A reconfigurable multi-precision quantization-aware nonlinear activation function hardware module for DNNs

Open Webpage

Qi Hong, Zhiming Liu, Qiang Long, Hao Tong, Tianxu Zhang, Xiaowen Zhu, Yunong Zhao, Hua Ru, Yuxing Zha, Ziyuan Zhou, Jiashun Wu, Hongtao Tan, Weiqiang Hong, Yaohua Xu, Xiaohui Guo

Published: 01 Jan 2024, Last Modified: 05 Nov 2025Microelectron. J. 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•The mathematical properties of NAFs were then exploited to significantly reduce the hardware storage requirements using symmetric multiplexing techniques, achieving a 50 % reduction in storage cost with negligible error.•Three preset precision configurations were introduced, complemented by a reconfigurable data path and a universal address generator, enhancing the hardware architecture's flexibility.•A customized floating-point format quantization with an error fine-tuning algorithm, has been proposed to maximize accuracy stability in computations. This approach maintains the flexibility and cost-effectiveness of the hardware implementation.