Is there a way to get niri to release some of its GPU memory with NVIDIA cards? I did read the wiki

[nickjj]

Edit: This affects AMD GPUs too, not just NVIDIA, see #3146 (comment).

I have a GeForce 750 Ti which only has 2 GB of memory and I noticed a lot of instability opening Firefox and Ghostty, especially when recording videos with OBS. Windows would lock up and become unresponsive in semi-random ways. All of those things use the GPU.

nvidia-smi was showing niri using 1 GB of GPU memory on a fresh boot and then I found https://github.com/YaLTeR/niri/wiki/Nvidia. This helped a lot. On a fresh boot niri now uses around 75 MB. A huge difference for the better.

But now after opening and closing a few windows, recording a few videos with OBS, etc.. the GPU usage of niri is at 310 MB and it seems to be climbing even if nothing is open except 1 Firefox window (which uses its own GPU memory). I'm basically in a fresh boot scenario except I have opened and closed stuff during this session so it feels like something isn't getting cleared at the niri level. The wiki mentions niri shouldn't use much more than 100 MB.

If I spam open a few windows that don't use the GPU like Thunar, niri's GPU allocation goes up. This makes sense since they are occupying resources. Opening 8 or so Thunar windows raises the memory by about 100 MB total. However after I close them, this memory goes back down, however it won't drop below about ~300 MB since "something" raised the baseline. For reference I have 2 displays both at scale 1.0 but when recording videos I do change the scale to 2.0 for only the display being recorded.

Edit: I don't have hard evidence but I just rebooted and if I resize windows (height and / or width) then niri will continue to use more and more GPU memory. Even if you close the window / process that was resized, the memory does not get reclaimed. Is there a leak somewhere?

Is there anything that I can do to clear this memory in a safe way without disrupting niri / open applications without logging out or rebooting?

I launch niri through uwsm in my zprofile like this if uwsm check may-start; then exec uwsm start -F -- niri-session -l; fi.

Before I understood what the problem was, I defined these 3 env vars in uwsm's env file. Figured I'd include this in any case they have an impact:

# Attempt to stabilize NVIDIA drivers on Wayland.
export GBM_BACKEND=nvidia-drm           # Force GBM to use NVIDIA DRM for buffer allocations
export __GLX_VENDOR_LIBRARY_NAME=nvidia # Ensures OpenGL / EGL apps use NVIDIA driver
export WLR_NO_HARDWARE_CURSORS=1        # Disables wlroots hardware cursor (compositor only)

I really don't know what they are doing or if they have a positive or negative outcome. It was just rummaging through many pages of Google results and Chat GPT to identify why my system was not stable. Should I even be setting these?

Edit:

Here's the nvidia-smi output:

$ nvidia-smi
Tue Dec 30 09:36:58 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.119.02             Driver Version: 580.119.02     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce GTX 750 Ti      Off |   00000000:01:00.0  On |                  N/A |
| 42%   48C    P0              2W /   38W |    1139MiB /   2048MiB |      3%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A             869      G   niri                                    294MiB |
|    0   N/A  N/A             941    C+G   walker                                  129MiB |
|    0   N/A  N/A            1043    C+G   .../lib/xdg-desktop-portal-gnome          2MiB |
|    0   N/A  N/A            3791      G   Xwayland                                  2MiB |
|    0   N/A  N/A           36422      G   /usr/lib/firefox/firefox                414MiB |
|    0   N/A  N/A           52853      G   ghostty                                 154MiB |
+-----------------------------------------------------------------------------------------+

What's interesting is only xdg portal and Xwayland popped up after I used OBS. OBS is not included here because I closed it. Is it normal for them to continue running in the background?

[nickjj]

I managed to not crash for a little over 1 day and niri is using almost 600 MB of memory just using my computer normally.

$ nvidia-smi
Sun Jan  4 15:37:00 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.119.02             Driver Version: 580.119.02     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce GTX 750 Ti      Off |   00000000:01:00.0  On |                  N/A |
| 42%   48C    P0              2W /   38W |    1405MiB /   2048MiB |      4%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A           62365      G   niri                                    578MiB |
|    0   N/A  N/A           62452    C+G   walker                                  129MiB |
|    0   N/A  N/A           62846      G   Xwayland                                  2MiB |
|    0   N/A  N/A          204811      G   ghostty                                  67MiB |
|    0   N/A  N/A          210533      G   /usr/lib/firefox/firefox                250MiB |
|    0   N/A  N/A          211011      G   ghostty                                 190MiB |
+-----------------------------------------------------------------------------------------+

niri is climbing with no limit, something is leaking somewhere but I don't know where or what is causing it.

