https://bugs.kde.org/show_bug.cgi?id=521675
Bug ID: 521675
Summary: Regression in Plasma 6.7.0: kwin_wayland multi-GPU
swapchain synchronization causes compositor freeze
during heavy secondary GPU compute (CUDA)
Classification: Plasma
Product: kwin
Version First 6.7.0
Reported In:
Platform: CachyOS
OS: Linux
Status: REPORTED
Severity: normal
Priority: NOR
Component: compositing
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---
***
If you're not sure this is actually a bug or valid feature request, instead
post about it at https://discuss.kde.org
If you're reporting a crash, attach a backtrace with debug symbols; see
https://community.kde.org/Guidelines_and_HOWTOs/Debugging/How_to_create_useful_crash_reports
Please remove this comment after reading and before submitting - thanks!
***
DESCRIPTION
A regression was introduced in Plasma 6.7.0 regarding how kwin_wayland handles
multi-GPU setups. Under the new multi-GPU explicit sync and swapchain behavior,
KWin automatically creates background rendering pipelines and maps processes
across all available system GPUs, even secondary, display-less cards used
strictly for compute.
When a heavy parallelized compute workload (such as llama.cpp processing a
large context window) saturates the execution queues of the secondary GPU,
KWin's compositor frame-scheduling logic misses its synchronization deadlines.
This causes the compositor to time out, crashing the llama server/process.
STEPS TO REPRODUCE
1. Boot into a Plasma 6.7.0 Wayland session on a multi-GPU setup (e.g., primary
RTX 5070 Ti driving displays, secondary RTX 5060 Ti as an internal compute
node).
2. Run nvidia-smi to confirm that kwin_wayland is now spawning background
graphical processes on the secondary compute GPU.
3. Initiate a heavy multi-GPU CUDA workload via llama.cpp using both cards
simultaneously.
OBSERVED RESULT
The compositor hits a synchronization watchdog timeout on the secondary card,
causing a rendering freeze, shell panic, and a crash of llama server or process
utilizing multi-GPU processing.
EXPECTED RESULT
The compositor should cleanly isolate desktop workspace rendering to the
primary card or gracefully handle compute-induced sync delays on secondary
cards without triggering watchdog for heavy multi-GPU workloads.
SOFTWARE/OS VERSIONS
Operating System (available in the Info Center app, or by running `kinfo` in a
terminal window):
KDE Plasma Version: 6.7.0
KDE Frameworks Version: 6.3.0
Qt Version: 6.7.1
ADDITIONAL INFORMATION
Downgrading back to Plasma 6.6.5 completely fixes the issue. Under 6.6.5,
nvidia-smi confirms that kwin_wayland restricts its processes entirely to GPU 0
(the primary display card) and leaves the secondary compute card completely
uninhibited, preventing any cross-talk timeouts. Using the KWIN_DRM_DEVICES
environment variable to manually bypass this in 6.7.0 unfortunately results in
a black screen at boot.
--
You are receiving this mail because:
You are watching all bug changes.
