Skip to content

Periodic, non-fatal GPU work-submission stall (~7-9s cycle) across multiple renderers on RTX 5080 — distinct from Xid 109/crash reports #1260

Description

@kilhaasi

Summary:
Reproducible, perfectly periodic stall where the GPU alternates between ~97-98% and ~17-25% utilization on a strict ~7-9 second cycle, across multiple unrelated game engines, with no Xid error, no crash, no driver recovery — this appears distinct from the Xid 109 CTX SWITCH TIMEOUT crash reports (#1080, #1097, ValveSoftware/csgo-osx-linux#3290) and the general desktop stutter in #777.

System:

  • Bazzite (Fedora Kinoite-based), KDE Plasma 6, Wayland
  • CPU: Intel Core Ultra 7 265K
  • GPU: RTX 5080 (GB203), 16GB VRAM
  • Driver: nvidia-open, tested 610.43.02 and 595.71.05 (both from the experimental-bleeding-edge branch per in-game diagnostics — no mature/stable branch currently exists for this GPU)
  • Display: 5120x2160, tested at 120Hz+VRR and fixed 60Hz/no-VRR — no effect on the pattern

Reproduction matrix (all at max settings, 5120x2160):

Title Engine Renderer path Result
The Last of Us Part I Naughty Dog proprietary DX12 via Proton/VKD3D-Proton Periodic stall (FPS ~120↔17)
RoboCop: Rogue City Unreal Engine 5 DX12 via Proton/VKD3D-Proton Periodic stall
Counter-Strike 2 Source 2 Native Vulkan, no Proton Periodic stall (FPS ~60-100↔15)
Team Fortress 2 Source 1 Native OpenGL/Vulkan, no Proton Clean, no oscillation

Proton is therefore not a required factor (CS2 reproduces natively) — the discriminator looks more like renderer complexity (multi-queue/async-compute usage), present in the three affected titles and absent in TF2's simpler forward renderer.

Key measurement: during the low-utilization phase, GPU clocks/power state do NOT drop (stays P0, ~2780-2820MHz, no HW/SW slowdown flag) — the GPU has full clock/power headroom throughout, it simply isn't being handed work. This also reproduces identically in a completely static scene (game's own pause menu / title screen with no simulation running), ruling out scene-content-driven load variance.

Ruled out: VRAM (never saturated), system RAM/swap, disk I/O (PSI = 0), HDR, native-Wayland vs Xwayland, CPU P-core/E-core pinning (all threads pinned to P-cores via taskset — no change), network/telemetry latency, other GPU clients (nvidia-smi pmon confirms only the game process has load), VRR/refresh-rate/fps-cap combinations, Proton version (Hotfix vs 10.0 — the latter caused an unrelated single Xid 109 hang on savegame load, not the periodic pattern), driver version (610.43.02 vs 595.71.05, same branch).

Happy to provide nvidia-bug-report.sh output, raw nvidia-smi CSV logs, or test further permutations if useful.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions