Author: Eugene Nalimov
Date: 12:51:17 01/21/00
Go up one level in this thread
On January 21, 2000 at 14:39:18, Tom Kerrigan wrote: >On January 21, 2000 at 14:03:58, Eugene Nalimov wrote: > >>In his IEEE Micro article Hsu estimated his evaluation function as an equivalent >>to ~40,000 general purpose CPU instructions. (Or is the entire procesing of one >>node? In any case, that doesn't matter - all other work can be done in 0.5-2k >>instructions). > >So let's say you have a nice new Pentium III running at 800MHz. If one >instruction takes one clock cycle, that translates to 20000 NPS. > >If each instruction takes 2 clock cycles (an absolute worst-case scenario) >that's still 10000 NPS. > >If you have the world's best evaluation function, I think 10000 NPS should be >enough for a competitive program. And if it isn't, well, the DB program is >already parallel... > >I don't see why FHH doesn't do this. > >-Tom DB search was designed in the assumption that it's *fast*. For 20knps (or even for 100k) it's necessary to rewrite a lot of stuff - e.g. remove singular extensions, add null move (or something other based on "null-move observation"), etc. And in the process Hsu will find that removing slowest parts of the evaluation function will make the program stronger, as nps will go way up. In the end he'll end with one more PC chess program - exactly as Bob did. And first version of the program will be not the strongest one, as speed/knowledge tradeofs will differ wildely from Hsu's instincts. Eugene
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.