Shift and Accumulate Convolution Processing Unit

Convolutional Neural Network (CNN) is the state-of-the-art learning technique for image understanding in several artificial vision systems. Extensive uses of memory storage and bandwidth with high computation capacity boost up the performance of a CNN. Convolution is the fundamental operation in CNN models and hence a large number of multiplier-accumulator (MAC) unit is required to compute the convolution operations adequately. MAC processing incurs extreme computational complexity by consuming significant amounts of time, energy and area. These limitations in computation have led to the exploration of different architecture of MAC to meet the demand of CNN processing. To increase the overall speed of a CNN model, an architecture called SAC (shift-accumulator) is proposed that reduces the number of overall convolution computations. Further, the SAC architecture is especially designed for convolution operations used in image sharpening, edge detection, blurring etc. The proposed SAC architecture facilitates faster convolution computations and reduces the overhead in terms of area and power than the conventional MAC architecture. Although the accuracy is slightly reduced, it does not significantly affect the efficacy of computational unit used in CNN processing.