Author: Robert Hyatt
Date: 05:51:29 08/02/01
Go up one level in this thread
On August 02, 2001 at 07:15:01, Vincent Diepeveen wrote: >On August 02, 2001 at 04:12:24, Dann Corbit wrote: > >>On August 02, 2001 at 04:03:25, Joshua Lee wrote: >> >>>A bunch of posts mention Hardware being Deep Blue's strength but can't you take >>>the program if you knew what the eval, parameters, etc and put it on a regular >>>computer? You wouldn't get the same nodes obviously , but it would play chess >>>the same way right? >> >>I think you will have to take an entirely different track, trying to emulate >>Deep Blue in software. > >>Because they had *LUDICROUS* compute power at their disposal, they could make > >They got actually 200M nps and with their simplistic eval which doesn't >even know doubled pawns for example, you get easily at a fast K7 now >like 1M nodes a second. Vincent, that is a stupid statement to make. (doesn't even know doubled pawns). Their eval, for example, knows more about bishops of opposite color and pawn endings than yours ever will. I listened to them talking to a GM about this. And what they were doing was _not_ simple. We played a game at an ACM event and we thought the game was drawn (Cray Blitz knew about opposite bishops, but only like crafty does... they are generally not winnable). Hsu said "this one is winnable." After the game, he, I, Gower and a GM (I don't remember who now, it was several years back) sat down and the GM explained why it was winnable and why deep thought was correct in its evaluation and we were not. That was just _one_ example. So to keep this "incredibly simply eval" nonsense alive is poor judgement. > >Fritz easily gets it and i can only agree with Frans statement that >from any perspective seen fritz is better as deep blue. There goes my opinion of you, now... :( > >Fritz gets 990k nps to 1.3M nps at my dual P3-800. > >Now if you simply do next: select engine parameters and set the nullmove >to 0 ply reduction, then fritz is not using nullmove. > >That factor of 6 you lose by not using hashtable last 6 plies i will >even not take into account here, but obviously if you compare >any program fullwidth versus using nullmove, then the comparision >can be done very easily. > >That deep blue was in hardware only meant it was harder to make, >its design started in 1993 or something? How about 1985? It played in the 1986 ACM event. > >Back then fastest computer was a 486. Getting a 12 ply fullwidth >search looks very good then. Nullmove was doubted by anyone, >so definitely you can compare with nowadays hardware, but take >into account the huge losses. > > - no hashtables > - fullwidth > - many extensions > - extensions just before qsearch (see IEEE99) > - a completely untested program I would hardly call it untested. It played hundreds of GM games in exhibition matches all over the world. I personally watched two such matches at conferences I was attending. > >That it crashed so little can only be a major compliment for >the hardware designer. I do not know much about hardware >techniques, but if you receive a chip in hardware just 2 weeks >before the event starts, i can imagine that those first tests >whether it worked a bit, were pretty nervous times. > >Compare that with a very well tested engine that are getting used >nowadays in software. > >hard to compare a nearly untested thing which is just 2 weeks before >event starts ready to use. > >Then if you find a bug you can at most disable its parameter, but you >cannot fix the pattern. > >>choices nobody else would ever dream of. All of the top PC programs agressively >>prune. They may not use null move, but something else like it instead. If you >>can hit a peak of one billion NPS, you can do things differently. > >Even in 1997 getting 12 ply was very good, considering that except some >programs which forward pruned bigtime or were completely preprocessor >based, were getting that depth by then. > >But by todays standards it makes no sense to say an old machine is that >well. > >It's like saying that a 1980 build computer from hundreds of >millions of dollars is going to outgun a dual 21264 from 20000 dollar. > >>It really does not make any sense to emulate deep blue on PC software. >>Unless you want to wait three months between moves. > >Comparing with the same kind of search with around 1.5M nodes a second, >you would need like 200M / 1.5M * 6 = 200 / 9 = 22x more time. > >I think a factor 22 is very realistic taking into account >hashtables kick butt also for the b.f.
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.