https://bugs.kde.org/show_bug.cgi?id=521814
adri365 <[email protected]> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |[email protected] --- Comment #31 from adri365 <[email protected]> --- I can confirm that `KWIN_RENDER_NODES` successfully resolves the NVIDIA dGPU runtime power management issue on my HP OMEN 16 with hybrid AMD/NVIDIA graphics. **System configuration:** - Laptop: HP OMEN 16-ap0007ns - CPU: AMD Ryzen AI 7 350 - Integrated GPU: AMD Radeon 860M - Dedicated GPU: NVIDIA GeForce RTX 5070 Laptop (8 GB) - Distribution: Kubuntu 26.04 - KDE Plasma / KWin: 6.7.5 - Qt: 6.11.2 - Mesa: 26.2.4 - Linux kernel: 7.3.0-9-generic - NVIDIA driver: 580.178.04 - Display server: Wayland **Before applying the workaround:** KWin correctly rendered the desktop using the AMD Radeon 860M, but the NVIDIA RTX 5070 remained in D0 (active). Stopping both `nvidia-powerd` and `omen-space-daemon` did not resolve the issue. KWin was still accessing the NVIDIA devices. **Workaround:** I configured the following environment variable for the `plasma-kwin_wayland.service` systemd user unit: `KWIN_RENDER_NODES=/dev/dri/by-path/pci-0000:06:00.0-card` This restricts KWin's directly opened render nodes to the AMD GPU. **Results:** - The NVIDIA RTX 5070 successfully enters D3cold. - Runtime power management reports `suspended`. - The desktop continues rendering correctly using the Radeon 860M. - Both `nvidia-powerd` and `omen-space-daemon` remain active. - Running `nvidia-smi` wakes the NVIDIA GPU to D0. - Closing `nvidia-smi` allows it to return to D3cold within a few seconds. - A direct CUDA Driver API test successfully initialized CUDA, created a context, allocated 128 MB of VRAM and performed a device memory operation. - After destroying the CUDA context and exiting the application, the RTX 5070 automatically returned to D3cold. Interestingly, `kwin_wayland` still appears among the processes accessing NVIDIA device files, even while the GPU successfully enters D3cold. This confirms that the `KWIN_RENDER_NODES` workaround enables proper NVIDIA runtime power management on this recent AMD/NVIDIA hybrid laptop without disabling CUDA or NVIDIA services. Thank you for implementing this option! -- You are receiving this mail because: You are watching all bug changes.
