Modeling neural networks on the MPP

A network of fixed-connection-weight neuronlike elements is simulated on the massively parallel processor (MPP) in two ways. First, the square connectivity matrix of a 128-neuron network is mapped onto the square MPP processor array. This allows a highly parallel simulation in which 128 MPP processors were active at all times. Second, a 128-by-128 array of neurons is mapped onto the 16.384 MPP processors. Here the MPP processor limits neuron connections somewhat, but all MPP processors are active at all times and a large speedup is obtained. The first simulation, based on mathematics (weight matrix), produces a significant speedup but tends to obscure the second faster simulation based on mapping the physics (entire physical description) of the neural network onto the MPP. The authors suggest that alternative mappings onto the MPP should be sought and examined carefully.<
>