A simulation environment for very large neural networks

Computational requirements in signal processing may cover a wide spectrum, from binary operations to floating point. A task-oriented approach to solving such complex problems may prove costly if the application domain is not specifically restricted. The present approach consists in combining performance speed-up through parallelism with enough flexibility in an attempt to cover a wide field of application. The paper provides a brief overview of a reconfigurable multiprocessor system, called Shiva, and an example of how a computationally intensive problem such as large neural network simulation can be mapped onto it. This is illustrated by a speech synthesis application