NON-VON's Applicability to Three AI Task Area 1 David Elliot Shaw Department of Computer Science Columbia University ABSTRACT NON-VON is a massively parallel machine constructed using custom VLSI chips, each containing a number of simple processing elements A preliminary prototype is now operational at Columbia University The machine is intended to provide highly efficient support for a wide range of artificial intelligence and other symbolic applications This paper briefly describes the current version of the NON-VON machine and presents evidence for its applicability to the execution of OPS5 production systems, a number of low- and intermediate-level computer vision tasks, and certain "difficult" relational algebraic operations relevant to knowledge base management Analytic and simulation results are presented for a number of algorithms The data suggest that NON-VON could provide a performance improvement of as much as two to three orders of magnitude over a conventional sequential machine for a wide range of AI tasks 1 Introduction A strong case can be made that the most formidable challenges now facing artificial intelligence researchers have little to do with problems of computational efficiency Indeed, the development of high-performance hardware is clearly not a sufficient condition for the implementation of systems having human-like cognitive capacities A number of researchers now believe, however, that as progress is made in understanding and mechanizing intelligent behavior, the availability of such hardware may ultimately prove to be a necessary condition in the case of certain AI tasks of practical importance Several researchers [Fahlman, 1980, Shaw, 1980, Hilis, 1981; Stolfo and Shaw, 1982, Moto-Oka and Fuchi, 1983; Deenng, 1984] have proposed machines 1 This research was supported in part by the Defense Advanced Research Projects Agency, under contract N000390-84-C-0165 bv the New York State Center for Advanced Technology in Computers and Information Systems at Columbia University, and in part by an IBM Faculty Development Award. intended to accelerate various operations relevant to one or more AI applications The problem is complicated, however, by the fact that many integrated AI systems will ultimately require the performance of a number of different computationally intensive operations, ranging from the various signal processing algorithms employed in low-level computer vision and speech understanding through a wide range of inferencing and knowledge base management tasks that may be involved in high-level reasoning The central goal of the NON-VON project is the investigation of massively parallel computer architectures capable of providing significant performance and cost/performance improvements by comparison with conventional von Neumann machines in a wide range of artificial intelligence and other symbolic applications In this paper, we provide some evidence of NON-VON's range of applicability within the space of AI problems by briefly describing certain algorithms, asymptotic results, analytical performance projections, and instruction-level simulation data involving three distinct AI task areas rule-based inferencing, computer vision, and knowledge base management The following section provides a brief overview of the NON-VON machine family In Section 3, we report the results of a detailed projection of the machine's performance in executing OPS5 production systems, based on data derived from Gupta and Forgy's investigation of six existing production systems at Cargnegie-Mellon University [Gupta and Forgy, 1983, Gupta, 1984] In Section 4, we review a number of low- and intermediate-level computer vision algorithms for the NON-VON machine that have been developed and simulated at the functional and instruction levels. Section 5 presents the results of a projection of NON-VON's performance on certain AI- relevant database management benchmarks formulated by Hawthorn and DeWitt [1982], using the analytical techniques employed by those researchers The final section summarizes our results and attempts to characterize the essential architectural and algorithmic principles which would appear to be responsible for NON-VON's strong performance in these superficially dissimilar AI application areas