Hsu's chess chip: from ChipTest to Deep Blue

In June 1985 Carnegie Mellon graduate student Feng-hsiung Hsu set out to build a chess move generator on a single chip in the 3-micron technology available to universities. The first working chips arrived about ten months later; the chip produced up to 2 million moves a second, ten times the 64-chip module of the Hitech machine.

With the chip the program first searched about 50,000 positions a second; ChipTest played in the 1986 North American championship and, by AI Magazine, won it in 1987 with a perfect score. Deep Thought had two processors of 450,000 positions a second each, 700,000 together, and in November 1988, tying for first with Grandmaster Miles, beat Grandmaster Bent Larsen. The 1996 Deep Blue had 36 IBM RS/6000 SP nodes and 216 next-generation chess chips at 1.6-2 million positions a second each, 50-100 million overall. The 1997 machine had 30 nodes and 480 chips in 0.6-micron CMOS, about 1.5 million transistors each, 2-2.5 million positions a second each. By Hsu, sustained speed reached 200 million positions a second; by Campbell, Hoane and Hsu, in the 1997 match the average speed in searches longer than a minute was 126 million and the maximum sustained 330 million. The chip itself, by the abstracts of Hsu's two papers of 1987, at the International Solid-State Circuits Conference in New York (25-27 February) and in the IEEE Journal of Solid-State Circuits (October): 36,000 transistors on a die of 6.8 by 6.9 mm in 3-micron double-metal CMOS, dissipating under 0.5 W, built mainly as an 8 by 8 array of circuits of fewer than 550 transistors each, and the first chip of a two-chip set for a special-purpose chess computer. The abstracts do not name ChipTest; that this 3-micron chip, made through MOSIS, served in ChipTest, Deep Thought and Deep Thought 2 is Hsu's own account in a Computer History Museum oral history of 2005. What the record does not claim: anything from the body of the two 1987 papers, which were read only as abstracts.