Abstract: On-device fixed-point deep neural network (DNN) accelerators are widely used, but the multiply-accumulate (MAC) units that perform the atomic operations in DNNs have become inefficient due ...