(An) energy-efficient deep reinforcement learning SoC for mobile platform

Abstract

Recently, deep neural network (DNN) has been actively researched from simple recognition tasks to precise control for robot or autonomous systems, which are treated as a task that only human can do. Unlike recognition tasks, the real-time operation is essential in action control, and it is too slow to use remote learning on a server communicating through a network. New learning techniques, such as reinforcement learning (RL), are needed to determine and select the correct robot behavior locally. In this paper, we propose a low power deep reinforcement learning (DRL) SoC, supporting CNN and learning-optimized RNNs. The adaptive reusability of weights and inputs, and data encoding/decoding techniques reduces power consumption and peak memory bandwidth of DRL processing by 31% and 41%, respectively. The 65nm 16mm2 chip achieves a peak 2.16TFLOPS/W at 0.73V and 204 GFLOPS at 1.1V with 16b data.