Author: Uri Blass
Date: 08:36:07 06/30/04
Go up one level in this thread
On June 30, 2004 at 09:13:34, Anthony Cozzie wrote: >On June 29, 2004 at 18:11:21, Robert Hyatt wrote: > >>On June 29, 2004 at 18:06:27, Vincent Diepeveen wrote: >> >>>On June 29, 2004 at 17:57:07, Djordje Vidanovic wrote: >>> >>>>On June 29, 2004 at 15:50:00, Vincent Diepeveen wrote: >>>> >>>>>On June 29, 2004 at 12:48:06, Robert Hyatt wrote: >>>>> >>>>>>On June 29, 2004 at 12:28:03, Vincent Diepeveen wrote: >>>>>> >>>>>>>On June 29, 2004 at 11:56:05, Robert Hyatt wrote: >>>>>>> >>>>>>>>On June 29, 2004 at 11:25:08, Jaime Benito de Valle Ruiz wrote: >>>>>>>> >>>>>>>>>On June 29, 2004 at 08:26:15, Zach Wegner wrote: >>>>>>>>> >>>>>>>>>>One important point is that crafty uses bitboards, so it will have an additional >>>>>>>>>>speedup on a 64 bit processor. >>>>>>>>> >>>>>>>>>I might be wrong, but I think that Fritz also uses bitboards (I don't know about >>>>>>>>>Shredder). >>>>>>>>>Anyway, what about the compiler? And what extra efficiency do you expect from a >>>>>>>>>64-bit processor? Enough to outsearch all other engines running on similar >>>>>>>>>processors? >>>>>>>>> >>>>>>>>>Jaime >>>>>>>> >>>>>>>> >>>>>>>>AMD has tested Crafty to answer this question. They compiled it for 32 bits, >>>>>>>>and for 64 bits. The 64 bit version runs 47% faster than the 32 bit version, >>>>>>>>everything else remaining constant. >>>>>>> >>>>>>>Going from 32 to 64 bits also the number of registers moves up from 8 to 16 >>>>>>>which is a big speedup also for Crafty and is inside that 47%. >>>>>> >>>>>> >>>>>>Does not change their statement or measurement... >>>>>> >>>>>>And the question will be whether or not the commercial guys use a beta compiler >>>>>>to produce accesses to those extra registers. If, as you always claim, fritz is >>>>>>in ASM (rather than what Frans claims, that it was rewritten into pure C a >>>>>>couple of years back) then it won't be able to touch those registers either... >>>>> >>>>>Fritz will be opteron assembly, read my lips... >>>>> >>>>>Please show me a statement from Frans where he quotes it is written entirely in >>>>>C. I cannot remember that at all. They (reporters) just ask whether 'fritz' is >>>>>written in C. The answer is probably 'yes' to that (interface). >>>>> >>>>>>Time will tell... >>>>>>>>Fritz does not use bitboards by the way... >>>> >>>> >>>> >>>>Fritz will hit 25 million kn/s, reaching ply 17-18 in the middlegame... :-) And >>>>it will be Opteron ASM, by the way :-) >>> >>>I'm not so sure about speed. Commercially they will say it's C++ when that is in >>>their interest to say, they will say it's C when that's in their interest and it >>>will be assembly when it is in their interest to say so. >>> >>>But the *engine* is in assembly, let's be clear there. >>> >>>I would expect about 15 million nps, except that last years Frans has put more >>>knowledge into evaluation. I remember he said already around 1999 he had an >>>evaluation exactly 2 times bigger than crafty, so i guess that'll be at least 4 >>>times bigger by now. >> >>2x bigger doesn't mean much when comparing bitboard to non-bitboard. >> >>Not that I believe any of that anyway. >> >>At the moment, recent benchmarks suggest Crafty is significantly faster than >>Fritz on equal hardware which suggests that he is doing something different than >>in the past, as 3 years ago he was 2x-3x faster than me in raw NPS. >> >>But it doesn't matter. We play what we have, and see what happens... >> >> >>> >>>That will slow it down. >>> >>>Also it's doing seemingly checks in qsearch which slows it down. >>> >>>I expect somewhere around 9-10 million nps or so. That's about 1000 cycles a >>>node. >> >>That is dead on my quad 2.4ghz opteron speed. I bet fritz is closer to 6M based >>on recent speed reports here... > >From http://www.beepworld.de/members39/computerschach2/chessmarks.htm: > >Athlon64 2.2 3400 1405 kn/s Gian-Carlo Pascutto >Ahtlon64 3000+ 1195 KN/s Joachim Rang (2 Ghz, 512 Kb Cache) > >This means that 1 processor 32-bit Fritz should get about 1250 KN/s. Assuming >perfect nps scaling gives 5M nps. However, I'm sure Frans has been busy >rewriting to AMD64 assembly; the big question is how much that gives him. > >anthony I have no idea what Frans was doing but I think that at the level of chess programs it is not a good idea to waste a lot of time on optimizations. maybe it is good if you care only about blitz but if you care about long time control then if they need to choose between spending time on optimizations and spending time on implementing better pruning and better extensions than the second possibility seems to be better because I do not think that chess programs are close to implement the best algorithm. Uri
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.