Author: Ed Trice
Date: 19:55:43 01/12/04
Go up one level in this thread
Hello Paul, >Last month's perft activity reminded me of some related code I started on ages >ago and never finished, so over xmas and new years I completed it and decided to >try to confirm the perft 11 result. > >The program just completed, after a little over 11 days, obtaining perft 11 of >2,097,651,003,692,820 >Unfortunately, this is not the same result as the distributed perft project got: >2,097,651,003,696,806 > >Which is not to say the distributed perft is wrong of course. It is quite >possible I have a silly bug, or hardware problems (the CPU is a recent upgrade). > I can share with you an unexplained result in a similar long computation when building checkers endgame databases over prolonged periods of time. When we compared our results to Dr. Jonathan Schaeffer of the Chinook team, of the 132 billion entries we were comparing, a few thousand were different. It turned out that his databases had some errors in the win-loss-draw results. His database results still "verified" when he ran his tests, and we have no idea how the errors crept in. You can see this in his paper which was just published in Graz 2003 alongside the WCCC. http://www.GothicChess.org/papers.html Anyway, it appears with so much of your calculations being done in RAM, my first guess would be you may have encountered a theoretical limit to reliability factor. 11 days * 86400 seconds per day * the number of computations done per second, when factoring into account your RAM access, well one wrong bit can throw your whole computation off. If there is a failure rate where a transistor packet switches states from +0.5V to -0.5V once every hundred trillion operations...well you may have encountered it. Just a SWAG -- Scientific Wild Ass Guess.
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.