This guy has a way to reduce power requirements and chip area for a neural network chip with a special operator for neural network weights.
I also showed before how to do no multiply random projections.
Addition, subtraction and a few bit hacks can get you quite far if you are smart about it.