一种FPGA集群轻量级深度学习计算架构设计及实现

Design and Implementation of Lightweight Deep Learning Computing Architecture for FPGA Cluster

传感器技术的发展带来了边缘、端设备功能的迅速迭代升级,也带来了战场前端的数据量成倍增长。针对边缘、端设备数据量的急剧增长和芯片计算处理能力的矛盾,结合Map/Reduce框架,提出了一种基于现场可编程门阵列(Field Programmable Gate Array,FPGA)计算集群资源的深度学习架构,能够实现多个深度学习算法的并行快捷部署和应用。该轻量级深度学习计算架构同时满足军事应用对“端”的智能处理能力提出的新要求,即不仅局限于数据采集和智能的应用,还必须具备分布式并行智能实时计算的能力。该FPGA集群轻量级深度学习计算框架部署不同类型算法容易,实时性高(ms级任务响应),可扩展性好,在多种类异构传感器、大场景大数据吞吐量的军事场景及森林防火等民用场景有广泛的应用前景。

With the development of sensor technology,the functions of edge or terminal equipment are rapidly upgraded,and the data quantity of front-end battlefield increases exponentially.According to the contradiction between the inability of chips and the sharp growth of data volume on edge and terminal equipment,combined with the Map/Reduce framework,a deep learning architecture based on field programmable gate array(FPGA)computing cluster resources is proposed,which can deploy multiple applications with deep learning algorithms and can be widely used in military scenes and civilian scenes such as forest fire prevention.