Simulation of artificial neural networks on the hierarchical network of hypercubes

Issues related to the suitability of message-passing architectures for simulating neural networks have been examined. Performance analyses of ring, mesh, binary tree, hypercube, hypernet, and extended hypercube architectures for simulating artificial neural networks have been carried out. The studies have revealed that the performance of the extended hypercube (a hierarchical interconnection network of hypercubes) is better than that of the ring, mesh, binary tree, hypernet and hypercube topologies. The performance of a neural network simulator implemented on the hierarchical network of hypercubes has been measured in terms of CUPS (connection updates per second) and was found to exceed the performance of the Connection Machine-2 and Intel´s WARP machine