If complex problems are to be solved in reasonable computation times, then large scale parallel processing is necessary. For many of these problems, the density of the global communications dominates the performance of the parallel implementation. In these cases, the design of the interconnection network for the processors is known to play a significant part in the efficient implementation of problems on large T800 transputer systems. This paper presents a new genetic algorithm for generating optimal configurations, augmented by simulated annealing for selected refinement of difficult cases. These configurations have the further advantage that they satisfy the best known criteria for producing configurations that perform well on real applications. The paper concludes by describing the impact this might have on the design of future T9000 transputer configurations.