Abstract:
This paper proposes a low-power computing unit based on heterogeneous approximate structure, which is suitable for deployment in binary convolutional neural network (BCNN...Show MoreMetadata
Abstract:
This paper proposes a low-power computing unit based on heterogeneous approximate structure, which is suitable for deployment in binary convolutional neural network (BCNN). We propose a heterogeneous approximate structure based on approximate adders, which reduces the hardware resources consumption of BCNN deployment without loss of accuracy through error inter-compensation of different adders. The proposed low-power computing unit based on this heterogeneous structure can complete convolution operations in BCNN and greatly reduce power consumption. Implemented under 22nm ULL process, compared with the standard accurate computing unit, the proposed computing unit can reduce the power consumption by 33.92% and delay by 46.36%, while reducing the accuracy by less than 1% in GSCD.
Published in: 2022 IEEE 16th International Conference on Solid-State & Integrated Circuit Technology (ICSICT)
Date of Conference: 25-28 October 2022
Date Added to IEEE Xplore: 01 December 2022
ISBN Information: