Author: James Robertson
Date: 14:54:35 09/26/98
Go up one level in this thread
On September 26, 1998 at 17:30:09, John Coffey wrote: >On September 26, 1998 at 17:21:58, James Robertson wrote: > >>What is a reasonable number of NPS for a program to search if it has an average >>sized evaluation procedure? I.E. what does Crafty get? Chessmaster? Comet? My >>program started out searching a very nice number of nodes per second, but as I >>added stuff that number decreased dramatically from about 600,000 to 160,000. By >>the time I finish the evaluation function (which is currently like 20 lines), >>I'm afraid it will be down to like 5,000 NPS on my P233! At what point should I >>get worried? >> >>James > >This is an interesting question, because Robert Hyatt told me the other day >that it was the size of the tree and not the NPS that really matter. When >I looked at the Crafty source code, I was surprised at how much code that >gets executed for every single node. But the main purpose of that code is >to cut down the size of the tree (using null moves, hash tables, move ordering, >etc.) That tree grows exponentially, so cutting it from say 6^N down to 3^N >is a huge improvement regardless of the number of nodes per second. > >Fritz 5 is suppose to use about a 1,000 clock cycles per node. Crafty gets >80,000 nodes on a Pentium 200 mmx, so that works out to be about 2,500 clock >cycles per node. But Crafty is still very very strong. > >John Coffey Yeah, my tree size has diminished a lot since I added all that move ordering stuff. How on earth does Fritz take only 1,000 clock cycles per node?? James
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.