Author: Vincent Diepeveen
Date: 08:36:41 10/14/02
Go up one level in this thread
On October 13, 2002 at 22:00:01, Robert Hyatt wrote: >On October 13, 2002 at 10:21:05, Gian-Carlo Pascutto wrote: > >>On October 13, 2002 at 10:05:38, Robert Hyatt wrote: >> >>>But not on an 8-way using the fusion chipset. the quads scale well due to 4->way memory interleaving. The 8-way still uses 4-way interleaving ahd produces >>>a significant memory bottleneck... >> >>Want to bet that, when actualy efficiency is not a consideration, >>I can design a parallel system that scales well _in NPS_ no matter >>what memory bandwidth is available? > >I'll take that bet, so long as it is a _real_ parallel search using shared >memory. > >Note that we are talking about a real chess program. Not some junk thing that >just searches the same stuff everywhere... > >IE something comparable to Fritz. I do not see why fritz would not get that 3.5 MLN nodes a second at 4 processors or a dual P4 2.8Ghz when optimized for it. If he simply has kicked out the stupid MMX stuff and replaced it by less MMX stuff and decent P4 code (so no branches) he can get 3.5MLN at a dual 2.8 P4 with DDR ram and 533 bus. For sure is that his program fits within the caches... > >> >>I could start by not sharing the hashtables. >> >>You get the idea... >> > > >No I don't, because we are talking about _fritz_ here. A _real_ parallel search >engine. Nobody in their right mind would do that unless forced to do so due to >some oddball constraint having nothing to do with chess. > > > > > >>-- >>GCP
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.