Hi,
We are seeing intermittent GPU crashes in a UE 5.6 shipping build using Direct3D 12. The crashes appear to be random and do not consistently point to one shader or rendering pass.
The NVIDIA Aftermath dumps report several different failure types:
- “Shader Instruction Error in this shader location”
- Compute shader failures, including locations in/around NaniteClusterCulling.usf,SlateElementPixelShader.usfSlateElementPixelShader.usf,SlateElementPixelShader.usf
- Vertex shader failures, including locations in/around BasePassVertexShader.usf
- GPU hangs / device state: Hung
- GPU page faults / device state: Error_DMA_PageFault
- Failures during Nanite rasterization
- In some cases, fragment shader failures, including VirtualShadowMapCompositePS
Our project also makes extensive use of Nanite World Position Offset (WPO) and Nanite Spline Meshes. Because several crashes involve Nanite-related culling, rasterization, and vertex-shader paths, we would especially like to know whether there are known UE 5.6 limitations, fixes, or diagnostic steps for these two features.
The reported source location varies between crashes. For example, one dump points to HZB texture sampling in a compute shader, while another points to scene-data unpacking in a vertex shader. This makes us suspect that the reported shader line may be where the GPU error becomes visible, rather than the root cause.
We have reproduced this on different NVIDIA GPUs, including an RTX 4060 Laptop GPU and an RTX 2070, on Windows 11. We tested multiple NVIDIA driver versions, from 591.74 through the 616.xx series, but the issue still occurs.
At the moment, we do not have a reliable minimal reproduction case. The crashes are intermittent and can occur in different rendering paths.
Has anyone seen similar UE 5.6 D3D12 crashes involving Nanite, HZB/cluster culling, Base Pass vertex shaders, or Virtual Shadow Maps? Are there known UE 5.6 fixes, hotfixes, engine changelists, recommended diagnostic options, or feature toggles that could help isolate this issue?
We also reproduce the issue more frequently on AMD GPUs. However, the AMD crashes provide much less diagnostic information than the NVIDIA Aftermath dumps, so we have not been able to identify the affected shader, pass, or GPU operation on AMD yet.
This may indicate a cross-vendor D3D12/engine issue, but we currently cannot determine whether the higher AMD crash rate has the same root cause as the NVIDIA shader instruction errors, hangs, and page faults.
We can provide additional Aftermath dumps, DRED information, logs, and command-line/configuration details if needed.
Thanks.
NVdmp.7z (3.7 MB)