Please describe your issue

Marvel Rivals hangs and then dies with Unreal's render-thread watchdog assert. The hang is preceded by ~1 second by an NVIDIA Xid 13 graphics exception naming the game's own GameThread, and the crash lands exactly 60 seconds after the Xid — which is GTimeoutForBlockOnRenderFence.

This happened twice in three days of play. Both occurrences have an identical signature and there is no other Xid in the log across ~14 hours of captured play.

Unlike #3180 (Space Marine 2, Ampere, Class 0000c797, a 3D class), this fault is on a compute class on Blackwell, and the kernel reports the specific failing scheduler checks rather than only the class id.

Kernel message (both occurrences, verbatim):

NVRM: Xid (PCI:0000:01:00): 13, Graphics Exception: SKEDCHECK06_REGISTER_COUNT failed
NVRM: Xid (PCI:0000:01:00): 13, Graphics Exception: SKEDCHECK16_CTA_THREAD_DIMENSION_ZERO failed
NVRM: Xid (PCI:0000:01:00): 13, Graphics Exception: ESR 0x407020=0x802000 0x407028=0x100 0x40702c=0x3ab712 0x407030=0x0
NVRM: Xid (PCI:0000:01:00): 13, pid=1657731, name=GameThread, Graphics Exception: channel 0x00000027, Class 0000cec0, Offset 00000000, Data 00000000

SKEDCHECK16_CTA_THREAD_DIMENSION_ZERO suggests a compute dispatch reaching the GPU with a zero thread-block dimension. On D3D12 a Dispatch() with a zero thread-group count is legal and is a no-op.

Unreal's assert, from the crash dump (CrashContext.runtime-xml):

Fatal error: [File:E:\workspace\S10.0_release\UnrealEngine\Engine\Source\Runtime\RenderCore\Private\RenderingThread.cpp]
[Line: 1387] GameThread timed out waiting for RenderThread after 60.00 secs

Timeline — both occurrences

On the second occurrence I had a bpftrace tracer and a per-thread sampler running. They independently corroborate the sequence:

• 02:01:36 — Xid 13 (kernel log)
• 02:01:37 — frame output stops; the tracer records GameThread, RHIThread and RenderThread all entering ntsync waits. It never records the freeze ending.
• through the hang — 1–3 of the process's 241 threads are runnable (sampled live at 5 Hz). Every other thread is asleep.
• GPU — utilisation drops 91% → 1%, power 334 W → 36 W, P0 → P8, measured across the hang itself (265 samples at 5 Hz). The GPU is starved, not hung: nothing is submitted after the channel fault.
• 02:02:40 — UE's 60 s watchdog asserts, crash handler runs.

So the causal chain looks like: the channel takes a graphics exception → the submitted work never completes → RHIThread/RenderThread never finish → GameThread's render fence never signals → Unreal kills the process at 60 s.

What I checked and ruled out

• Thermal / power throttling — no. The GPU is idle at ~36 W during this hang. clocks_event_reasons shows no thermal or power-cap reason (the Reliability bit is set in 73% of all recorded play on this card, so it is this card's normal state, not a signal).
• Memory pressure — no. 61 GB RAM, 39 GB free at crash time (from the crash context).
• Wayland — no. Also reproduces on X11.
• Machine-wide stall — no. The game's own network traffic continues throughout the hang (median 111 pkt/s received over 116 samples, its UDP sockets still open); other processes run normally.
• Bus lock / split lock traps — present on this system (15,172 over ~9 h, 10,665 of them from GameThread) but zero during the hang or any freeze window, so anti-correlated.
• Overclocking — none applied. Stock clocks.

What I am NOT claiming

• I have not shown this is the same problem as #3180. Same Xid number, but that report is an Ampere 3D class (0000c797) and this is a Blackwell compute class (0000cec0).
• This machine also gets frequent sub-5-second freezes (~7/h) with the same thread-blocking pattern but no Xid at all. Those are almost certainly a separate issue and I am not reporting them here.
• I have not tried to isolate which shader or dispatch produces the fault, and I do not have a minimal reproducer. It is not reliably reproducible on demand — roughly once per multi-hour session.

Software information

• Game: Marvel Rivals (Steam appid 2767030), Unreal Engine 5.3.2-3870120+++depot_marvel+S10.0_release
• Renderer: Direct3D 12, feature level SM6 (per UE crash context: RHI.RHIName: D3D12, RHI.FeatureLevel: SM6)
• Settings: default. DLSS frame generation is off (bDlssFrameGeneration=False); DlssPreset=1 is set.
• Launch options: SteamDeck=0 DXVK_NVAPI_VKREFLEX=1 PROTON_ENABLE_WAYLAND=1 PROTON_VKD3D_LOWLATENCY=1 mangohud %command%
  (also reproduces without Wayland)

System information

• GPU: NVIDIA GeForce RTX 5090 (Blackwell), device id 2B85
• Driver: NVIDIA 615.71.09 (Open Kernel Module)
• CPU: AMD Ryzen 9 9950X3D
• RAM: 64 GB
• Kernel: 7.2.5-1-cachyos
• Proton: proton-cachyos-11.0-20260703-slr
• D3D12 layer actually loaded: vkd3d-low-latency initial-release-1-g565d695 — a third-party
  fork (netborg-afps/vkd3d-low-latency) based on
  vkd3d-proton v3.0.1, which PROTON_VKD3D_LOWLATENCY=1 substitutes for the bundled
  vkd3d-proton. Verified by sha256: the d3d12core.dll in the game's prefix is byte-identical to
  the fork's copy, not to the bundled vkd3d-proton (vkd3d-1.1-5438-g3dfc6f0), which is NOT loaded.
  Corrected after the initial post — I first reported the bundled version by mistake.
  I have not yet reproduced with stock vkd3d-proton; doing that is the obvious next step and I will
  report back. If the fault does not occur on stock vkd3d-proton, this belongs in the fork's tracker
  and I will move it.
• Distribution: CachyOS

I still have the full crash dump (CrashContext.runtime-xml, UEMinidump.dmp) and the kernel log for both occurrences, plus the frametime, per-thread and GPU-state captures around the second one. Happy to attach or run anything specific — including with VKD3D_CONFIG options or a newer vkd3d-proton if that would help narrow it.