Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Match: Deep Rybka 4 SSE42 x64 vs Deep Rybka 4 x64
- - By rarara (**) Date 2010-05-28 17:03
Hi!

I am currently running an engine mathc between the two engines named in the title. My PC: Win 7 Home Premium 64-bit (no large pages), Intel Core i7 860 @ 2.80 GHz, 8 GB Ram  (1333).
Settings for tourney: 2 minutes and 1 second increment, 100 MB Hash, no opening book, no endgame tables.
Result so far:
Deep Rybka 4 SSE42 x64 - Deep Rybka 4 x64:
+1  =33  -6

I know, everyone yell now "not enough games". But still: Is it possible that the non-SSE42 version scores better? Can it depend on my (probably bogus) settings? What might be the reason for that?

I thought the SSE42 version should have a 5% performance plus?

Regards
rarara
Parent - - By Uly (Gold) [mx] Date 2010-05-29 05:51

> I thought the SSE42 version should have a 5% performance plus?


Set both to one CPU, let them analyze the opening position. If the SSE42 version is faster then... it's faster. You don't need to play several games for this.
Parent - By George Tsavdaris (****) Date 2010-05-30 09:27
Make this sticky for this topic. Obviously you will prevent some people from wasting their time or CPU resources. :-)
Parent - - By M ANSARI (*****) [kw] Date 2010-05-29 08:07
This is either a statistic coincidence or you have something seriously wrong with your setup.
Parent - - By Uly (Gold) [mx] Date 2010-05-29 08:09
Another member reported a Child process dying due to overclocking, maybe that's a place to look in.
Parent - By M ANSARI (*****) [kw] Date 2010-05-29 08:55
Yes, since Rybka 4 uses processes and not threads, something like that could really kill performance.  Anyway I have setup a match between SSE42 and default, and so far it is going exactly as expected with the SSE42 +14 ELO ahead into around 60 games.  I will let it run for 200 games using the silver suite at 2 0, which means running the 50 sets of positions twice from both sides.
Parent - By Dadi Jonsson (Silver) [is] Date 2010-05-29 08:13

> no opening book


This is not good.
Parent - - By M ANSARI (*****) [kw] Date 2010-05-31 06:49 Edited 2010-05-31 06:51
I finished the test and results are exactly as expected.  The SSE42 is around 7 ELO stronger after about 200 games with colors reversed.  Of course this is a 2 0 match against basically the same engine, so there will be more draws that there should be due to equal evaluation.  If both engines see 0.00 they will most likely go for a draw.  Most likely the real improvement is a little bit more than 7 ELO, especially against non Rybka opponents with differing evaluations.

1   Deep Rybka 4 SSE42 x64   +29/=146/-25   51.00%  102.0/200
2   Deep Rybka 4 x64            +25/=146/-29   49.00%   98.0/200
Parent - By Schiffermueller (*) [de] Date 2010-05-31 09:14

> The SSE42 is around 7 ELO stronger after about 200 games with colors reversed.


You should play a couple of thousend games to measure a 7 ELO difference. But in this case I would measure the speed up at a singel core because it is deterministic. From this the ELO difference can be calculated.

> Of course this is a 2 0 match


Don't use x+0 TC. It is too sensitive for TC-management. I use 1/4+1/4 TC with Arena-GUI to get a sufficient number of games in a few hours.
Parent - - By Banned for Life (Gold) Date 2010-05-31 16:28
Accepted wisdom is that a speed advantage for one side when the same engine plays against itself is magnified, rather than reduced. LK has speculated on this forum why this is so, but it is well supported by testing. So in all probability, your value would be overstated (but 200 games isn't nearly enough for this type of measurement).
Parent - - By M ANSARI (*****) [kw] Date 2010-05-31 17:13
This is true, but the 200 games were 50 positions with alternating colors for 2 runs.  So pretty neutral setup without too many unknowns.  The only factor is the non-deterministic behaviour of MP engines and I guess a better test would have been to use 1 core each.  I looked at many of the games and many were quickly drawn due to the evaluation being 0.00 on both sides.  I think if we remove that ability to allow draws the results would be higher in favor of the faster side.  Anyway this was just to quickly see if anything seemed wrong and if so to investigate in more detail.  This has not happened so as far as I am concerned it is not an issue.
Parent - By Leto (***) [us] Date 2010-06-11 20:18
What node counts are you getting for each version at the starting position?
Parent - - By ernest (****) [fr] Date 2010-06-12 12:38
Don't kid yourself, Majd!
You got the "expected" result, but the standard deviation of your result is 3.7
Which means that within 1 standard deviation (not to say 2, which gives the 95% usual probability),
you could as well get 98.5/200 instead of your 102/200 :wink:
Parent - By M ANSARI (*****) [kw] Date 2010-06-12 15:20
True of course, but that was just a quick check to see if something unusual was happening ... like maybe a bug in the SSE42 code etc.  That doesn't seem to be happening and most likely there is nothing wrong with the SSE42 version.  Had something larger than expected occurred I would have tested much deeper.
- - By Everest [br] Date 2010-05-30 09:21
Omg! I bought today, Deep Rybka x64, ill test it too.

