(An) energy-efficient deep reinforcement learning processor with dual-mode weight compression and floating-point computing-in-memory

Abstract

The deep neural network (DNN) has shown impressive performance in various applications such as computer vision, and natural language processing. However, such a high performance of DNN comes at a high computational cost. For calculating DNN, a large size of weight and feature map should be stored in memory, and multiplied by each other after being loaded from memory. Recently, the portion of memory access for the total time and total power consumption has been increased in the DNN accelerators, due to the stagnant improvement of memory, which is known as the "memory wall" problem. Especially in the case of deep reinforcement learning, memory optimization is essential because it has mainly utilized multiple numbers of fully-connected layers simultaneously. In this paper, we propose a deep reinforcement learning accelerator that optimizes memory bandwidth and memory power consumption. The proposed accelerator optimizes memory bandwidth by dual-mode weight compression and optimizes memory power consumption with floating-point in-memory computing architecture.