Author: Robert Hyatt
Date: 16:06:27 11/21/98
Go up one level in this thread
On November 21, 1998 at 15:15:01, Eugene Nalimov wrote: >On November 21, 1998 at 13:03:07, Robert Hyatt wrote: > >>On November 21, 1998 at 11:55:18, Amir Ban wrote: >> >>>On November 21, 1998 at 10:34:55, Robert Hyatt wrote: >>> >>>>On November 20, 1998 at 23:25:37, James Robertson wrote: >>>> >>> >>>>> >>>>>Deep blue was searching 250,000,000 nps right? Would it take roughly 160 seconds >>>>>(40,000,000,000 / 250,000,000) for Deep Blue to search the same number of nodes? >>>>> >>> >>>You have to divide the Deep Blue NPS by some factor due to the unavoidable loss >>>when doing parallel search. This factor depends on the number of processors and >>>the efficiency of the algorithm. I think we can safely assume a factor of 5, at >>>least. >> >>It's not that bad. Hsu reported roughly 30% loss... ie if he searches 250M >>nodes per second, that is something close to a single processor searching >>about 170M nodes per second. I see similar numbers in crafty between 1-8 >>processors and also saw similar (actually maybe a little better) numbers >>with Cray Blitz... > > >I beleive that Murray Campbell said that their "effective" node >rate is about 25-30% of "nominal" because of parallel overhead. >They have not 8 or 16 processors, but hundreds... > >Eugene > true, but they don't do a 256-way parallel search... the split into N threads on the SP, then each SP procssor splits into 16 threads for the 16 chess processors, effectively... And since the last iteration of the db chips didn't do hashing (Hsu said they could, but there wasn't time to get it working before the Kasparov match) there's not a lot of interaction between those 256 threads... only between the 16 groups of 16... I don't see where they would have any huge scaling problem... and notice that the 256M nodes per second is very conservative... each chip runs at 2.4M nodes per second, and there were either 256 or 512 of them in the last machine. I had heard 512, which should produce 1 billion nodes per second. That is likely where you got the 25-30% figure as that converts to 250 M nodes per second... >> >>> >>> >>>> >>>>just don't overlook the *huge* difference in the "shape" of the trees. The >>>>tree by dark thought is basically shallow and wide, when compared to deep >>>>blue. Because at this point Dark Thought has searched 20 plies, about >>>>*double" the depth of Deep Blue... Yet apparently DB went far deeper along >>>>the critical lines (seems singular extensions and other things they do work >>>>very well here)... >>>> >>> >>>It's true that Deep Blue did brute force (no forward pruning like null move), >>>but you expect a 20-ply null-move search to find what an (incomplete) 11-ply >>>full-width search supposedly found. >>> >>>One way to compare DB to other search engines is to consider the following: DB >>>gets fail-high on Qb6 after 1 sec., and this resolves after 5 sec. >>> >>>Amir >> >> >>I think their search is difficult to understand. IE I'll point back to the >>position I posted last year on r.g.c.c about the c5 move in a game against >>Cray Blitz, in Orlando at the 88 or 89 ACM event. They played c5 after >>failing high to +2.x, the game went *10* full moves further before *we* >>failed low to -2.x... I was looking right at their output and they had >>this incredibly long PV showing that the bishop was going to be lost. They >>saw it 20 full plies before we did. Lots of micros tried this position last >>year, and almost all would play c5 (as we expected that reply ourselves in >>the real game). But *none* had any clue that it was winning material.. even >>when they went far into the variation... >> >>The stuff they do with singular extensions and threats really shines in some >>positions... and probably costs them dearly in others... But they have the >>horsepower to pay the price when it doesn't work, and then they kill us when >>it does.. >> >>In Cape May (94 ACM) we ran with the full singular-extension algorithm >>enabled and promptly lost the first game we played, because we searched out >>the bottom of our 60 ply limit, and we didn't detect that...
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.