一种神经网络并行处理器的体系结构

An Architecture for Neural Network Parallel Processor

神经网络处理系统所能实现神经网络模型的种类越多其通用性越好,应用范围就越广泛.提出了一种神经网络并行处理器的体系结构,能以较高的并行度实现典型的前馈网络-BP网络和典型的反馈网络-Hopfield网络的算法.该处理器以SIMD(Single Instruction Multiple Data)为主要计算结构,并结合这两种网络算法的特点设计了一维脉动阵列和全联通的互连网络,能够方便灵活地实现处理单元之间的数据共享.实验结果表明该体系结构有效地提高了神经网络的运行速度.

If a neural network processing system can realize more kinds of neural network, its commonality is much better and it can be used in more application fields. We propose an architecture for neural network parallel processor to broaden its application fields ,which can realize the BP neural network and the Hopfield neural network with a higher parallelism. The processor is based on SIMD（Single Instruction Multiple Data）architecture. Combining the characters of the two algorithms,we design systolic array and a full interconnection net to realize the data sharing more easily and more flexibly. The experiment result shows that the runtime of the neural network is enhanced effectively.