Author: Vincent Diepeveen
Date: 05:35:36 12/25/05
Go up one level in this thread
On December 25, 2005 at 08:25:12, Vincent Diepeveen wrote: >On December 23, 2005 at 18:22:29, Jonas Cohonas wrote: > >>As far as know deep blue was based on an actual chess playing program (please >>correct me if am wrong) Dark Thought, does anyone recall the k/Ns for Dark >>Thought on a regular PC at that time? > >>I am asking this because we could maybe make an estimate of how Dark Thouhgt >>would perform k/Ns wise on PC's of today and get a better understanding of what >>the real speedup was on Deep Blue. > >>Maybe i am confusing the program Deep Blue was build up around and maybe it is >>not comparable (apples and oranges as Bob would put it), but if feaseble it >>would make for an interesting comparison/experiement. > >>Regards >>Jonas > >Hello Jonas, > >Dark Thought had an assembly port to alpha. The biggest improvements in >darkthought and optimizations were done by Peter Gillgasch, someone who really >understood chessprogramming and was state of the art chessprogrammer at the >time. > >In world champs 1997 darkthought searched at a 767Mhz alpha 21164. This machine >has 8KB L1 cache. Fastest PC at the time was a 300Mhz PII, which you could not >buy yet in a shop. Worldchamps 1997 was in october in Paris. > >Darkthought had a maximum of 1 million nodes a second. > >However please realize how restricted its search was last plies. > >To quote Peter Gillgasch: "it's so stupid that everything gives a cutoff, >which makes search very easy". Ferret was also there in 1997 at a 767Mhz alpha. > >It reached 14 ply in the game against me, which started as a complex middlegame >in sicilian (diep black side). > >In 1999 programs were improved a lot. Darkthought was still the same. Ernst A >Heinz had made a fight with Peter Gillgasch, though i do not know cirumstances >at all. Whatever happened there is perhaps interesting for insiders to know, but >what matters is this: Peter Gillgasch had left Darkthought, so no chance of it >ever improving. > >Deep Blue was a far simpler program. It has been made by hand in the hardware >logics. So it has not been made in verilog, but using the utmost lowlevel >components that are there in hardware. So you must make a program then by cut'n >pasting logical hardware blocks. This is the hardest way to program a machine. > >That Hsu got a chessprogram like that to work is really amazing. > >It didn't use nullmove and had a very limited evaluation function. Well you can >see that from the games and its logfiles of course. Deep Blue initially was a >simple piece square table program. Later on they took over a few Gnuchess 4.0 >components, but of course in hardware programming is so hard, that you must >expect gnuchess to have more knowledge than deep blue. > >Of course very important to realize is that in 1997 and before it was normal for >programs to be true beancounters. Later on chessprogrammers discovered that a >bit more agressive tuning of evaluation gives great play. The honour of that >invention really goes to Chrilly Donninger. > >It is very easy to figure out the search speed of Gnuchess 4.0 at todays >program, however gnuchess 4.0 is way more sophisticated than Deep Blue, yet it >has a few obvious bugs in easy eval terms. Gnuchess for example has nullmove, >deep blue doesn't have that. > >Just imagine how hard it would be to implement nullmove in hardware. > >Deep Blue didnt use killermoves in hardware. Imagine a 4 ply search in hardware >which already is terrible inefficient, then also not using killermoves. > >A single search of 4 ply in hardware without any sophisticated move ordering, >with a huge quiescencesearch and no nullmove and no killermoves, no history >moves, no move ordering at all except captures first, you can try yourself too. > >It is terrible inefficient, but good for your nodes a second. You can easily >search 4 million nodes a second like that at todays processors. There is no >bottleneck. No hashtable, nothing to stop search speed. > >But you will search just like Deep Blue around 10 ply first move out of book. > >Vincent Note that if you search like deep blue you will get that 10 ply only if you have 480 processors @ 2 million nodes a second with a terrible speedup. Note that Deep Blue severely restricted extensions and also forward pruned in hardware in order to get that 10-11 ply. There is a pdf document describing their extension system. Their dual threat mechanism is pretty clever taking todays search depths into account. Vincent
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.