Abstract: A precision-scalable neural processing unit, considering the quantization-sensitive of each neural network layer, has large hardware redundancy in multiplication units and shift logics. In ...