Machine Architectures for Artificial Intelligence Computing,

The authors pointed out several issues that should be considered as part of an evaluation of potential AI machine architectures. They have then exemplified several activities of the advanced architecture community that focus on improving the run-time performance of AI systems. One can conclude that when the architecture matches the algorithm, the performance increase is substantial over that of von Neumann architectures. For example, when ASPRO is given an exact match forward chaining production system to execute, the performance increase is linear over von Neumann. If that same machine architecture were presented many procedure calls during execution, performance would degenerate to that of a sequential architecture. Real-time AI system designers should address the issue of matching their algorithms to machine architectures from the beginning of the system development process. The higher the real-time performance requirements become, the more critical this ongoing evaluation becomes.