Could we provide input in this process? I'd really like to make sure ST avoids the mistakes they made with the CORDIC and FMAC engines by making them too specific and not general enough (i.e. with CORDIC I could be able to reprogram it on the fly using DMA so I can autonomously use it for multiple tasks and for FMAC not being able to load the X1 and X2 values autonomously and sequentially mean that the opportunity to use it as a general purpose vector multiplier are lost. If you'd allowed those two things, you'd have had a beautiful hardware FFT capability which was missed). I'd just really like to make sure that the "ST Neural Processing Unit" can be used for more general vector/matrix operations on a LEAST signed 16bit data, preferably 32bit float, since this would mean that you can REALLY make use of it BOTH for NN/AI and ALSO for audio synthesis/processing/motor control etc. Please let us help you by providing feedback before it's too late and you've already committed missed opportunities to silicon.