lindsay@MATHOM.GANDALF.CS.CMU.EDU (Donald Lindsay) (07/14/90)
In article <2555@acorn.co.uk> RWilson@acorn.co.uk writes: >In recent article cdshaw@cs.UAlberta.CA (Chris Shaw) writes: >>... the major reason Deep Thought >>beats excellent chess people is because of its specialized hardware. The >>program operates by a well-known brute force algorithm. My main point is that >>Deep Thought would be useless if the program relied on some set of greasy >>assembler tweaks on some two-bit general purpose CPU. > [...] >Which machine beat Deep Thought? I'm glad you asked... a 68030 coded in >assembler. > >General purpose computing is improving more rapidly than special purpose >hardware, even of the wackier types. No, they're on the same curve, because both benefit equally from improvements in fab technology and in design tools. The general purpose systems are usually further along that curve, because companies put more designers on their flagship items, and put fab improvements there first. Notice also that small production runs are a poor idea if a new fab process needs months of tweaking. (Small production runs are a great idea if one is tweaking a new design tool.) As for chess machines in particular ... I've held a Deep Thought board in one hand. It's a dead ordinary VME board, containing two independent processors. The current system uses a few boards, each containing microcode and FPLA code that makes greasy assembler look straightforward. But then, it was designed by some very good people, who saw that as the "high level" alternative to full-custom VLSI (which they also did). I would point out that at the Master level, chess ratings are not linear in the machine power. The actual truth is currently being argued about. Dr. Berliner's HiTech has picked up quite a few rating points in the last year, strictly through AI (algorithm) improvements (mostly endgame stuff that runs on the host - I think). The designers of Deep Thought argue differently, and their current multiprocessor is rated a GrandMaster, rather higher than the 68030 mentioned above, and in fact higher than all but a few hundred humans. The designers argue that a ply (move/response) is worth 200 to 250 rating points, and that each extra ply takes 4x to 6x the speed. That is, they argue that ratings go as the log of the machine power. They also argue that they would beat Kasparov if they could go 12 plies instead of their current ten plies. They are going to get their chance: they have been funded to build a seriously parallel machine, with on the order of 1,000 move processors. They hope for a parallel speedup on the order of N/2 to N/5, and a processor speedup - perhaps to 3-5 M nodes/sec. It should be a very interesting decade. -- Don D.C.Lindsay