The Next Generation of Chess Playing Machines is Introduced by IBM
By Tets Maniwa
San Jose, CA--March 3, 1997--IBM (San Jose, CA) displayed the next generation of chess playing computer at the Association for Computing Machines conference, ACM97. IBM calls the latest version of their chess playing machine Deep Blue Junior (DBjr.). DBjr runs on a single RS/6000 with 16 special chess accelerator chips installed in the machine. It evaluates ten million chess positions per second and has some built-in learning capabilities to improve the analysis algorithms with time and experience.
DBjr will play a match with Chess Grand Master Garry Kasparov from May 3 through May 11 this year in New York. Kasparov has intuition and flexibility on his side, while DBjr has tactical accuracy, machine endurance, and a total detachment from distractions. Look for this to be a close match.
Deep Blue, the previous generation of chess machine, won one game and tied two other games before losing the match with Kasparov last year. It is a 32-node RS/6000 SP with eight of the chess ASICs per node, for a total of 256 accelerators. The Deep Blue machine evaluated up to 200 million chess positions per second in a parallel processing mode. The chess ASIC is a specialized coprocessor with three main sections: feature extraction, move generator, and move evaluation section.
IBM is using the game of chess as an end application for their research into enhanced computer through-put. Dr. Murray Campbell, research scientist at the IBM T. J. Watson Research Center in Yorktown Heights, NY views the work on the chess program as applied research into high-performance special-purpose computers. By combining special purpose hardware with new software, they are addressing classes of regular, parallelizable problems such as molecular algorithms for pharmaceuticals, data mining of large databases, and detailed financial models.
IBM achieved order of magnitude increases in throughput without incurring too much processing overhead when adding the additional coprocessors. In designing DBjr, they went from eight coprocessors per node in Deep Blue, to 16 processors per node, with almost linear increases in processing capability.
The knowledge gained in the programming and hardware interfaces to multiple coprocessors helps to address the need for significant increases in computing to solve some of the more difficult problems. The increased compute capabilities will enable the user community to address difficult tasks in a completely different manner than currently possible.
Return to Headlines