Author: Robert Hyatt
Date: 11:22:29 02/18/02
Go up one level in this thread
On February 18, 2002 at 13:01:55, Uri Blass wrote: >On February 18, 2002 at 12:24:43, Robert Hyatt wrote: > >>On February 18, 2002 at 11:46:12, Aaron Gordon wrote: >> >>>Actually you can hit 1M nps in Crafty on a regular old AMD Thunderbird at >>>1.65GHz. That means you'd need ~330GHz to hit 200mnps. Of course if you were to >>>run a system such as this (say, 256 cpu's) in a cluster then you'd lose a heck >>>of a lot of NPS. Also lets say we used Myrinet and TTL_Papers and 'total' gained >>>a speedup of 128 x one cpu with 256 cpu's. I'm not sure how realistic that >>>number is but it 'seems' alright if you consider using an experienced cluster >>>designer along with good code. This will of course put you at 128mnps using the >>>1.65GHz tbirds. By the time you get something like this built there will be 2GHz >>>AMD Thoroughbred cpu's (0.13 micron AthlonXP's). >> >>Figure speedup = 1 + (N-1)*.7 for reasonable numbers of processors. I >>don't know that that will hold for N very large, say 128. >> >>Also, even if a CPU could run at 330ghz, we need a memory breakthrough or >>else it won't be 330 times faster than a 1ghz cpu today. >> >> >> >> >>> >>>If you figure 1 * 2 / 1.65 then that 2GHz XP would put you at 1.21212~Mnps. That >>>x 128 = 155.15Mnps. While not as fast as Deep Blue I think most of todays >>>programs should outplay Deep Blue with a little tuning (like cutting back on the >>>selectivity/pruning a bit). In the case of CT14 & Fritz7 actually running at >>>this sort of nps then most definately it will exceed Deep Blue strength. Perhaps >>>even around 50-80mnps. >> >> >>There is a _huge_ difference between what Deep Blue "knew" and what the two >>programs you mention "know". And knowledge is important against strong >>human players... > >Deeper blue had not enough time to tune their evaluation >I do not believe that the evaluation of it was better than the top programs of >today. > >The quality of knowledge is more important then the quantity. I _specifically_ mentioned the two programs he quoted. They are missing basic knowledge, such as drawn endings and so forth. Yes they play actively in the middlegame, but the point for playing actively is to catch your opponent in a short-sighted mistake. DB would likely not be guilty of such. And if you can't win in the middlegame, you had _better_ be good in the ending because they certainly were. And neither of the two programs mentioned impress me as being super endgame players. Very few do in fact... > >I agree that knowledge is important in comp-humans games > >Kasparov knew nothing about Deep blue. >Humans today know a lot about ct14 or F7 because they are commercial and humans >can play against them and learn their weaknesses. > >This knowledge is important in comp-human games. > >Uri It certainly is. And were I a GM and playing against DB, I would be studying anti-computer strategies if I had not tried them, under the assumption that DB might be tricked by one or more of them just like the other programs frequently are... I don't believe DB is unbeatable. I do suspect that if humans play it and use what I would call "normal chess" it might be nearly unbeatable. And we don't have enough evidence relating to how it might do against an anti-computer expert... So it will be a mystery for a long while...
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.