Author: Paulo Soares
Date: 23:38:21 04/18/99
Go up one level in this thread
On April 18, 1999 at 06:46:19, Ernst Walet wrote: >On April 18, 1999 at 01:23:25, Ernst Walet wrote: > >> >>And CM6000 gives the following result with 8MB (and 32MB) hash: >> >>1:04 (1:05) 1/9 Bxe6 -4.40 >>1:45 (1:45) 1/9 Rxh6 -3.24 >>2:27 (2:24) 1/10 Rxh6 -0.75 >>4:08 (3:46) 1/11 Rxh6 -0.75 >> >>so solves it in 105 sec. >> >>Ernst-J. >> >>The reason to test it with 8MB and 32MB hash is that I want to determine if The >>King engine in Tascbase 2.1a suffers much from it's limitation of 8MB hash. >> >>So far I've found, in CM6K, on a K6-200, a penalty of 3% at 3 min., a penalty of >>5% at 3 min., a penalty of 14% at 10 min., and a penalty of 17% at 30 min.. >> >>What penalty should I expect at a overnight analysis? I expect that the curve >>flattens out at longer time controls at maybe 25% or so. Still 25% of about 8 >>hours gives a total time of 10 hours. Maybe a programmer could shine a light on >>this. >> >>What I also noticed is that The King, as far as I can tell, does not rely on the >>hash table for it's evaluation. It still finds moves at the same ply depth with >>the same, or about the same score with different hash sizes, only the time >>differs. Rebel for example sometimes finds moves at a certain ply depth in a >>certain time with 60MB hash, but doesn't find it with 40MB hash at all, or much >>later at much deeper ply depth. >> >>My conclusion sofar is that The King does not rely on hash tables for evaluation >>and that Rebel does, is this right? >> >>Kind regards, Ernst-J. > > >While trying Nolot #10 I've noticed a time diference between 32MB and 8MB up to >50%! This startled me a bit, meaning that on a lets say PII 450, sometimes >after 10min., there is a difference of 40% and after 20min. even 50%. Where >does this end, 100%? > >My results for 8MB (and 32MB) on a K6 200 are: > > 0:26 (0:27) 1/8 -0.09 Bb3 0% > 0:44 (0:44) 1/8 0.06 Rxf7 0% > 1:47 (1:46) 1/9 0.68 Rxf7 0% > 5:53 (5:18) 1/10 -1.64 Rxf7 11% > 8:32 (7:38) 1/10 0.00 Bb3 12% >23:26 (16:57) 1/11 -0.05 Bb3 38%! >37:58 (26:42) 1/11 0.37 Rxf7 42%! >66:52 (44:37) 1/12 0.08 Rxf7 50%! > >I stopped after about 2 hours of anaysing, completing ply 12. > >Is this difference common for other programs or? One could say that a maximum >of 8MB hastable size in Tascbase is out of date..., I do. > >Kind regards, Ernst-J Ernst, Very interesting his tests, I never had seen nothing seemed. I do not understand which is the size of optimous hashtable. I read that the CM6000 programmer considers one than enough the use of 16Mb, for SSDF games. I read also that Rebel8 is better with 8Mb than with 32Mb (if does not fail the memory) for the SSDF games. Regards, Paulo Soares
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.