would consider “working” for my use case - would have to work on Fedora.

7

I would just say as Windows works properly it’s certainly something the kernel is doing wrong with some of these eGPU devices. It’s a needle in a haystack WHAT though without comparison of PCIe traces from Windows to Linux with exact same hardware setup.

Right now the only two things I know for sure are different:

  • Windows enables ASPM, Linux doesn’t.
  • Linux enables more AER than Windows does.

But changing both of these has not improved the situation.