Energy-Efficient Task Partitioning for CNN-based Object Detection in Heterogeneous Computing Environment

Abstract

Along with the high accuracy and the various use-cases of CNN, the number of services which are based on CNN continues to grow. Thanks to the development of GPU, a hardware accelerator for parallel processing, CNN have become powerful despite of its large amount of computations. In recent year, many studies have suggested using FPGA as a CNN accelerator due to its advantages over GPU. However, using these two accelerators together can greatly improve the processing performance because GPU and FPGA have complementary characteristics. Although there are some scheduling algorithms in the literature for the heterogeneous platform, they do not consider power efficiency and compliance with the deadline of an application at the same time. This paper found that the most power efficient accelerator is different for each sub-layers of CNN. It confirmed that task partitioning in the unit of sub-layers can improve the energy-efficiency of the system. Based on this finding this paper proposes an energy-efficient adaptive task partitioning scheme for CNN-based service. Experimental results show that the proposed scheduling consumes less energy than EDP method while satisfying the requested deadline of tasks.