Author: Vincent Diepeveen
Date: 11:00:09 05/06/01
Go up one level in this thread
On May 06, 2001 at 13:32:42, Roy Brunjes wrote: >On May 06, 2001 at 12:16:06, Laurence Chen wrote: > >>On May 06, 2001 at 11:37:02, Jorge Pichard wrote: >> >>>On May 06, 2001 at 10:52:39, Jorge Pichard wrote: >>> >>>>Has anybody done any tests using a dual motherboard vs an equivalent Single >>>>processor system? I would like to know if anybody has significant evidence as to >>>>whether Deep Fritz performs better using a Dual motherboard vs a Single >>>>Processors. Another intersting experiment would be to test Chess Tiger 14 using >>>>an AMD Athlon 1.3 GHz vs Deep Fritz using a dual motherboard 700 or 750 MHz. >>>> >>>>Please provide result if anybody has done experiment specially using DF. >>> >>> >>>>PS: My question should have been what dual Intel motherboard performs equal to >>>an Athlon 1.3 Ghz? Probably the best way to test it would be to use Deep Fritz, >>>but I have a wild guess that an Athlon 1.3 Ghz can outperform a dual 800 MHz. >>> >>>>Pichard. >>Why not measure equal CPU's, say, Dual 1 GHz against a single 1 GHz? If you >>think that it's an unfair test, than why should you think that a dual 800 MHz is >>equivalent to a 1.3 GHz system? > >For the simple reason that you do not get a linear scaling of performance with >multi-CPU systems running a single application. I.e. 800MHz x 2 does NOT give >1.6 GHz of performance to a chess program. The best scaling I have seen is >about 75% in moving from 1 to 2 CPUs. This would mean that you get 800 MHz of >performance from the first CPU, but only 800 * .75 or about 600 MHz of >performance equivalent from the second 800 MHz CPU. That would put the two CPU >system (each CPU at 800 MHz) at about 1.4 GHz of performance for this >hypothetical application. that's right there is no lineair scaling initially. In fact there is always overhead. however branching factor gets slightly better because if in some lines computer sees a need to split, then it might be that the second processor delivers a cheap cutoff. Now depending upon how you manage the tree you can get a good speedup. Of course if the overhead is more, for example because a program is faster in nodes a second, then it's tougher to get 2.0 Deep fritz gets about 1.9 at 1 hour a move I measured at fast levels, rapid about 1.4, very close to what i get in blitz. Diep scales sooner to 1.9 and then gets over it and gets to 2.0 soon. At 2 minutes a move in a game diep gets 2.0 easily, especially under linux, as that works parallel more efficient as long as i don't compare the nps i get there with what i get under windows. At faster levels however this isn't the case. At bullet i can be happy if diep gets better speedup as 1.2 actually. 1.4 might be seen at blitz, at rapid 1.6 to 1.7 and that slowly goes up to 2.0 for 2 to 3 minutes a move. this was a long and hard work however to get. Fritz curve for speedup is quite similar only it needs more time. Interesting to see is that after 5 minutes of search the total number of nodes needed to get to a certain depth is LESS when searching parallel as when searching single cpu. Of course the big overhead takes care it doesn't have faster search times then! I was happy to see this phenomena happening as i have it also with DIEP! Best regards, Vincent >But it is not even that easy. Each program will have its own scaling >characteristics, so it is not as easy to determine the performance of Chess >Program X in a multi-CPU environment by using a formula alone. The best method >is to test each program to see how much it benefits from the second (and third, >and fourth and ....) CPU. > >Roy
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.