Chapter 5 - Compute Models and System Architectures

Abstract

This chapter reviews the models and organizational styles for large-scale, highly parallel computing resources and emphasizes how they can be used in the organization of reconfigurable computers. The restrictions and stylizations in some compute models may limit the freedom in expressing an algorithm. However, limiting expressive freedom that would lead to poor reconfigurable solutions is one of the ways that a good compute model provides assistance and guidance. If the limitations make a solution hard to express, that can be a good indication that the solution approach is not well suited to a reconfigurable platform or that the compute model is not the natural choice for the task. A good compute model helps in decomposing a problem into components that can be designed and validated independently. This helps in avoiding the quadratic explosion in complexity arising from potentially interacting resources and avoids the linear cost necessary to program each resource independently. One of the most important tools provided by each compute model is a way to reason about correctness, which ultimately facilitates scaling, implementation adaptation, and optimization because it defines what transformations are possible without impacting correctness. As the compute model helps in understanding the natural composition and parallelism in the application, the system architecture deals with how to organize the implementation. The choice of architecture depends on technology costs and resource availability compared to the application resource and performance requirements.