A neuromorphic GaN system for intelligent computing on edge

Abstract

Recently, artificial intelligence has gained great success in moving toward to edge devices. However, the involved deep neural network (DNN) computations which extremely resource-intensive and power-hungry became a big challenge to the edge computing. The programmable ReRAM based neuromorphic engine opened an opportunity for efficient DNN computing; however, memory utilization and communication latency in the training process implementation is still significant in the current research, and as yet, unmet challenges. The problem becomes more severe in DNN with complex training process such as the generative adversarial network (GAN) which is widely adopted in the intelligence computing on edge. In this work, we target to solve these challenges by designing an efficient GAN computing system on ReRAM neuromorphic engine with a online training framework and an optimized backward computation and a cross-parallel computation flow to execute the training process efficiently. The system performance is evaluated and compared with traditional GPU accelerator, and our results show 2.8× speedup and 6.1× energy-saving.