On the design of a multiprocessor architecture for neural network simulation

Criteria are outlined for sizing the memory in a parallel architecture for NN (neural networks) simulation. Although a particular topology and a certain learning algorithm are considered, the design guidelines illustrated can be seen as being generally valid. The performance parameters qualifying an NN simulator are mainly speed and its capacity to represent networks of extended size. It is shown that both parameters affect the field of application and the sizing of the memory of each processing element. Guidelines are presented for a trade-off between network size and the amount of memory necessary for each CPU, pointing out the implementational limits which are created when the system grows in size. >