Audio2Face-3D SDK Build Failure on DGX Spark - x86_64 Dependencies Downloaded Instead of ARM64

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

I am attempting to build the Audio2Face-3D SDK on DGX Spark (ARM64 architecture) but the build system downloads x86_64 dependencies, making compilation impossible.

The fetch_deps.sh script downloads x86_64 versions of all dependencies instead of ARM64:

  • cmake@3.24.1+nv3-linux-x86_64

  • ninja@1.10.0+tc5-linux-x86_64

  • gtest@1.13.0-0-linux-x86_64-static-release

  • benchmark@1.9.1-1-linux-x86_64-release

  • zlib@1.3.1-0-linux-x86_64-static-release

  • cnpy@20180601-0-cci-linux-x86_64-static-release

  • tbtsvd@0.0-21e06aa-1 (contains x86 SSE intrinsics: xmmintrin.h)

When attempting to build after this, compilation fails with architecture mismatch errors.

Alternative Attempts

I also tried using ONNX models directly with:

  • PyTorch: No GB10 (sm_121) support on ARM64

  • ONNX Runtime: Only CPU execution provider available on ARM64

  • TensorRT 10.8/10.14: Engine build fails with “unknown embedded device” and “genericShader nullptr” errors

Question

How do I build the Audio2Face-3D SDK for ARM64/DGX Spark?

Specifically, I need:

  1. ARM64-compatible packman configuration or package sources

  2. Build instructions specific to ARM64 architecture

  3. Confirmation that GPU-accelerated Audio2Face inference is supported on GB10

  4. Alternative approaches if native SDK compilation isn’t available

3

Audio2Face-3D does not have support for ARM architectures so will not work on DGX Spark