Hello
Since starting gaming on Linux exclusively about a year ago, I had crashes with "ring gfx_0.0.0 timeout" every once in a while. After some updates it got better and I was able to game for weeks without any problem, then it started again. Through the year I was able to play from 30 minutes to multiple hours until it crashed (if it did), but now I wasn't able to play more than a few minutes if the game didn't crash in the menu already. I tried all the possible solutions like adding different parameters to the kernel cmdine, add variables to the game launch command, try different Proton versions, downgrading Kernel etc. but nothing really helped (downgrading mesa stopped my ssdm to even start so I can't tell if that would have helped).
I was on EndeavourOS and thought "maybe with all the tinkering I made it worse" so I reinstalled it from scratch, same problem. Changed to CachyOS, problem persists. Thought about changing to a completely different distro but read that they all can have this problem.
So I found this Github page that first didn't help but pushed me in the right direction. On EndeavourOS I already tested with CoreCtl but didn't have any luck. So I tried LACT this time and tried a few things (should also work with CoreCTL, I just wanted to test different things to rule out errors on my side).
I have a RX8600 XT which should go up to 2360MHz. I had de graphs open on my second monitor and started the game. The game crashed as soon as the GPU clock went above 1600Mhz. I already feared that my GPU was dying. So after the next reboot, I set the max GPU clock speed to 1600MHz, way under the normal limit but I was able to play for about 20 minutes without any crash. I stopped the game, changed the clock speed to 2000MHz and it cashed instantly.
So I looked at the GPU voltage and it stayed at 856mV all the time, way to low for that clock speed. The my Sapphire Nitro+ the factory default is 850mV minimum and 1150mV maximum (Screenshot from igorslab.de):
As I'm too stupid to find a way to change the voltage to a fixed value and I didn't want to play around with the offset because "the offset of what exactly?" I searched for articles with similar problems and found a solution in a forum thread about a completely unrelated problem: Instead of setting the Power Profile Mode to "Auto" or "3D Full Screen" (which would make sense and is used automatically when playing games), I changed it to "Compute". The voltage now goes up to 1150mV when playing and I didn't have a crash for multiple hours!
I still have the command line options from the git-repo active but I think only amdgpu.ppfeaturemask=0xf7fff is really necessary, maybe even amdgpu.ppfeaturemask=0xffffffff would work. I'm gonna test this another day, now I'm just happy that it works.
Hope it helps someone else out there! ๐
TL;DR:
- Enable overdrive to manually change settings (see LACT documentation)
- Install LACT or CoreCtrl (and enable the daemon)
- Change Power Profile Mode to COMPUTE
Update: Shortly after posting here I had a crash again, because of course this happens after writing about a solution...
I then saw that the maximum clock speed was at 2500MHz while stock is 2360MHz. So I changed that. My PC then crashed while on desktop but in the now frozen graph I saw that the target clock speed still was at 2500. I rebooted and also changed the kernel cmdlime back so that the only addition is amdgpu.ppfeaturemask=0xffffffff. Stable again since then.
Oh no, how did we use it for decades and were even able to talk to Google Chat users.