Design of Unsigned Approximate Hybrid Dividers Based on Restoring Array and Logarithmic Dividers

Weiqiang Liu,Tao Xu,Jing Li,Chenghua Wang,Paolo Montuschi,Fabrizio Lombardi
DOI: https://doi.org/10.1109/tetc.2020.3022290
2022-01-01
IEEE Transactions on Emerging Topics in Computing
Abstract:Approximate computer arithmetic has been extensively studied due to its advantages to further reduce power consumption and increase performance at reduced accuracy. Although a number of approximate adders and multipliers have been studied, only a few approximate dividers have been proposed. A logarithmic divider (LD) has low complexity and accuracy. while an exact array divider (EXD) has a high complexity. Therefore, in this article, an approximate hybrid divider (AXHD) is proposed. It takes advantage of both LD and EXD to achieve a tradeoff between hardware performance and accuracy. Exact restoring divider cells are used to generate the most significant bits (MSBs) of the quotient for attaining a high accuracy while the other quotient digits are generated by using a LD as an approximate scheme to improve figures of merit such as power consumption. area and delay. To further save hardware resources, a so-called eliminated approximate hybrid divider (E-AXHD) based on AXHD is also proposed. In this improved design, a reduced width divider is used to replace the EXD in AXHD. Specifically, for a 16-by-8 design, n/(n+1) array division is used to replace the n/8 array division (n < 8). The proposed AXHD and E-AXHD are evaluated and analyzed using error and hardware metrics. The proposed designs are also compared with EXD, LD and previous approximate dividers. The results show that the proposed designs outperform previous approximate dividers by considering both energy and error. The proposed hybrid dividers are of particular interest for error tolerant applications such as image processing and machine learning.
What problem does this paper attempt to address?