Author: Jesper Antonsson
Date: 15:02:59 11/24/01
This is speculation, as I'm not a chess program author myself, but I wonder if anyone has experimented recently and could explain things to me. Machines of today has a tremendous memory bottleneck and I asked myself the other day how come the large hash tables used today are beneficial. I remember figures from long ago when hash tables were said to give a speedup of perhaps 3%, but today when processors are 10-20 times faster than main memory, they should give less, perhaps even be detrimental? If you turn of hash tables entirely, how much of an increase does this give in NPS on a 1 Ghz+ processor? Nothing? A lot? Has anyone experimented with small hashtables, carefully tuned to fit in cache, and used perhaps only in shallow parts of the tree, and compared the results to the standard "use as much as you have"-approach? *Especially* in lightning games, where a huge hash table won't be filled anyway, a cache-tuned table could perhaps perform better? And by the way, does anyone bother to try to make sure his/her engine itself fits in instruction cache and that the search-function is so aligned that it won't get pushed out of instruction cache by more seldomly used functions? br, Jesper
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.