GPU Page fault crash on LumenRadiosityHardwareRayTracingCS

Development Rendering
, ,
1

Hi, we’re currently running on a consistent GPU page fault crash that occurs when players load into a new map in our game. We have not seen this issue in 5.6, prior to upgrade to 5.7. This seems to occur with the `LumenRadiosityHardwareRayTracingCS` shader attempting to access already deleted resources.

I have attached a couple screenshot from NVIDIA NSight showing where the call is. It looks like it’s while loading vertex attribute (LoadVertexPositionFloat3) from TraceLumenMinimalRay. Our maps do load a lot of assets so I wonder if that could have caused the BLAS to be slow to construct.

Machines encountered GPU crash (so far):

  • NVIDIA GeForce RTX 3080, Driver Version: 32.0.16.1047, Date: 5-19-2026, and Driver version: 32.0.15.9186, Date: 1-20-2026
  • NVIDIA GeForce RTX 4070 Ti, Driver Version: 32.0.16.1047, Date: 5-19-2026
  • NVIDIA GeForce RTX 4070 Ti SUPER, Driver Version: 32.0.15.9186, Date: 1-20-2026
  • NVIDIA GeForce RTX 5080, Driver Version: 32.0.16.1047, Date: 5-19-2026

I’m sort of at a loss on how to resolve these crashes. Perhaps there have been fixes that we can pull from future patches. Looking for suggestion and advice from the team. Thank you,

-Trung

[Attachment Removed]

2

Steps to Reproduce
In 5.7, GPU page fault crash on NVIDIA GPUs when loading into a new map.

[Attachment Removed]

3

Hello,

This could be a residency issue that has since been fixed by the refactors to the system in 5.8. If you are able to reproduce the issue, can you see if the crashes stop with r.RayTracing.UseReferenceBasedResidency=0? This will make all RT geometry to stay resident and increase the amount of VRAM used which could cause OOM crashes, but could at least verify the problem area.

We have an internal issue that looks like it’s tracking a bug in a similar place in 5.7.4 and the associated fix in 5.8 is associated with this change:

CL#53307123 (31317c) [HWRT] Fix stale RayTracingBuildRequestIndex after ProcessBuildRequests Reset(). Also remove bRemoveFromRequestArray and let the calling code handle that.

There’s also this possibly related one:

CL#53436089 (e23503) [HWRT] Fix PendingStreamingRequests not being processed when rendering is disabled/paused.

- this prevented always-resident from being streamed-in while loading screens are active in some circumstances.

However, this system was refactored somewhat for 5.8 so the fixes for 5.7.4 may not be as straight forward as integrating that CL.

[Attachment Removed]

4

Thank you Alex. We’ve been playtesting for the last couple days with r.RayTracing.UseReferenceBasedResidency=0 and that seems to make the crash go away. That’s a positive option.

We can take a look at these couple CLs mentioned above and determine if they should help as well.

Thank you.

[Attachment Removed]