Posit Arithmetic and Its Applications in Deep Learning Computation

Posit is designed as an alternative to IEEE 754 floating-point format for many applications. It has non-uniformed number distribution, and it can provide a much larger dynamic range than IEEE floating-point format. These make posit especially suitable for deep learning applications. In recent years, more and more posit based deep learning hardware accelerators appear in the literature. In this lecture, the basics of posit format and the corresponding posit-based arithmetic units available in the literature, including adder, multiplier, multiply-accumulate unit, and quire operator, will be discussed. Then, several posit-based deep learning processors for deep learning inference and training will be discussed. Finally, the trends and challenges of posit arithmetic units and posit based deep learning processors will be discussed to motivate more related research works.