A Comparison of Two Application-Specific Architectures for 2-d Mesh Computations

This paper considers the question of whether a mesh-connected machine is always better than a multi-pipelined machine for iterative 2-d mesh computations. Optimal throughput is determined as a function of a unified measure of resources (cost). The resulting performance curves for the two architectures show that there is a cost below which the pipelined architecture is an order of magnitude faster than the mesh, and above which this relationship is reversed. This methodology of comparing architectures using throughput-versus-cost modeling may prove useful in other contexts.