通用可伸缩并行神经计算机系统NeuroC的设计和实现

THE DESIGN AND IMPLEMENTATION OF A SCALABLE GENERAL PURPOSE PARALLEL NEURO-COMPUTER NeuroC

ＮｅｕｒｏＣ是一个通用的并行神经网络计算机系统，它的规模是可以伸缩的．针对神经计算，系统设计了一套具有可选择广播通信能力的非等时访问共享存储器系统．本文首先对神经计算的需求进行了分析，接着讨论了计算单元的选择、存储器的组织与通信的实现，然后介绍了系统硬件的主要结构．文章中还简要地阐述了系统中软件的组成和结构，最后本文对ＮｅｕｒｏＣ的特点进行了总结．

This paper presents the architecture and the realization of NeuroC,which is a general purpose neuro-computing oriented parallel computer system that has one to four processor elements. It has a single memory addressing space in a distributed memory system which is called the nonuniform-memory-access (NUMA) shared memory system. First, the requirement of general neural network computing is discussed, and some statistics of neuro-computing is get. Then, the architecture of NeuroC and its characteristics is introduced. After that, a detail description of its realization is given. The main structure of the software is also introduced.