Author: Moritz Berger
Date: 13:43:10 02/25/98
Go up one level in this thread
On February 25, 1998 at 13:07:04, Enrique Irazoqui wrote: >But your idea is false. The difference in speed between Fritz 5 with 20 >MB tables and 100 MB tables is not greater than 35%, or some 20 or 25 >Elo points. You like better a Fritz 5 playing with 20 MB? OK, say it is >10 or 15 points weaker than with SSDF's 44 MB. That's all. Aaaah ... BUT: Fritz does not only run faster, but "smarter". I have observed on several occasions when running testsets that Fritz took way longer with 98 MB than with 12 MB, extending on certain moves for minutes and minutes that Fritz/12MB considered only for a second and then abandoned. This suggests more accurate information from "non polluted" hash tables, forcing the extended search in a critical position. a quote from the chessbase.usa site: > On a PC with 64MB of RAM you would be limited to around 45MB of hash tables. > In comparison with our 100MB, Fritz would be roughly 40-50 Elo points weaker > in tournament games. The 35% speedup from 20 to 100 MB you gave is correct also in my experience, but don't forget that there's more to it than just speed. Hiarcs 6, e.g. doesn't show any speedup at 40/120 tc with this hash table increase ... To all those who think that hash tables don't matter: Don't forget that Fritz is about the fastest program around, easily filling up 10 MB hash tables in 10 second on my P233MMX. Even at Blitz, using 100 MB instead of 15 MB gave me a speedup from 20s to 18s in one position. One more word about "fair SSDF testing conditions": The SSDF tested Fritz mostly on machines with 8 MB RAM: Subtract 4 MB for Windows, 3 MB for the Fritz GUI and you are left with about 512-1024KB hash tables for Fritz. Small wonder that Fritz made some peculiar moves with a 100-fold "oversaturated" hash table ... No replacement policy in the world could remedy this ... I guess that the hash table limit alone was responsible that Fritz ran at 25% of the speed it could have achieved on the same machine with e.g. 32 MB hash tables. And don't forget the "accuracy" issue ... Just for fun, I tested Fritz 4.01 at 40/120 time controls on 2 P166 machines(remember, the SSDF mostly used the initial 4.00 release with its well known engine and tablebase bugs): It scored 51% vs. Hiarcs 6/Fritz engine/identical opening positions (40 games) 50% vs. Rebel 8 (29 games) 60+% vs. M-Chess 5 (20 games) etc. I do not question the motives of ChessBase to make sure this time that Fritz will get reasonable hardware (remember, Frizt 4 was rated 2285 on P90 when it entered the list !!!!). RAM prices are such that 64 MB of any kind of RAM (SDRAM, EDO, you name it) can be had for about $125. I do not think therefore, that when testing a program's playing strength giving it a 64 MB machine is unreasonable (just figure the total price you paid for that machine in relation to such a moderate RAM upgrade). Moritz
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.