COMPUTATION UNIT AND METHOD FOR PERFORMING COMPUTATIONS FOR DEEP NEURAL NETWORKS

摘要

A computation unit for performing computation of a neural network layer is disclosed. A plurality of processing element (PE) units are arranged in an array. A first input value is provided in parallel in the input dimension of the array during a first processing cycle, and a second input value is provided in parallel in the input dimension during a second processing cycle. The calculation is performed based on the weights stored by the PE unit. An adder is coupled to the first set of PE units to produce a first sum that is a result of the calculation by the first set of PE units during a first processing cycle, and produces a second sum that is a result of the calculation during a second processing cycle. A first adder is coupled to the first adder, stores the first sum, and moves the first sum to a second accumulator before the second sum is stored.