Author: Moritz Berger
Date: 16:59:55 10/17/97
Go up one level in this thread
On October 17, 1997 at 17:16:37, Enrique Irazoqui wrote: < snip > >We'll see nothing really. 11 rounds won't prove much one way or another. >500 games per program played by SSDF mean much more. In there F5 will >not be so far from top. 50 points at the most if it runs with enough >RAM. > >Enrique We should maybe try to make it clear about which "Fritz 5" we are talking: - Fritz 5 on a K6/200 is slower than on my P5/166 (tested here with Beta2 version of Fritz 5 in my Gigabyte DX 586 Intel 430HX dual mainboard). -> Hiarcs 6 OTOH is 2.5-3.5 times faster on a K6/200 than on my P5/166 (nothing to do with "good" or "bad" programming - just different instruction pipelines/schedulers and HUGE difference in memory bandwith consumption of the 2 engines) - Fritz with e.g. 2 MB hash tables is 3-4 times slower at 40/120 than Fritz with 100 MB hash tables (which also get sometimes 100% filled on my P166 at 40/120 time control). Thorsten knows about this, of course ;-) -> Hiarcs is quite insensitive to less hash tables, also when you use e.g. Hiarcs 6/DOS on a 16MB RAM machine you get 15 MB HT, with Fritz 5 on the same machine (Windows also takes some RAM, add some for the graphical user interface ...) you will maybe get 1 MB HT). - Opening books: Fritz 5 comes with old (throw-away) .FBK books and a (small, too shallow for comp. vs. comp. games) 100 MB opening tree on CD which is not optimized for play vs. other programs. -> I use Fritz 5 with my own 1.5 Gigabyte opening tree or the Fritz 5 PowerBook tree (600 MB). -> The tree has to be "writable", i.e. on harddisk and not on CD for the learning function to be able to store learning values in it. This also makes a big difference if you e.g. play a sequence of 20 or 40 games against the same (computer) opponent. By chosing the "right" or "wrong" configuration, it is very easy to manipulate results in favour of either Fritz or Hiarcs, I guess the deviation would be at least +- 100 ELO points. My advice therefore is: Never trust results where the poster doesn't quote at least: - CPU used - RAM allocated for hash tables - book settings (book type, learning enabled?, book source, ...) - maybe also some figure like "it takes 20 seconds from initial position with fixed depth 8 ply to play 1.e4" to verify the correct speed of other components (2nd level cache, RAM settings and type (EDO, SDRAM, ...)). It's all slightly technical and complicated, but without sufficient information you are ill advised to believe *any* results of the aforementioned kind. Moritz P.S.: Fritz 4.01 with 64 MB hash tables vs. Hiarcs 6 (Fritz-Engine) with 32 MB hash tables on P166 with Fritz4.fbk and reversed colours from the same opening: result: Fritz got 52% out of 40 games 40/120 with permanent brain disabled.
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.