Performance of parallel neural network simulations

Journal Article (Academic article)

A general constraint satisfaction neural network simulator is implemented on two different parallel machines, the BBN Butterfly and the INMOS Transputer. The general constraint satisfaction algorithm is partitioned similarly for both implementations. The simulator on the Butterfly is written in C under the Uniform System programming environment and the MACH operating system. The simulator on the Transputer is written in Occam 2 under the Transputer Development System programming environment. The performance of each simulator is evaluated with various sizes of the neural network on different numbers of processors for each machine. The communication overhead of neural network simulation is large enough that both implementations show signs of reaching the speedup limit even for modest numbers of processors. The T800 Transputer is a much faster computing engine than a Butterfly processor; therefore, it begins to reach its speedup limit much faster

Duke Authors

Cited Authors

  • Board, ; A, J; Jr, ; Lu, S-J; J,

Published Date

  • 1990

Published In

  • Transputer Research and Applications 2. Natug 2 Proceedings of the North American Transputer Users Group

Start / End Page

  • 185 - 200

Conference Location

  • Durham, NC, USA