Development of DNN Accelerator and Its Application in Avionics System

Future avionics systems must have high-performance intelligent data processing capability, especially for various intelligent application scenarios, such as image/radar target recognition, big data analysis, control and decision. Deep neural network (DNN) is a widely used AI algorithm due to its accurate fitness of nonlinear functions. It can introduce huge computation and memory costs as the increase of network depth. Recently, DNN hardware accelerator is proposed as a domain-specific accelerating platform. It can provide a flexible and reconfigurable operating environment for different algorithms, including convolutional neural network and recurrent neural network, with the advantages of high performance, low latency, and low power cost. The key technologies of DNN accelerators mainly include: parallel computing architecture, intelligent processing elements design, efficient memory structure, and dataflow scheduling optimization. This paper studies the design methodology of DNN accelerators, summarizes the key technologies, and discusses the prospective applications in avionics computing systems. This paper will provide some technique supports for the design of next-generation intelligent airborne computing systems.