7900XTX eGPU much slower on Linux after upgrading to 2.8k laptop screen

I have a Radeon 7900XTX GPU in an eGPU enclosure connected to my 13" Framework (AMD 7840U). eGPUs are temperamental, but with newer kernels it has been working fairly consistently for AI applications. But this week I got the new 2.8k display and installed it, and much to my surprise it somehow seems to affect the eGPU’s performance a lot, despite there being no displays plugged into the eGPU.

I didn’t think to take exact benchmarks from before I swapped out the laptop display, but it is approximately 1/6 speed that it was before for ML inference applications like LLMs. Here is a screenshot of radeontop with the eGPU trying its best:

Notice the shader clock in particular. It used to be that this was always near 100% when the eGPU was in use. On the upside, the eGPU is now very quiet, because it is barely working. So that is nice! But less nice is that it is no longer useful for most things.

Any ideas? I really like the new display, and am confused about how it could interfere with my eGPU, but I guess it isn’t totally out of the realm of possibility since they are both graphics-related things.

I am using Fedora Silverblue, currently version 40.20241024.0, with kernel 6.11.4. I believe I can rule out coincidental software changes since I haven’t updated the OS in a week, and it was working fine with this exact version of Silverblue before the display upgrade.

BIOS info:
Version: 03.05
Release Date: 03/29/2024

1 Like

There are some dGPU performance patches that that aren’t in 6.11.y that could impact this. Try setting the “3d fullscreen” profile.

6.12-rc5 sets this by default.

1 Like

Worth checking this : https://forge.quad.moe/quad/win3-resources/src/branch/master/docs/eGPU.md#ensuring-your-amd-egpu-runs-at-its-peak-speed

1 Like

Thank you! That was either the problem, or led me to the problem, still not 100% sure. It led me to install CoreCtrl, which I used to switch to 3d fullscreen, but then I also had to change the memory clock. No idea why, but it seems like it was at 96MHz. I’m pretty lost on how that could be caused by installing a new laptop display, but at least it is fixed now!

1 Like

It’s a coincidence most likely. Maybe you had an unaffected kernel before and upgraded but hadn’t rebooted in a while or something.