[nickjj]

I tested it with KDE Plasma (Wayland). It's not leaking memory.

After opening an image with mpv with that setting mentioned earlier, the compositor's GPU memory goes up let's say 10 MB or however much it needs to display it and after closing mpv the memory is reclaimed. On niri, it's never reclaimed.

[nickjj]

I started a post on Reddit at https://www.reddit.com/r/niri/comments/1q4lqy6/i_believe_i_found_a_reproducible_memory_leak_in/ asking for folks to verify with different devices.

So far someone else has reported they can reproduce it with an nvidia 3070ti, Driver Version: 580.105.08, CUDA Version: 13.0 while using niri commit b5640d5.

I won't update this discussion with every reply but wanted to include 1 just to get the ball rolling.

Hopefully in a couple of days we have a bigger sample size of devices / drivers but so far it's looking like something is malfunctioning at the niri / smithay level.

Do you want me to make a proper issue out of this or are there things you'd like to see verified first?

[nickjj]

Turns out there's an existing issue open from 6 months ago in niri with a range of different systems and GPUs (AMD included) reporting the issue #1869. Their experience is exactly what's happening to me too.

Seems to be related to Smithay/smithay#1562 and I replied there with ways to reproduce using mpv since opening videos allocates a decent amount of memory, we don't have to worry about sub-1 MB size allocations that won't show up in GPU monitoring tools.

[thegabriele97]

On cosmic-comp, I noticed that simply opening and closing Nautilus increases VRAM usage by about 6 MB each time. This happens with many applications.

I realized it a few days ago because my game was stuttering due to high VRAM usage, and I noticed that even with nothing open—just the desktop—around 2 GB of VRAM were already in use.

[nickjj]

Thanks, yep it seems related to that Smithay leak which also likely affects niri.

There has been pop-os/cosmic-comp#1179 open for a while with lots of people reporting the same thing as you.

I experience the exact same symptoms as you.

[nickjj]

@YaLTeR I am able to reproduce this leak using a brand new system with an AMD GPU. I reported it on Smithay in the same thread that has been open for over a year.

The same exact reproducible steps causing the leak on my desktop with NVIDIA happen with AMD too. I am using the same distro and all of the same system / user configs (minus hardware differences).

I am quite confident you'll be able to reproduce this too. You just need to boot up, look at your GPU memory usage for niri, use your system normally for a few days, close everything and look at your GPU memory for niri. It will be quite a lot higher.

You can always open and close mpv images or videos to fast track this leak which also aligns perfectly with Smithay/smithay#1562 which mentions memory doesn't get released when closing windows.

[jadegong]

The same issue, and I did as the niri wiki said, add /etc/nvidia/nvidia-application-profiles-rc.d/50-limit-free-buffer-pool-in-wayland-compositors.json file content, but the problem still exists. The niri costs over 800MB gpu memory after 3 or 4 days.
nvidia-smi output:

Thu Jan 22 16:27:16 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.48.01              Driver Version: 590.48.01      CUDA Version: 13.1     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4060 Ti     Off |   00000000:01:00.0  On |                  N/A |
|  0%   40C    P8              7W /  165W |    1990MiB /  16380MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A            1087      G   /usr/lib/Xorg                            37MiB |
|    0   N/A  N/A            1302      G   /usr/bin/niri                           880MiB |
|    0   N/A  N/A            1443      G   Xwayland                                 26MiB |
|    0   N/A  N/A            2086      G   ...rack-uuid=3190708988185955192        477MiB |
|    0   N/A  N/A            8842      G   kitty                                   134MiB |
|    0   N/A  N/A          390397      G   kitty                                   134MiB |
+-----------------------------------------------------------------------------------------+

nvtop output:

[nickjj]

This PR #3404 appears to have resolved this, there was a leak. At the time of this comment, it's not in the latest stable release (v25.11) but you can get it from the AUR with the niri-git package until the next release is available.

[YaLTeR]

Awesome, that's good to hear