AMD Radeon RX 7600 XT

General Tutorial & Course Discussions
,
1

Hello,

I have a problem with RealityScan 2.2.

When I try to create a Normal Detail or High Detail mesh, I get the following error:

“RealityScan could not find a usable NVIDIA graphics card. Some features require an NVIDIA GPU with CUDA compute capability 3.5 or higher.”

My system:

  • GPU: AMD Radeon RX 7600 XT (RDNA 3 / gfx1102)
  • RealityScan: Version 2.2
  • OS: Windows 11
  • AMD driver: 25.10.30.02

According to the RealityScan 2.2 release notes, RDNA 3 GPUs should now be supported.

Preview Mesh works correctly, but Normal Detail and High Detail always fail with the NVIDIA/CUDA error.

Is this a known issue, or is there any additional configuration required to use AMD GPUs for Normal and High Detail reconstruction?

Thank you.

2

Adding a second data point to this thread — same failure stage on gfx1101, but presenting as a hard crash rather than the NVIDIA message.

SYSTEM

  • GPU: Radeon RX 7800 XT (gfx1101)
  • CPU: AMD Ryzen 7 7700X 8-Core Processor
  • RAM: 32 GB
  • OS: Windows 11, build 26100
  • Adrenalin driver: 26.7.1
  • RealityScan: 2.2.0.119430

REPRO
27 NEF images of a small matte, well-textured object. Alignment completes cleanly into a single component. Preview Detail mesh completes normally. Normal Detail crashes every time, roughly 3 minutes into the process. It also crashes with Image downscale set to 4, so this does not look like a memory or resource ceiling. Cache is on an SSD with ample free space; no other GPU applications running.

CRASH
Unhandled exception, access violation reading address 0x0. Faulting module is amdhip64_7.dll.

ExceptionAddress: 00007ffb4f16936e (amdhip64_7!hipGetCmdName+0x16d1e)
ExceptionCode: c0000005 (Access violation), read from 0x0
FAILURE_BUCKET_ID: INVALID_POINTER_READ_c0000005_amdhip64_7.dll!Unknown

Stack (top frames):
amdhip64_7!hipGetCmdName+0x16d1e
amdhip64_7!hipRegisterTracerCallback+0xb7bd6
amdhip64_7!hipRegisterTracerCallback+0x8f233
amdhip64_7!hipRegisterTracerCallback+0x8cea3
amdhip64_7!hipMemcpyToSymbol+0x48
RealityScan_MvsDepthMapsV02_Hip!CapturingReality::ModelCalculation::DepthmapsV02::GpuHip::CU_V02_G_loadCameraMatricesToConstantMemoryAsync+0x59f
RealityScan+0x1c54935

Faulting instruction is cmp qword ptr [rbp],0 with rbp=0.

OBSERVATION
The call is hipMemcpyToSymbol from the HIP depth-map module while loading camera matrices into constant memory. A null dereference at that point is consistent with the device symbol failing to resolve — i.e. no code object present for gfx1101 in the fat binary, with the runtime dereferencing the null result instead of checking it.

The 2.2 release notes list gfx1101 (RX 7800 XT, RX 7700 XT, PRO W7700) as supported. The original poster here is on gfx1102 and sees Preview succeed while Normal and High fail. Different surface symptom, same pipeline stage, both non-gfx1100 RDNA 3. Worth checking whether only gfx1100 was validated for the depth-map kernels.

Crash report uploaded via the in-app dialog:
RealityScanCrash_9761_61423-667_2_2_0_119430_RS.dmp

3

Hello, thank you for reporting this. We’ll look into it.