The best computer chessplayer of 1990 (Deep Thought) has reached grandmaster strength. How strong a player can be built within five years, using today's techniques?
Deep Thought is a chess engine implemented in VLSI that
searches roughly 500,000 positions per second. The speed of
Chiptest-type engines can probably be increased by about a factor of 30
through design refinements and improvements in fabrication technology.
This factor comes from assuming a speed doubling every year, for five
years. Our own results imply that an additional factor of 250 speedup
due to coarse-grain parallelism is plausible. This is assuming
something like a 1000-processor machine with each processor being an
updated version of Deep Thought. This means that a machine capable of
searching 3.75 billion (
) positions per
second is not out of the question within five years.
Communication times will also need to be improved dramatically over the nCUBE-1 used. This will entail hardware specialization to the requirements of chess. How far communication speeds can be scaled and how well the algorithm can cope with proportionally slower communications are poorly understood issues.
The relationship between speed and playing strength is well-understood for ratings below 2500. A naive extrapolation of Thompson's results [Thompson:82a] indicates that a doubling in speed is worth about 40 rating points in the regime above 2500. Thus, this machine would have a rating somewhere near 3000, which certainly indicates world-class playing strength.
Of course nobody really knows how such a powerful computer would do against the best grandmasters. The program would have an extremely unbalanced style and might well be stymied by the very deep positional play of the world's best humans. We must not fall prey to the overconfidence which led top computer scientists to lose consecutive bets to David Levy!