Empowering edge mining on smartphones with reconfigurable fabrics

With the prevalence of sensors in smartphones, a large volume of time series is being produced, which contains useful information about personal lifestyles and habitats. Considering privacy and security, the mining of such data is best done on the edge. However, the tight energy constraint imposed by batteries makes this task challenging. Recently many studies focus on edge mining of time series on low-power co-processors for high energy efficiency. However, none of them has explored the opportunity of reconfigurable fabrics, which are well-known energy-efficient and low-cost solutions. In this paper, we propose an efficient smartphone architecture to empower edge mining on smartphones. A configuration lib is associated with the fabrics which can be configured according to the current bottleneck of edge mining tasks. Two widely-adopted algorithms, artificial neural networks (ANNs) and Dynamic Time Warping (DTW) for edge mining, are evaluated. The proposed architecture is implemented on the off-the-shelf smartphones and low-power FPGAs. Experimental results show that compared with smartphones, the proposed architecture can achieve a speedup of 7.8x-21.1x and an energy efficiency improvement of 10.6x-27.1x.