Tachyon – SNPE / QNN DSP runtime fails with incompatible binaries (needs Ubuntu 22.04?)

Tachyon
1

I’m trying to run a YOLOX-based object detection model on a Tachyon (QCM6490) using Qualcomm’s SNPE SDK. I was able to convert and quantize the model successfully on my x86 host, and CPU inference on the Tachyon works fine — just really slow (about 2 seconds per frame).

The problem is when I try to run with --use_dsp I get this:

The selected runtime is not available on this platform. Continue anyway to observe the failure at network creation time.
error_code=2008; error_message= QnnBackend_DeviceCreate() failed; QNN_COMMON_ERROR_INCOMPATIBLE_BINARIES: Loaded libraries are of incompatible versions; error_component=QNN; line_no=2073; thread_id=547613409296
Segmentation fault

Couple things I’ve noticed:

  • The Tachyon ships with Ubuntu 20.04, but SNPE officially needs Ubuntu 22.04 (Python 3.10, newer libc++).

  • There’s no libQnnHtp.so anywhere on the device.

  • There’s basically no documentation for setting up the onboard AI engine — the only mention is a tiny blurb in the camera section.

So I’m wondering:

  • Is DSP/HTP inference even supported out of the box on Tachyon right now?

  • Do I need a different BSP or runtime package to make QNN work?

  • Or should I be upgrading the OS to Ubuntu 22.04 to match SNPE’s requirements?

Would love to know if anyone has actually gotten SNPE DSP inference working on Tachyon, or if this is just not set up yet.

2

You might find these projects on Hackster.io useful—check it out!