Hypercube algorithms for neural network simulation: the Crystal_Accumulator and the Crystal_Router

Abstract

We discuss communication algorithms relevant for neural network modeling on distributed memory concurrent computers with a hypercube topology. Full, intermediate (medium range) and sparse network connectivities are analyzed. We point out that the flexible hypercube topology allows for the efficient implementation of the broad class of network algorithms with variety of connectivity patterns. We find algorithms index, crystal_router, fold and crystal_ accumulator which are appropriate in different cases depending on the sparseness and regularity of the network interconnect. We show that there are natural analogies between biological communication paths and the interconnects of a distributed memory concurrent computer.