Mobile/embedded DNN and AI SoCs

摘要

Recently, Deep Neural Networks are changing not only the technology paradigm in electronics but also the society itself with Artificial Intelligence technologies. In this presentation, firstly, the status of AI and DNN SoCs will be reviewed from two perspectives; the data-center oriented and the mobile and embedded AIs. This dichotomy shows clearly the possible application areas for the emerging future AIs. Especially, mobile and embedded deep learning hardware will be introduced together with CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network). In addition, real CMOS chip implementation results of mobile/embedded DNNs will be explained with measurement results. Secondly, KAIST's approach integrating both sides of brain, right-brain for "approximation and adaptation hardware" and left-brain for "precise and programmable Von Neumann architecture", will be explained with novel design methodology. The deep neural networks and the specialized intelligent hardware (mimicking right brain) capable of statistical processing or learning and the multi-core processors (mimicking left brain) performing the precise computations including software AI are integrated on the same SoC. Based on this brain-mimicking SoCs, the object recognition and the augmented reality applications are implemented with low-power and high-performance for wearable devices such as smart glasses, autonomous vehicles, and intelligent robots.