WebGPU performance kept degrading slowly from 60 FPS to about 5 FPS since February 2026; persists reboot
Categories
(Core :: Graphics: WebGPU, defect, P5)
Tracking
()
| Tracking | Status | |
|---|---|---|
| firefox-esr115 | --- | unaffected |
| firefox-esr140 | --- | unaffected |
| firefox150 | --- | fixed |
| firefox151 | --- | fixed |
People
(Reporter: Fox, Unassigned)
References
(Blocks 1 open bug, Regression)
Details
(Keywords: regression)
Attachments
(1 file)
|
44.90 KB,
text/plain
|
Details |
User Agent: Mozilla/5.0 (X11; Linux x86_64; rv:149.0) Gecko/20100101 Firefox/149.0
Steps to reproduce:
I've started developing my own THREE.js-based webapp on 1st of February 2026.
I compared the THREE.js WebGL and WebGPU renderer and decided to develop with the WebGPU renderer.
Back then I noticed that WebGL was capped at 120 FPS and WebGPU was capped at 60 FPS and WebGPU had some stuttering on some THREE.js examples. This was a minor issue to me.
Specifically I compared performance here:
- https://threejs.org/examples/?q=clearco#webgl_materials_physical_clearcoat (WebGL)
- https://threejs.org/examples/?q=clearco#webgpu_clearcoat (WebGPU)
However, since then Firefox WebGPU performance (in my app, but also the linked example) kept degrading from day to day, even without any code changes which should affect it. Reopening the tab or Restarting Firefox did not solve the issue.
I was unable to reboot my machine during this time.
It kept getting worse every week:
By end of February I was barely getting 30 FPS in WebGPU while still reaching 120 FPS in WebGL.
Around mid March, performance had degraded to about 10 FPS, still 120 FPS in WebGL.
Now, early April, it's been at 5 FPS all week now, really hindering my workflow - WebGL still at 120 FPS.
I always thought I could fix the WebGPU issues by rebooting, but I have done that now and the FPS remains at about 5 FPS, even after a reboot. So something must have broken or degraded over time.
I might have upgraded packages during that time, but performance also got worse while I definitely wasn't updating.
Clicking the FPS counter in the clearcoat example above reveals that "Scene" rendering time is about 3ms while "Miscellaneous & Idle" is about 150-200ms.
So something appears to waste time between draws.
However this isn't just the case in THREE.js or its clearcoat example, most WebGPU based pages are slow now:
- https://webgpu.github.io/webgpu-samples/?sample=rotatingCube is showing considerable stutter (not sure about exact framerate).
- https://webgpu.github.io/webgpu-samples/?sample=metaballs is also at 7 FPS but was working fine when I checked it in February.
It doesn't seem to affect compute-only pages like the one mentioned in https://bugzilla.mozilla.org/show_bug.cgi?id=1982280 (which does not appear to affect me on Linux, but also doesn't seem to be affected by my issue).
Native applications glxgears (OpenGL) and vkcube (Vulkan) seem to be unaffected.
Power saving or performance settings also don't affect the issue.
This is on:
- Laptop: ASUS Vivobook S15 OLED S5506
- CPU: Intel(R) Core(TM) Ultra 7 155H
- GPU: Meteor Lake-P [Intel Arc Graphics]
- Linux 6.19.10-1-cachyos (last full upgrade about an hour ago)
- Firefox 149.0 (64-bit)
- mesa 2:26.0.3-2
- wayland 1.24.0-1.1
- plasma-desktop 6.6.3-2.1
Actual results:
Performance degraded over time, across Firefox restarts and even persists machine reboot.
It was okay in February but as I kept using WebGPU, most WebGPU based pages have slowed down considerably.
Expected results:
Performance should not get worse from day to day.
I expect 120 FPS or at least 60 FPS, even after considerable usage of WebGPU.
Comment 1•7 months ago
|
||
The Bugbug bot thinks this bug should belong to the 'Core::Graphics: WebGPU' component, and is moving the bug to that component. Please correct in case you think the bot is wrong.
Comment 2•7 months ago
|
||
Will you please bisect Nightly builds using mozregression to determine when this was broken? That can help us narrow down what might have caused this issue much more quickly.
Comment 3•7 months ago
|
||
Also, you've provided a lot of great details for your environment, thank you! Just for completeness, could you please attach the report from about:support as an attachment to this bug, using the text from Copy raw data to clipboard? You can paste directly into comment entry here, and Bugzilla will ask you if you want to create an attachment.
Also also, hello fellow CachyOS fan! :saluting
I feel this might be a caching issue (where some cache slowly filled up, with horrible lookup times or similar), because it gradually became worse, even between version updates. I'm certain that webgpu used to work fine, even with public Firefox releases from 2026... initially. Only using webgpu daily worsened the situation over time, which feels like some persistent data structure kept growing / getting worse (the use of which might have been added way before I noticed the issue).
So I bisected this from 2025-01-01 to 2026-04-10.
