A number of physical systems (e.g., N body Newtonian, Coulombian or Lennard-Jones systems) can be described by N2 interaction terms. Completely connected neural networks are characterised by the same kind of connections: Each neuron sends signals to all the other neurons via synapses. The APE100/Quadricsmassive parallel architecture, with processing power in excess of 100 Gigaflops and a central memory of 8 Gigabytes seems to have processing power and memory adequate to simulate systems formed by more than 1 billion synapses or interaction terms. On the other hand the processing nodes of APE100/Quadrics are organised in a tridimensional cubic lattice; each processing node has a direct communication path only toward the first neighboring nodes. Here we describe a convenient way to map systems with global connectivity onto the first-neighbors connectivity of the APE100/Quadrics architecture. Some numeric criteria, which are useful for matching SIMD tridimensional architectures with globally connected simulations, are introduced.
Using Feed Forward Neural Network to Solve Eigenvalue Problems