Dear NXP team,
1) I am interested to know if NXP has a plan in opening up the access to NPU for general purpose computing. I came across the reference manual Section 24.2:
eIQ Neutron NPU does not have any user-configurable registers. All NPU configuration and control functions must be performed using NXP's eIQ Toolkit software. See References for links.
That means we cannot use NPU for other general purposes, which is kind of a waste. Although NPU is specially designed for machine learning, it can also be used by other signal processing and statistics applications. For example, Add Tensor-Tensor, Convolution 2D/1D, Maximum, Mean, Minimum and etc. can be used in many applications, they are kind of generic.
It would be great if NXP can consider opening up the registers access so that we can manually configure it to perform general purpose computing, rather than relying on the eIQ tool which only support machine learning computations. Imagine that if this is possible, it will be like CUDA-NVIDIA 18 years ago! This should definitely benefit NXP.
2) If there is no plan to open up the NPU for general purpose computations, is there any other ways we can use eIQ tool to generate some "fake" CNN model and run it for general purpose computation?
Thanks.
