Light-Weight Design and Implementation of Deep Learning Accelerator for Mobile Systems

Deep neural network (DNN) has been widely used in various fields along with the great attention of artificial intelligence (AI). Accordingly, the study and development of deep learning accelerators have been actively performed for working the DNN fast and efficiently. The accelerators usually consume high computing power and require large memory footprint. However, low power and memory-efficient accelerators with light-weight design are essential for mobile systems due to the limitations of hardware resources. Mobile system is necessary, in particular, in an era when internet of things (IoT) becomes important. Therefore, we proposed and implemented the light-weight design of deep learning accelerator for mobile systems. Implemented results prove that the proposed design can be sufficiently embedded in mobile systems with over 50% smaller memory size than existing accelerator design for mobile deep learning applications.