Author: Bas Hamstra
Date: 02:05:05 01/13/03
Go up one level in this thread
>i hope you realize that there is a big contradiction between 'extensive pawn >structure analysis', fast search, checks in qsearch, good move sorting and fast >nps. Yes, probably. You start to wonder how Frans did it in Fritz5. But F5 proves that is IS possible. At least checks/speed/sorting. Although sorting I'm not sure of, with F5's extremely heavy dependency on hashsize. >It is not hard to make something search 10 million nodes a second. It will be >horrible inefficient though. It is not hard to make something search deep with >million nodes a second (single cpu) and just psq. But fastest of all is and >will remain of course replacing qsearch by a SEE. I am not so sure of that. It would save a lot of qnodes, but my experience is that it also destabilizes the search so bad that the overall effect on depth could even be negative. But I am not going to bloat nps at the cost of efficiency of course, it should have a normal FH rate. >Just add the SEE to eval and that's it. Very well possible with PSQ progs >Basically you want to make 'extensive pawn structure analysis'. I would call it >dumb pawn patterns, because you do only consider pawns versus pawns in your >hashtable. >what's the use of that? You can search way faster by not doing checks in >qsearch. Crafty is best example of it. Also no pawnevaluatoin there of course >and a big lazy eval. Ed's 0.50 pawn which is ridicioulous small margin, >springs to mind. I hope you realize that if you search at a few million nps, >that getting from main memory a cache line with the pawnstructure code is very >expensive. it is about 400 clocks or so. Very expensive! then you also waste >system time to figuring out which checks to try. Very expensive. Slowing you >down a lot. i hope you see the problem. Yes, difficult problems to solve, I know. But a question: since when is one hashprobe 400 clocks? I don't uderstand. Best regards, Bas.
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.