Systolic Array based Multiply Accumulation Unit for IoT Edge Accelerators

Abstract

Accelerator is a hardware that runs along with the processor and executes the key functions much faster than the processor. The Main purpose of the Accelerator is to increase speed. Deep Neural Networks has achieved wide results in the various Machine Learning Applications Such as image, video, text classification and language translation. The purpose of DNN Accelerators is to speed up the most complex Computation i.e., matrix multiplication. Systolic array Based Accelerator seems like multiply Accumulate unit with Systolic Array based multiplication followed by Adder and accumulator. Multiply Accumulate Unit comprises multiplier, adder and Accumulator. Multiplier is designed used systolic array and that output is given as one of the inputs to the adder followed by Accumulator. In this paper general Matrix based Multiply Accumulate Unit is compared with systolic array based Multiply Accumulate Unit using Xilinx ISE 14.5, various parameters like area, delay and speed are compared. Systolic Array based Multiply Accumulate Unit consumes less area of 49%, less delay of 35% and in turn provides high speed when compared with general matrix multiplier-based multiplier Accumulate unit.