Author: Vasik Rajlich
Date: 03:54:13 01/12/06
Go up one level in this thread
On January 11, 2006 at 06:30:54, Uri Blass wrote: >On January 11, 2006 at 05:12:34, Vasik Rajlich wrote: > >>On January 10, 2006 at 06:40:43, Uri Blass wrote: >> >>>On January 10, 2006 at 04:43:57, Vasik Rajlich wrote: >>> >>>>On January 09, 2006 at 21:41:17, Arnon Yogev wrote: >>>> >>>>>On January 09, 2006 at 19:38:59, robert flesher wrote: >>>>> >>>>>>why not 64 bit providers a huge speed gain from what I have read. Even if its >>>>>>only 30-50 elo that is alot. An extra ply here or there can spell death. I >>>>>>prefer to see testing under fair conditions. >>>>> >>>>> >>>>>I don't think that this particular type of hardware advantage can cause such a >>>>>difference in results. >>>>>Do you honstly think that the 32 bit version would score below 50% just because >>>>>of it? hard to believe.. >>>>> >>>>>For example, take a look at the latest CEGT results : >>>>> >>>>>Rybka 1.0 Beta 32-bit Vs. Deep Shredder 9 2CPU 512MB >>>>>== >>>>>104 + 50 = 25 - 29 >>>>>! >>>>> >>>>>The barely 80 kn/s fish slaughters the double CPU'ed kn/s monster. >>>>> >>>>>Hiarcs versus the same opponent scores + 16 = 15 - 19 which is pretty decent, >>>>>but not slightly close to Rybka's result. >>>>> >>>>>From my personal exeperience I can tell you that while playing online for the >>>>>first time i could pass the 2600+ rank with my antique P4 1.8 playing against >>>>>_much_ faster machines. >>>>>all this thanks to Rybka. before that era I could barely pass the 2450 rate with >>>>>any other engine.. Rybka just brought back all the fun =). >>>>> >>>>>I don't ask Hiarcs to win every match it set against, But I think that we could >>>>>except more from a commercial release .. Surely just to give some more >>>>>competition. (I'm not speaking just about Rybka, so spare the "it was not >>>>>available at the time"..) >>>>> >>>>>AY. >>>>> >>>>>PS if the setting is really better than default, Why on earth it was not >>>>>discovered before the release ? I remember enrico said that they ran more than >>>>>1000 long games for the beta version before the release, so I mean, none of the >>>>>beta-testers could have just think about changing some parameters? that's their >>>>>jobs and why they are betatesters anyway.. >>>> >>>>Enrico's explanation is pretty good. I want to add three things: >>>> >>>>1) 60% speedup (as you get going from 32-bit Rybka to 64-bit Rybka) gives around >>>>30 rating points. >>> >>>I think that it is dependent on the time control. >>> >>>http://www.husvankempen.de/nunn/eloblitzall.html >>> >>>I see bigger difference at blitz and I think that there is also a bigger >>>difference for weaker programs than rybka. >>> >>>I guess that the advantage is bigger when the engine does more mistakes. >>>It is also possible that the 30 elo is misleading because it is harder to >>>improve result against weak opponents and the difference is going to be bigger >>>when the opponents becomes stronger. >>> >>>Unfortunately no opponent in the CEGT list is stronger than Rybka and rybka >>>plays only against weaker opponents(it is possible to change the situation by >>>giving the opponents more time and have CEGT rating for things like Fruit2.2*3 >>>when you give Fruit2.2 120 minutes against 40 minutes of other opponents) but >>>unfortunately I do not expect it to happen). >>> >>>Uri >> >>Interesting. Indeed, I took a look, and found the following: >> >>Blitz, Rybka 64 bit vs 32 bit: 58 ELO >>40/40, Rybka 64 bit vs 32 bit: 31 ELO >>Blitz, Shredder 9 2 CPU vs Shredder 9 1 CPU: 61 ELO >>40/40, Shredder 9 2 CPU vs Shredder 9 1 CPU: 51 ELO >> >>The number which is out of line with the others is the first one, not sure why >>it's so high. >> >>Also, some small evidence overall for diminishing returns. >> >>Vas > >If you take weaker programs you can find bigger improvements from 64 bits 2 cpu. >Not enough games were played but it seems to be 100-200 elo improvement from the >combination of 2 cpu and 64 bits. > >I am not sure if it is diminishing returns or that the difference for programs >in the top of the list is underrated. > >14 Zappa 1.1 64-bit 2CPU 512MB 2754 >94 Zappa 1.1 2607 > >30 Deep Sjeng 1.6 2CPU 512MB 2708 >92 Deep Sjeng 1.6 1CPU 2608 > >26 Crafty Cito 1.4.2 64/bit 2CPU 2715 >134 Crafty Cito 1.2-1.4.1 2533(best Crafty 1 cpu) >146 Crafty Cito 1.4.2 2514 (not enough games but I guess no big difference to >1.2-1.4.1) > >Uri Interesting. Here it might be not enough games, Zappa played just 103 and Sjeng just 48. There is also a Shredder 9.12 with a difference of 184 ELO (but 79 games total for both versions). Anyway, there are probably some other factors as well. For example, every engine throws away some points - getting crushed in the opening, giving away half a point in the endgame, etc. So there are going to be some boundaries on what search can do, and those boundaries may be smaller between more mismatched engines (for example). Vas
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.