Disclaimer: Claude Sonnet 5.5 helped me diagnose, fix and document this.
Summary
With two DisplayPort monitors at 2560x1440 @ 144Hz, the right monitor flickers and shows artifacts. It does not happen if either monitor is at 120Hz. Locking the memory clock with nvidia-smi -lmc 14001,14001 eliminates it, so I believe the memory clock P-state transitions are not fitting inside the vertical blanking window when two high-refresh displays are active.
System
- GPU: GeForce RTX 5070 (GB205), PCIe link currently 2.5GT/s x16 at idle
- Driver: 615.71.09 open kernel module (NVIDIA UNIX Open Kernel Module)
- OS: Fedora 44, kernel 7.2.7-200.fc44
- Desktop: KDE Plasma 6.7.5 / KWin Wayland, nvidia-drm modeset=Y fbdev=Y
- Board: ASRock B550M-C, BIOS P3.90
- Display: 2 x 2560x1440 monitors on DP-1 and DP-2 (models: [fill in]), VRR off, HDR off
- Packaging note: userspace from negativo17 (nvidia-driver 615.71.09-5), kernel module from RPM Fusion (akmod-nvidia 615.71.09-3)
Display: 2 x LG UltraGear 27" monitors (identical), both on DisplayPort (DP-1 left, DP-2 right)
Panel: 2560x1440, 10 bits per color, EDID range 48-144 Hz
EDID: vendor GSM, product code 23423, version 1.4, made week 1 of 2021
Settings: VRR off, HDR off
Behavior
| Left | Right | Result |
|---|---|---|
| 144 | 144 | flicker/artifacts on right |
| 144 | 120 | fine |
| 120 | 144 | fine |
| 120 | 120 | fine |
It started immediately after boot and was reproducible on demand by switching refresh rates (kscreen-doctor output.DP-2.mode.2560x1440@144). No Xid errors, link-training failures or kwin errors in the journal. The previous driver here was 610.57.04, and I don’t know if it was affected.
Evidence for reclocking
At 144/144 with no lock, sampled with nvidia-smi at 1Hz, memory clock moved from P3 (7001 MHz) to P5 (810 MHz) to P8 (405 MHz), then to P0 (14001 MHz) with ~4% GPU load, and the flicker occurred during this. At 144/120, the GPU sat in P8 (405 MHz) steadily with no flicker. After sudo nvidia-smi -lmc 14001,14001, both monitors at 144Hz ran with P0 / 14001 MHz fixed and no flicker.
Workaround
Run both at 120Hz, or lock the memory clock.
Is this a known issue with multi-monitor high-refresh reclocking on Blackwell? Happy to test any driver or kernel parameters. nvidia-bug-report.log.gz attached.
nvidia-bug-report.log.gz (198.6 KB)
Edit: Updated to kernel 7.2.8 and issue still present.