Multiprocessor simulation of neural networks with NERV

Authors:
R. Manner;R. Horner;R. Hauser;A. Genthner
Affiliations:
Physics Institute, University of Heidelberg, Heidelberg, West Germany;Institute for Theoretical Physics, University of Heidelberg, Heidelberg, West Germany;Physics Institute, University of Heidelberg, Heidelberg, West Germany;Physics Institute, University of Heidelberg, Heidelberg, West Germany
Venue:
Proceedings of the 1989 ACM/IEEE conference on Supercomputing
Year:
1989

Citing 8
Cited 0

The connection machine

The connection machine
Neural network models of learning and adaptation

Physica D
Implementing neural network models on parallel computers

The Computer Journal
Computing with structured connectionist networks

Communications of the ACM
An introduction to computing with neural nets

ACM SIGARCH Computer Architecture News
An integrated system for neural network simulations

ACM SIGARCH Computer Architecture News
The ART of Adaptive Pattern Recognition by a Self-Organizing Neural Network

Computer
The Heidelberg POLYP multiprocessor project

Proceedings of the conference on CONPAR 88

Quantified Score

Hi-index	0.00

Visualization

Abstract

A general-purpose simulation system for neural networks is computationally very demanding. This paper presents some estimations of the computing power required, the necessary interconnection bandwidth, and the requisite memory size. Next, the hardware architecture of the NERV multiprocessor system is derived that fulfills these requirements. Up to 320 processors 68020 are used in a single VME crate together with a Macintosh II as a host computer. This set-up provides a computing power of 1300 MIPS together with a friendly graphical user interface. To support the simulation of arbitrarily interconnected networks and of asynchronous update models, the VME bus has been extended by a broadcast feature and a global max finder. The software architecture is outlined. It consists of system software, utilities, and application software.