btw Deep Rybka 4 x64 SSE 42 x64 is > Deep Rybka 4 x64 ?

Cheers,
Parent - - By rarara (**) Date 2010-05-30 09:30
no, that is a difference. the engine with the sse42 in the name is using some special instruction set that is only available on nehalem or phenom processor-series. it makes the engine about 5% faster (at least that is what it should ...)
Parent - - By Everest [br] Date 2010-05-31 00:53
Ohh good I have Phenom II x4 965 BE Revision C3.
Parent - - By pokerpawn (***) [be] Date 2010-05-31 03:01
to check just download cpu-z
http://www.filehippo.com/download_cpuz/

this will tell u all u need to know
Parent - - By isilverman (***) Date 2010-05-31 13:23
Sorry for being slow, but the program's report on my CPU says the following:

Instructions: MMX, SSE (1,2,3,3W,4.1) EM64T

Does this mean I can or can't run the SSE version?

Thanks.
Parent - - By pokerpawn (***) [be] Date 2010-05-31 13:26
you can't cause there is no mention of SSE4.2
Parent - By isilverman (***) Date 2010-05-31 14:10
Thanks!
- By Pelle [se] Date 2010-06-11 19:25
I think you can run the "SSE42" variant if you have the POPCNT instruction. Try it!
- - By fischerbobyrobertjam [hr] Date 2010-09-08 22:25
does all rybka 4 have the same engine strenght and what is the diferenties betwen deep rybka4 and rybka4 sorry i am not good in computers but i have rybka4 and somebody tells me that he have the better software is this means that he have the faster, or better engine?
Parent - By Indrajit (***) [in] Date 2010-09-09 02:28
Engine strength will depend upon your hardware too. There are some differences between the different rybka 4 versions.

Differences between rybka 4 and deep rybka 4 is that, rybka 4 is a single processor engine (i.e., it can only utilize one core). Whereas deep rybka will use as many processors as you have in the computer. Hence deep rybka will much faster. Find out from your friend which version he has. If he has Deep Rybka, then certainly it will be faster.
Parent - - By Labyrinth (****) [us] Date 2010-09-09 12:58
Engines are same.

Rybka4 = 1 Brain used ONLY

Deep Rybka4 = 1 to 2048 Brains

Think time = T

If #Brains_used ComputerRed > #Brains_used ComputerBlue

   Then ComputerRed = Faster

    Faster = More Search in T = Often Better Strength
Parent - By Uly (Gold) [mx] Date 2010-09-09 17:35

> If #Brains_used ComputerRed > #Brains_used ComputerBlue
>
>    Then ComputerRed = Faster
>
>     Faster = More Search in T = Often Better Strength


Huh, would this work?

Faster = More Search in T = Random(6) = Better strength.
- - By fischerbobyrobertjam [hr] Date 2010-09-09 19:36
so if i analayse with rybka4  3 hours and he 1 hour with deep rybka4 he will defeats me ???
Parent - By Indrajit (***) [in] Date 2010-09-09 21:08
Most likely it will be a draw. Coz, your searching for 3hrs might give the same depth as your friend's search for 1 hr. You can probably try a blitz match and see who wins.
- By fischerbobyrobertjam [hr] Date 2010-09-09 21:37
thats i want to hear thnx!!:grin:
Up Topic Rybka Support & Discussion / Rybka Discussion / Match: Deep Rybka 4 SSE42 x64 vs Deep Rybka 4 x64

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill