将神经网络控制器用于VLSI设计的方法研究  

A Method of Neural Network Controller Implementation in VLSI Design

作　　者：詹璨铭[1] 

机构地区：[1]中国电子科技集团公司第三十研究所,四川成都610041

出　　处：《微电子学与计算机》2015年第7期11-16,共6页Microelectronics & Computer

摘　　要：探索在超大规模集成电路中应用神经网络控制器的方法.根据Petri网理论,将库所与变迁组合成神经节点,节点通过输入触发信号链接组成复杂控制网络.定义两种类型神经节点,一种是节点组成串行分枝,另外一种用于同步并发分枝.通过两种节点组合,形成三种基本网络结构,三种结构再次组合又可形成任意复杂控制器结构.根据控制器分枝内串行、分枝间并行的特点,设计编译软件,输入更抽象的分枝描述代码,自动生成对应神经网络控制器逻辑电路描述代码.VLSI设计中使用神经网络控制器,能够更接近了寄存器传输级电路,以及更精确地描述电路,还能提高设计性能与可靠性.复制神经节点减小单节点负载,可优化电路时序;复制节点还可构成冗余缓解空间单粒子翻转.神经网络控制器可以处理各种异常情况,提高功能容错性和可维护性.这种方法已经用在产品开发流程中,实践证明能有效提高开发效率和质量.This article presents an approach to neural network implementation in VLSI, which is called as neural network controller based on Petri net. The structure of the neurons in the network is uniform; it is triggered by external events, and output place and transition signals of petri net. Two types of neurons are introduced, one is standing for serial process, and the other is used in synchronization of processes. The dual types of neurons are chained together by stimulate inputs, and compose the fabric. The controller is designed to conquer the side effects of state machine, and improves the performance and reliability. Typically, the controller is a precise description of the circuits. It is optimized to timing closure against constraints much easier than state machine. Reduplication of each node in neural network decrease single event upset （SEU）. Finally, the controller is easy to rebuild. The new design flow has applied in practice, and proved effectively.

关 键 词：超大规模集成电路 神经网络 状态机 PETRI网 单粒子翻转 

分 类 号：TN402[电子电信—微电子学与固体电子学]