42.20: Fortnite client GPU hang (DEVICE_HUNG) in LumenReflections when launching a UEFN session; the reset also kills UEFN

General Issues and Bug Reporting
, , ,
1

Summary

Build: ++Fortnite+Release-42.20-CL-58011042 (client and UEFN)

Hardware / OS:

GPU: NVIDIA GeForce RTX 3080, driver 616.92
CPU: AMD Ryzen 9 5900X, 32 GB RAM
OS: Windows 10 22H2 (10.0.19045.7725)
Client: D3D12, Aftermath on, DLSS on, 2560×1080 windowed
Settings: Reflections quality 2, Global Illumination quality 2 (High)

Summary:
Since 42.20, launching a live UEFN session often crashes the Fortnite client with a GPU crash. Because the GPU driver resets, UEFN crashes at the same moment. This never happened on 42.10 with the same hardware, driver and project. It has happened four times on 23–24 Sep.

Please select what you are reporting on:

Unreal Editor for Fortnite

What Type of Bug are you experiencing?

Stability

Steps to Reproduce

  1. Open a UEFN project that uses Lumen.
  2. Launch Session. The Fortnite client starts or is already running.
  3. The client crashes, sometimes while still in the front end, before or while the island loads. UEFN often crashes at the same time.

Expected Result

It should not crash

Observed Result

the D3D12 device is removed, and both the client and UEFN write GPU crash reports.

Platform(s)

PC

Additional Notes

Evidence (from FortniteGame.log and the Aftermath dump):

DXGI_ERROR_DEVICE_REMOVED with Reason: DXGI_ERROR_DEVICE_HUNG (D3D12Viewport.cpp:537)
Active GPU breadcrumbs: Scene > RenderDeferredLighting > DiffuseIndirectAndAO > LumenReflections. Three of the four crashes stopped here; the fourth was in UpdateAllPrimitiveSceneInfos.
Aftermath: faulted warp in compute_02, “MMU Fault Error: A shader instruction caused an MMU fault when accessing memory”.
PageFault at GPU address 0x100000000. No tracked resources, heaps or recently released resources in range.
Video memory at the crash: 1.0 GB used of a 9.3 GB budget, so not memory pressure.
Crash callstack hash: 9B763042B39078C3CEC8672660F3BD2BB4756D27. The same hash appears on three separate crashes.
The latest crash happened right after front-end UI loading (the Reload playlist art request), about 90 minutes into a running client session.

Crash GUIDs:

Client: UECC-Windows-A87628C246A0AD9D85A8FAA530C9BA48, UECC-Windows-E050961B4E7C0BF6A3A7DC825900A8E5, UECC-Windows-DBEA8272432D0340089832BA48CC8110, UECC-Windows-40580E944EAA4155624A4895B3B24AB6
UEFN (same event as client E050…): UECC-Windows-B7C1975D462B69E6904E7FB9DE74ECCB