I purchased the Radxa Dragon Q6A (QCS6490) specifically because the Adreno 643 GPU would allow me to run face‑recognition models (AdaFace) with full FP32/FP16 accuracy without quantisation. Unfortunately, I’ve been unable to use the GPU for any AI inference – only the NPU works.
Rubik Pi has image with the proprietory Qualcomm Graphics driver installed. Radxa OS uses the open source drivers. The QNN Backend is searching for the qualcomm drivers for inference, which is not available in the Radxa Q6A.