> It would be fair to mention that this is happening for you with a decade old GPU using an EOL driver, which sucks, but is unlikely to be a common experience.

The drivers are still getting maintained by NVIDIA until August 2026. They also got classified as "legacy" on paper 1 day before I installed them.

The compositor memory leak is affecting a lot of people. Since COSMIC and niri both use the same one (smithay), there's threads on GitHub with people using modern GPUs, both NVIDIA and AMD who experience it. There's a lot of replies across all of the different open issues.

The GPU allocation issue on Wayland (separate from the memory leak) also has hundreds of replies on the NVIDIA developer forums with people using new NVIDIA cards with the latest drivers.

The thing is, most people don't talk about either of them because if you have 8+ GB of GPU memory and turn your computer off every night then you won't experience this problem since all GPU memory allocations get reset on shutdown. It happens to be more of a direct problem for me because I have 2 GB of GPU memory but that doesn't mean the problem isn't common. The root cause is still there. Even if I switched to an AMD GPU the niri / smithay memory leak would be present. Instead of rebooting twice a day, if the GPU had 8 GB of memory I'd have to reboot every 2 days (x4 basically).

Since I opened that issue on GitHub NVIDIA did acknowledge it and suggested I try their experimental egl-wayland2 library. I did try that and it hasn't fixed it fully but it has made GPU memory allocations more stable. It even fixed 1 type of leak in niri. This library is decoupled from the drivers themselves as far as I know. I mean, this same library could still be used for the 590 series, it's not 580 specific which means it's not dependent on your GPU model.

▲

formerly_proven 7 hours ago | parent [-]

> The compositor memory leak is affecting a lot of people. Since COSMIC and niri both use the same one (smithay), there's threads on GitHub with people using modern GPUs, both NVIDIA and AMD who experience it. There's a lot of replies across all of the different open issues.

But then this sounds like a bug in that particular compositor rather than the driver(s)?

fwiw, I have a modern nvidia card, and use the proprietary drivers, and Wayland (KDE/KWin), and that box has a few weeks of uptime.

	▲	nickjj 6 hours ago \| parent [-]
		> But then this sounds like a bug in that particular compositor rather than the driver(s)? The NVIDIA GPU memory allocation issue affects all NVIDIA cards, at least based on that forum post where there's a good amount of people replying with similar issues with a large combo of cards + drivers. You probably don't notice it because you rarely ever use all of your GPU's memory at once. Try running `watch nvidia-smi` and then open multiple copies of every hardware accelerated app you have available. Once you reach nearly the max GPU memory apps will either start crashing or fail to open and if you get semi-unlucky your compositor (including KDE on Wayland) will crash if it's the app trying to allocate resources to render the window. I've had plasmashell or kwin hard lock / crash many times with just a small amount of testing. The expectation is the driver allocates those memory resources to system memory instead of denying the memory to the app. It works correctly in X11 or if you have an AMD card, in Wayland and X11. The leak is separate and is compositor specific, possibly related to NVIDIA driver bugs (to some degree) but this leak wouldn't be experienced unless you used this compositor. This leak is the compositor doesn't release any GPU memory after any window is closed, so simply opening and closing apps will cause the leak. Combine this with the first problem and that's how you end up rebooting every few hours with a lower end GPU. AMD and NVIDIA are affected. The post goes into all of these details and there's reproducible tests, and even demo videos showing the first GPU memory allocation problem in Plasma Wayland but not Plasma X11. It also links to all of the related GitHub issues that I could find.