From Floats To Posits: A Conversion Framework
Andrin Vuthaj, Elie Alhajjar · 2023
Modern Artificial Intelligence (AI) applications for mobile and embedded applications are shifting away from the IEEE floating point system due to its many inefficiencies. Recently, a new data type called posits was introduced as a potential replacement for IEEE-754 Standard floating point numbers (floats). Posits are a core component of the third generation universal numbers (type III unums). They provide compelling advantages over floats, including larger dynamic range, higher accuracy, better closure, bitwise identical results across systems, simpler hardware, and simpler exception handling. In this paper, we aim to provide a framework for an efficient conversion between the two systems. Namely, we devise a novel conversion algorithm that compares accuracy between representations in terms of computational error and highlights the superiority of the new type over the currently used number systems.