Posit for DNN Architectures

This chapter discusses the Posit number system, which is proposed as an alternative to the floating point representation. The chapter explains how Posit uses bits more efficiently, allowing for better accuracy with the same number of bits and affording a better dynamic range. The chapter describes how the real number is represented in the Posit format and the common parameters of this representation. It also explains the available approaches for the selection of these parameters and the resultant trade-offs between accuracy and hardware efficiency. The chapter discusses the tapered accuracy of the Posit format, which makes it more suitable to represent normally distributed data efficiently, such as those found in deep neural networks. In addition, the chapter classifies DNN architectures that use the Posit number system and discuss the pros and cons of these architectures. Finally, the chapter highlights the proposed variants of Posit representation to make it more suitable for DNN hardware implementations.