NEW Artificial Intelligence Lab: Aura SDK (Alpha) is released with native Hexagon NPU offloading on Snapdragon X.

Troubleshooting & FAQ

Common installation errors, driver version mismatches, and compilation issues.

Below are solutions to common configuration problems encountered during setup and runtime integration of the Aura SDK.

Driver Loading Failures

If you encounter driver loading failures during engine initialization:

Error: Failed to register driver backend: libQnnHexagon.dll not found

Ensure the Snapdragon driver paths are added to your system environment variables. You must reboot or restart your terminal session after modifying environment variables.

Core Isolation & Memory Integrity

If the NPU engine initialization fails to load unsigned hexagon drivers:

Error: ADSP driver registration rejected.

Ensure you have disabled Memory Integrity in Windows Security > Device Security > Core Isolation and restarted your computer.

Compilation Target Mismatches

When lowering ONNX graphs using the Aura Compiler, ensure your Hexagon target matches your target hardware version:

  • Snapdragon X Elite / Plus: Use --target hexagon-v73 or later.
  • Snapdragon 8 Gen 2 / 3: Use --target hexagon-v70 or v73 depending on mobile SoC specifications.

Precision Loss During Quantization

If your model begins generating garbage characters or repetitive loops post-quantization:

  1. Increase the size of the calibration dataset (e.g. from 100 samples to 500 samples).
  2. Ensure the calibration dataset contains text patterns matching the target domain of your model.
  3. If perplexity remains high, consider fallback options for specific deep layers using the Aura Compiler override flags.