MCM划分的自组织神经网络  

A SELF-ORGANIZING NEURAL NETWORK FOR PARTITIONING ON MCM

作　　者：胡卫明[1] 徐俊华 何志钧[1] 严晓浪[2] 

机构地区：[1]浙江大学计算机科学与工程系,杭州310027 [2]杭州电子工业学院CAD研究所,杭州310027

出　　处：《计算机学报》1998年第7期642-649,共8页Chinese Journal of Computers

摘　　要：本文在提出一个直接和间接相联模块间相似性的表示方法的基础上，提出了一个基于自组织神经网络的性能驱动MCM划分的神经学习方法．算法求解如何在高层设计中将功能模块分配到MCM芯片中．算法不仅考虑了模块间的相似关系，还考虑了MCM的版图结构；具有芯片间连线数目最少和时钟周期最短双重优化目标；能使连线尽量产生在相邻近的芯片之间；能满足时延、散热和面积约束．文中还提出了一个层次神经网络模型和面积约束下的MCM划分的层次神经学习方法．本文的神经网络模型结构合理，学习速度快．In this paper,a method of expression for similarity relationship between blocks connected directly or indirectly is proposed, then a new approach, based on the self-organizing feature mapping neural network, for partitioning on MCM, is presented. The algorithm considers the problem of assigning functional blocks into slots on Multi-Chip Modules during high level design in order to have fast feedback on the impact of high level design decisions. The method is considering both similarity relationship between blocks and physical structure of MCM; having double optimal object, minimizing number of connections between chips and cycle time of system; making connections occur between chips as closer as possible; satisfying area constraints and thermal constraints. In addition,on basis of the hierarchical neural network model,this paper proposes a hierarchical neural learning algorithm for partitioning on MCM under capacity constraints. The neural network has reasonable structure and rapid training speed.

关 键 词：MCM划分 自组织神经网络 多芯片组件 

分 类 号：TN42[电子电信—微电子学与固体电子学] TP18[自动化与计算机技术—控制理论与控制工程]