神经网络并行MIMD处理器的研究及实现
Realization of a Neural Network Parallel MIMD Processor
-
摘要: 为了能高速地实现多种神经网络,拓展神经网络在工业控制中的实时性、嵌入式应用,设计了一种多指令多数据流(MIMD)的通用型神经网络处理器(APP)。处理器的处理单元组之间、处理单元组与乘累加协处理器之间均可以并行执行任务、处理单元组与其他存储器之间可以并行通信。在FPGA上仿真验证了处理器的功能,并实现了用于轧辊偏心在线控制的BP网络和用于字符识别的Hopfield网络等两种不同的拓扑结构。实验数据表明,该体系结构具有较高的并行性,其性能优于其他常见的通用型实现手段。Abstract: To expand the realtime and embedded application of the neural network in industry control, general purpose neural netowrk processor (APP) based on multiple instruction stream and multiple data stream (MIMD) is paoposed. The tasks among PEGs, PEG, and MAC coprocessor can be processed in parallel, and also communications among PEGs and other memorys can be carried out in paralled. The processor has been simulated on FPGA and is used to implement two different neural network:the BP network used in roll eccentricity online control and the Hopfield network used in character recogonition. The simulation result shows that the performance of the proposed APP is better than that of other general methods for neural network implementition.