Author: Vincent Diepeveen
Date: 08:59:24 10/01/03
Go up one level in this thread
On October 01, 2003 at 11:49:20, Drexel,Michael wrote: >On October 01, 2003 at 11:25:26, Vincent Diepeveen wrote: > >>On October 01, 2003 at 07:44:03, Peter Skinner wrote: >> >>>On October 01, 2003 at 03:11:15, Gian-Carlo Pascutto wrote: >>> >>>>On September 30, 2003 at 19:44:02, Peter Skinner wrote: >>>> >>>>>I think those would be good. >>>> >>>>I'd scratch all single-CPU programs. Not having SMP support is likely >>>>going to be a major problem in a tournament where the average number >>>>of cpus per program will be way over 2. :) >>> >>>Bigger doesn't always mean better. While I am sure Vincent will do well on 500 >>>processors, but I don't think he is a lock to win because of speed. >>> >>>Shredder proved in the past a single processor system can beat the >>>multi-processor systems. I see it all the time online. I use a single processor >>>system online and I like longer time controls, usually greater than 90 mins per >>>side and I win a great number of the games. >> >>Do next compare online. >> >>Get pocketfritz at a 50Mhz. >> >>Now get something 500 times faster >>ugh, 25Ghz and play with that online against pocketfritz. > >Your math is interesting: > >500 500Mhz processors are 500 times faster than a 3 Ghz single processor >system? :) The problem is that people calculate the wrong way. The equal of this 250Ghz supercomputer is not giving more time to a 2ghz opteron single cpu (P4 a lot slower than opteron/k7, so forget 3Ghz P4). The problem is your PC does not have 512GB ram. When loading factor > 0.5 a 2 times bigger hashtable for DIEP already saves 20% in node count. Trivially at the pc it is not possible with a high node count like fritz or junior gets to store the last few plies to hashtable. It is 400 ns to store something. another 400ns to get it out of hashtable and so on 800ns. At your 3Ghz that's 2400 cycles and as you might know they need LESS than that for 1 node. Also chance you get a cutoff is like 3-4% or so, so it is not worth it at all. So practical you speak about a probe or 100k a second which they can do to the hashtable. I'm doing millions of probes at the same time to the hashtable with DIEP. At a supercomputer you can do that all in parallel you know... ...128 different mainboards... So where there is a parallel loss of course thanks to algorithms, there is a major win back. Note this was already written down in the journal of ICCA by Rainer Feldmann. The difference is however that his program with simple evaluation would get 500k nps at this machine, missing mate in 7, using 500 processors and i get a lot more with better evaluation. Best regards, Vincent >Michael > >> >>Oh well even that is not a good compare, pc's too slow for this compare. >> >>>I didn't pick your program to place higher than Brutus because of processors. I >>>picked it due to the strength in longer games. I feel it is one of the best at >>>tournament time controls. >>> >>>Peter.
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.