NVIDIA graphics meta-issue

[ids1024]

Known issues with cosmic on Nvidia graphics can be linked here, to keep track of them all while having separate issues for otherwise unrelated problems.

• https://github.com/pop-os/cosmic-comp/issues/264
  • Known issue with wgpu on Nvidia wayland, also on Gnome. Suspected to be driver bug
    • https://github.com/gfx-rs/wgpu/issues/4775
• https://github.com/pop-os/cosmic-comp/issues/243
  • Gnome also seems variable about whether or not external Nvidia monitors work at start. (Notably, Nouveau works better)
    • But I have not verified if the issue is the same. Warrants more testing, and fix in Smithay if possible.
• https://github.com/pop-os/cosmic-comp/issues/211
  • Gnome has similar performance issues with Intel surfaces on an Nvidia output, but is more consistent.
  • Performance issues with other combinations may exist on 1000 series, but not 3000?
    • https://github.com/NVIDIA/egl-wayland/issues/72 mentions 550 driver fixes an issue with 1000 series and earlier. Should test again when that's available.
• Extremely/unusably slow performance on some Nvidia hardware?
  • I have not been able to reproduce this one. Needs more information.

[ids1024]

We should probably write up some organized documentation for testing graphics driver related issues, but some things to look at in testing:

• See if there are any errors from cosmic-comp (which will be written to both stderr and journalctl). As well as any dmesg errors from the Nvidia driver.
• For basic performance testing, look at fps of fullscreen applications with vsync enabled.
• Log the output of the process with WAYLAND_DEBUG=1 to see what Wayland calls the application is making and the compositor is responding with.
• switcherooctl lists env vars to try running an application on a particular GPU.
  • I'm not seeing the env vars it offers affecting Vulkan clients(?), but VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json forces the client to only use Nvidia cards.
• Note whether the application uses Vulkan or OpenGL, which GPU its running on, and which GPU the output it appears on is connected to.
• Comparing to other Wayland compositors is useful for determining if an issue is caused by a driver bug, or something we should be (be able to) fix. Comparing an X session is not so useful.

[RayJW]

• Extremely/unusably slow performance on some Nvidia hardware?

I have this issue on my setup. I wanted to try the new Cosmic Epoch on Pop!_OS 22.04 around two months ago, but it still doesn't work. It usually takes around 30–60 seconds to even load any output, and then it's just static. The clock seems to work I think, but the cursor is stuck to the upper left corner (0, 0 perhaps?). Sometimes I can get the launcher to pop with the super key but then with a huge delay of multiple seconds. If there is any information I can provide to diagnose this issue, I am happy to help, since it was mentioned that this couldn't be reproduced so far.

[ids1024]

After looking through https://github.com/NVIDIA/egl-wayland/issues/72, an NVIDIA engineer there mentions some users with Pascal cards seeing "extremely low framerates (0.2FPS)", and says that should be fixed in the 550 driver, when that's released.

That seems to be the same issue you're seeing, and has appeared on at least one system we've tested it on.

I'm kind of curious what hardware/circumstances that occurs on since it doesn't seem to be all 1000 series GPUs, but that should hopefully no longer be an issue when the 550 driver is released.

[RayJW]

It seems like this is probably exactly the issue I'm experiencing since I use Pascal. In that case, I'll report back if it still won't work once the 550 series driver is out. Out of curiosity, I saw it mentioned once but is EGL Wayland now installed by default on Pop!_OS because the testing repo has been discontinued but is it also installed automatically as a dependency now or do I have to take manual steps?

[mmstick]

@RayJW It is installed by default with cosmic's packaging

[luccahuguet]

I can reproduce https://github.com/pop-os/cosmic-comp/issues/264

and it does work only on integrated graphics. Discrete or hybrid does not.

my system is the following: OS: Pop!_OS 22.04 LTS x86_64 Host: 80YH Lenovo ideapad 320-15IKB Kernel: 6.6.6-76060606-generic Uptime: 12 mins Packages: 2489 (dpkg), 2 (nix-user), 48 (nix-default), 45 (flatpak), 11 (snap) Shell: nu 0.88.1 Resolution: 1920x1080 DE: COSMIC Theme: Adwaita [GTK3] Icons: Adwaita [GTK3] Terminal: zellij CPU: Intel i7-7500U (4) @ 3.500GHz GPU: NVIDIA GeForce 940MX GPU: Intel HD Graphics 620

[RayJW]

@RayJW It is installed by default with cosmic's packaging

Perfect, thanks for the heads-up!

[ryzendew]

only issue ive had so far on nvidia is WGPU which is fixed by forcing it to use GL but the performance is great

[ryzendew]

https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/967#note_2258144 Nvidia dev confirmed it's a driver bug

[LunNova]

One of my monitors get stuck on one frame while the other works using nvidia proprietary drivers.

nouveau + gsp + mesa nvk 24 seems to work really well!

[Drakulix]

One of my monitors get stuck on one frame while the other works using nvidia proprietary drivers.

I have seen nvidia outputs occasionally not initializing correctly after the first frame. Do you happen to have logs of that particular run? Is it reproducible on your system?

[LunNova]

Do you happen to have logs of that particular run?

No

Is it reproducible on your system?

Trying again today with nvidia proprietary, no. It was reproducible within one particular boot.

[tim77]

COSMIC Doesn't work at all on Nvidia 550.76-1 drivers. Doesn't launch.

• NVIDIA 3060
• nvidia-open driver
• OS: Arch Linux
• COSMIC version: latest git snapshot from https://aur.archlinux.org/packages/cosmic-epoch-git
• DM: SDDM with non-root setup

Log: cosmic-startup-nvidia-log.txt

[Drakulix]

COSMIC Doesn't work at all on Nvidia 550.76-1 drivers. Doesn't launch.

You are either missing the nvidia-drm.modeset=1 kernel flag or have simpledrm loaded for the nvidia gpu. This is a configuration issue on your end.

Also we don't test the nvidia-open driver, but that doesn't seem to be the issue here.

If you need more help debugging this, please post the output of drm_info on your system.

[tim77]

@Drakulix it works with nvidia-drm.modeset=1! Thanks a lot.