Hi,
I got similar benchmark results on SD835(CPU, GPU, DSP) whether quantize or un-quantize.
I'm not sure, but I guess SNPE use one kind of floating point bit per component on all occasions such as FP32.
(Please tell me if my opinion is wrong)
So I wonder it is possible to select FP32 or FP16 by giving a parameter according to quantization and processing unit.
Additionally, could you let me know floating point bit which NPE use since I'd like to choose the best combination of processing unit.
Thanks in advance.