Flosit: Float/Posit Coarchitecture Exploiting Value Location in Neural Network
Yuliang Chen, Ren-Xiang Liu, Tsung‐Chu Huang · 2024
Posit is the most significant universal number format for compansating for IEEE 754 in both high dynamic-range computing and high precision neural network. However, it suffers considerable overhead and time penalty. In this paper, we firstly exploited the value locality of computing data and develop the switching Float/Posit algorithm to save systematic computing time. Secondly, we develop the combined switching architecture, called the Flosit for selecting proper operations efficiently. Thirdly, the Flosit architecture is integrated to further reduce the area and power costs. From prelimianary simulations, the proposed Flosit can automatically have both advantages in speed of Floating-Point and high-dynamic range of Posit.