DFXP for DNN Architectures
摘要
This chapter examines the use of dynamic fixed point number systems for deep neural networks. We introduce the concept of DFXP and compare it to block floating point systems highlighting their similarities and differences. In addition, we review existing DNN architectures that use DFXP and compare their performance. Additionally, we discuss the various factors that impact DNN performance when using DFXP and explore different approaches for determining the optimal settings of these factors.