Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Performance - Rybka 3 vs. Deep Rybka 3
- - By Kedar (*) [cz] Date 2008-11-14 17:43
Hello everyone,

can anyone give me an information on performance differences probability between Rybka 3 (single core) and Deep Rybka 3 (4 cores)? I assume that Deep Rybka 3 on my 4 cores CPU will be much stronger that the basic version on a single core system, but for how much? Is it going to be something about 10 ELO points or 200 points?
On the single core system (CPU 2600+) I used the hash size 70 MB at maximum, now I can give it up to 2000 MB hash size, having CPU 4 x 2.2 GH.

Thanks for any hints.
Parent - - By Pia (****) [ru] Date 2008-11-14 17:58
I have same CPU 2600+ and CPU 4x2.4 recently. But I don't bother with that question. The last is about up to 10 times faster, that's all I need.
2x performace = about +50 Elo. 10/2*50=250 Elo at maximum.
Parent - - By Uri Blass (*****) [il] Date 2008-11-14 19:18
Your math is wrong.

If being 2 times faster is +50 elo then being 10 times faster is 50*log10/log2 elo and it is only 160-170 elo.

Note that there may be diminishing returns when things are dependent on time control so the +50 is not constant.

Uri
Parent - - By Uly (Gold) [mx] Date 2008-11-14 20:41

> If being 2 times faster is +50 elo then being 10 times faster is 50*log10/log2 elo and it is only 160-170 elo.


I'm glad my very wild guess of 200 points wasn't far from the truth (considering I don't even know how log works.) - though I calculated it for a 2Ghz Pentium 4 so perhaps I was closer.
Parent - By Carl Bicknell (*****) [gb] Date 2008-11-14 22:43

> (considering I don't even know how log works.) -


of course there is a cheaters way. If 2x speed = 50 elo then just count how many doublings there are. So for a 10x faster system it'd go: 1 - 2 - 4 - 8. That's three doublings of speed so 50 x 3 = 150 elo. Then just add a bit more because the system is 10x faster not 8 x faster and you get Yuri's 170 elo!

:)
Parent - By stvs (***) [gr] Date 2008-11-14 22:39 Edited 2008-11-14 23:07
with that way of thinking one supercomputer  1000 times faster will be: 1000/2=500*50 =+25000! elo? yuri have right. the formula x2 faster=50 elo gain dont work further  cuz the nature of the game.after deep search from 2 supercomputers who the one have the double speed
the elo gain will be tiny maybe 3-5 elo. but from yuri formula one supercomputer 1000 times faster than skulltrail we have: 50*log1000/log2=498 !!!
skulltrail =3200 elo +498=3698 near perfect chess? i have doubts about that formula..i think above from specific depth or speed the +50 elo gain
reduce dramatically ...my estimation about the fastest supercomp roadrunner is 3500 elo the most, otherwise we talk about perfect chess and its impossible...
Parent - - By Shaun (****) [gb] Date 2008-11-15 00:00 Edited 2008-11-15 00:06
I played some games directly between R3 1CPU and R3 4CPUs

CCRL 40/40

and

CCRL 40/4

Shaun

EDIT:

Please note the 'perf' displayed currently +14/+18 is relative to the R3 4CPUs rating and not the difference between R3 4CPUs and R3 1CPU, meaning that Rybka with 4CPUs performed better than expected based on its rating against its 1CPU brother.

R3 4CPUs is currently rated 63 and 89 elo higher than R3 1CPU in the CCRL 40/40 and 40/4 respectively.
Parent - By Roland Rösler (****) [de] Date 2008-11-15 04:45
It´s very interesting and surprising for me to see in 40/4 rating list, that after 6.5 losses :-) R3 sp wins a game (+188 -29 =223). Not bad for same engine with one-third of time.
In 40/40 rating list we see 4 wins for R3 sp (+19 -4 =31).

And my conclusion is: Chess is a complex game, also for the best engine. :-)
Up Topic Rybka Support & Discussion / Rybka Discussion / Performance - Rybka 3 vs. Deep Rybka 3

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill