A Posit Based Multiply-accumulate Unit with Small Quire Size for Deep Neural Networks

Abstract

Posit is a numerical representation that is especially focused on deep neural networks (DNNs). However, a specific register called quire is necessary for a posit multiply-accumulate (MAC) unit to ensure the calculation accuracy. In this paper, we proposed posit based MAC unit that can optimize quire size according to target applications. We also develop the DNN library to explore quire size of proposed MAC unit. Experimental result with ResNet-9 showed that we achieved the same level of accuracy as Deep Positron but with the area reduction of 43%.