Course Content

Design and Implementation of High Precision Floating Point Unit for Neural Network

Abstract

This article provides various comparator designs that provide comparisons to double, single, half, and bfloat floating-point values as well as provide comparison modes for 32 and 64 bit two’s compliment integer encoded numbers. The variety of different modes described are assessable via select signal to the proposed comparators. This comparator also houses a Rectified Linear Unit (ReLU) function to leverage performance in a machine learning environment. Many forms of machine learning architectures, such as Deep Neural Network (DNN) and Convolutional Neural Network (CNN), utilize the ReLU algorithm for weight updates to their respective computational layer networks. Providing a hardware level solution to these weight updates within these networks would produce faster results for the networks respective outputs due to the speed and reliability of hardware solutions over the traditional based software solutions found in the industry today.