The bisection gave me this https://phabricator.services.mozilla.com/D275914 (December 2025) - which is a generic wgpu update.
I also did more tests on other websites, measured using https://github.com/mrdoob/stats.js/ in Firefox:
- https://wgpu.rs/examples/?backend=webgpu&example=cube (10FPS)
- https://wgpu.rs/examples/?backend=webgpu&example=mipmap (10FPS)
- https://wgpu.rs/examples/?backend=webgpu&example=water (10FPS, visibly slow)
- https://wgpu.rs/examples/?backend=webgpu&example=bunnymark (10FPS, visibly slow)
I assume the next thing to do would be to bisect wgpu.
So I cloned the repo and ran the native bunnymark sample (cargo run --bin wgpu-examples bunnymark):
f7b650d55c6c4a546db236a0779f18fe4ee7d82f (latest) INFO wgpu_examples::framework] Frame time 8.33ms (120.0 FPS)
3f02781bb5a0a1fe1922ea36c9bdacf9792abcbc (firefox-bad) INFO wgpu_examples::framework] Frame time 8.34ms (120.0 FPS)
In other words: I was unable to reproduce it in wgpu standalone; animation is very smooth.
So it looks like this issue only exists in Firefox, but not the native wgpu examples?
If I should continue with wgpu bisection, please let me know how I could (quickly) bisect wgpu in Firefox OR how I could bring the native example closer to what Firefox does (to reproduce the issue for bisection).
Also happy to try other things (in/with Firefox) that might help analyze/reproduce/workaround the issue.
Probably irrelevant, but since my report, I also moved testing / development to Chromium which is not affected (which is using Dawn?).
I'd still prefer to return to Firefox (and also fix the issue to avoid any future user of my webgpu webapp being affected).
Looks like my problem is already fixed in nightly, but a new (unrelated) webgpu issue exists in very recent builds.
I wanted to run the profiler in a new browser session to collect data and chose random (new) builds in mozregression GUI for testing:
- 2026-04-10 does not exist yet
- 2026-04-09 does not seem to support webgpu (white canvas / no webgpu content) on https://wgpu.rs/examples/?backend=webgpu&example=bunnymark
- 2026-04-08 does not seem to support webgpu (white canvas / no webgpu content) on https://wgpu.rs/examples/?backend=webgpu&example=bunnymark
- 2026-04-07 has working webgpu (smooth animation) on https://wgpu.rs/examples/?backend=webgpu&example=bunnymark and even 120FPS (!)
So 2026-04-07 fixed all my issues with webgpu (60FPS and slowness getting worse); however, starting with 2026-04-08, webgpu is broken entirely for me.
Bisection of the fix (slow webgpu):
I bisected from 2025-12-26 to 2026-04-07 to look for the bugfix and found: https://phabricator.services.mozilla.com/D288337 (18th of March 2026)
(I also tested 2026-03-26 during bisection and got mixed results (so I skipped that build): it had smooth animation but some heavy stuttering / frame pacing issues; I'm not sure if this was bad luck or also something wrong around that time)
Bisection of the new issue (white canvas / broken webgpu):
bisected from 2026-04-07 to 2026-04-10 to look for the broken change and found: Bug 2026288. r=gfx-reviewers,lsalzman,bradwerth / Differential Revision: https://phabricator.services.mozilla.com/D291680
I don't have permission to see that change, so I can't actually comment on it to report my issue with it
I'll just leave this information here, also in case it is related to my original issue of webgpu being slow.
Please let me know if I have to take additional steps (like creating a new bug report for the broken webgpu in nightly - as I can't seem to find an issue about it).
Comment 8•7 months ago
|
||
:Fox: Woot, glad to hear the perf. issue doesn't happen for you! Could you please file a separate issue with the about:config and regression info you've got here for the blank WebGPU canvas?
Another question that's likely interesting: If you set dom.webgpu.allow-present-without-readback to false, does that resolve the blank canvas?
Todays nightly 2026-04-14 works fine (canvas no longer white & no performance regressions).
So, for completeness, I bisected it for the fix again.
White canvas (on my system) was fixed here: https://phabricator.services.mozilla.com/D293476
As the blank / white canvas issue is already resolved, should I still file a separate issue? If so, which version should I file the issue for? What info do you need from about:config?
During bisection, on one of the "bad" versions (= white canvas) I tried changing dom.webgpu.allow-present-without-readback to false (in "about:config" prior to visiting a page using WebGPU; no browser restart), but it didn't seem to affect the problem (canvas remained white on https://wgpu.rs/examples/?backend=webgpu&example=bunnymark).
Comment 10•6 months ago
|
||
As the blank / white canvas issue is already resolved, should I still file a separate issue? If so, which version should I file the issue for? What info do you need from about:config?
If both issues are resolved, then I don't see a reason to make you file another issue. Glad we could close the loop on this! ❤️ I'm going to mark this as RESOLVED.
Comment 11•6 months ago
|
||
Set release status flags based on info from the regressing bug 2005294
Updated•6 months ago
|
Description
•