Author: Robert Hyatt
Date: 21:06:36 04/22/98
Go up one level in this thread
On April 22, 1998 at 12:03:12, Carsten Kossendey wrote: >On April 22, 1998 at 11:20:17, Ernst A. Heinz wrote: > >>On April 22, 1998 at 10:47:27, Mark Young wrote: >> >>>I heard that IBM has come out with a 6 billion move a sec. version of >>>Deep Blue. Is this true? Is there going to be a rematch that I have not >>>heard about? If not why did they come out with a new version? >>> >>>Im in bed with the flu and have a high fever. So if I did not hear CNN >>>right forgive me. >>> >>> Mark Young >> >>Whoever reported this fell for a typically clever marketing gag of IBM >>... >> >>Today, IBM announced a new model of their RS/6000 SP parallel computer >>line with 332MHz CPUs. The announcement *claims* that these are up to >>5x times faster than the CPUs of the Deeper Blue machine and, together >>with the special chess processors, *would* result in a hypothetical >>chess machine calculating *up to* 1 billion = 10^9 moves/nodes per >>second. Below I quote an excerpt of the original announcement: >> >>"Company Press Release >> >> New IBM Technology Faster Than Deep Blue >> >> RS/6000 SP Capable of a Billion Moves Per Second >> >> SOMERS, N.Y.--(BUSINESS WIRE)--April 22, 1998-- Nearly one year after >> its historic victory over chess grand master Garry Kasparov, IBM today >> announced improvements to the RS/6000 SP that deliver five times the >> performance of Deep Blue. >> >> The improvement is the result of IBM's new 332 MHz microprocessor, the >> fastest chip available to date on the RS/6000 SP. If applied to the >> system that powered Deep Blue, this processor would increase its >> calculating power from 200 million to one billion chess moves >> per second." >> >>I personally do neither believe that the integer performance of new CPUs >>is 5x times higher than that of the ones used in 1997 -- maybe IBM >>intentionally >>compares the 332MHz CPUs to their 60MHz CPUs of the first match > >If my memory serves me right, DeeperBlue was using some sort of P2SC >nodes. While the P2SC provides exceptional floating point performance, >it stinks in integer performance. OTOH the P2SC based system have >usually much higher memory badwidth, so there seems to be a tradeoff >here. I'd guess the host part of EvenDeeperBlue would be faster by a >factor 3 or so now. > >>-- nor do I >>believe that their overall system performance scales linearly with the >>clock >>speed of the SP host CPUs. > >Not really, but since the host CPUs search the part of the tree near the >root they will see a speedup there, no doubt. > >>IMO, the "Deep Blue" part of the announcement is pure marketing hogwash >>... > >It is, they're not gonna have it take a bite at someone anywhen soon it >seems. > >>=Ernst= the only fly here is that if you search the "root part" faster, you overrun the chess processors. So if you can search a ply deeper on the SP hardware, you have two choices: (1) search one ply less on the chess procssors so they can take the positions as fast as you can feed them, otherwise they will become the bottleneck; (2) add more chess processors. Which would be easy. I'd bet the IBM announcement had something to do with "extrapolation" based on the g4 ppc chip, plus 8 or 16X the number of chess processors to keep up. It would be trivial for them to put that together, since the chess chips can be turned out in any quantity they want... whether they will do this or not is a matter for conjecture. I'd suspect not, at least until someone shows the "bean counters" how this is going to make even more money, publicity-wise.. but if that happens, bet your sweet bippy it emulate "Arnold"... "it'll be baacckkkkk." :)
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.