Page 3 of 4

Re: Komodo and SF: small android tourney

Posted: Wed Jan 07, 2015 11:58 pm
by stavros
i agree 1 sec/move is too short. 5 sec/move will be interesting.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 12:18 am
by Gurcan Uckardes
One visible evidence is the average nodes ramping up quickly during the first seconds. After that it's stabilized and changes very little.
I see the ramp up curve is way different from one engine to another, even among different versions of some engines. That warm-up phase might be something any tester should stay away.
I played 16000 bullet games and switched to a 30 seconds experiment. Elo differences between two TC were remarkable, up to +/-150 ELO.
Then i tried 15 secs to save time and checked elo effect. Almost nothing noticeable. 10-20 ELOs for each engine. They simply keep the same balance and distribution, be it 15 or 30 seconds.
Several comments of the community speak about "short" or "long" time controls. Sincerely i'm sceptical about that relative naming. How long is long and how short is short?
What matters to me is rather nodes/move. This is a criteria independent of the hardware, in line with CCRL's scaling factor. Same logic. They wanted to pin the nodes/move in relation to the reference cpu.
Therefore, even if i spend months to play games at rapid chess time control on Android, it means nothing more than bullet games one would play on TCEC hardware. Because SF reaches 16000-17000K on Martin's monster PC, while we're limited to 700-800K on a Galaxy Note 2.
I simply focus on the ratio between.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 7:23 am
by Laskos
Gurcan Uckardes wrote:One visible evidence is the average nodes ramping up quickly during the first seconds. After that it's stabilized and changes very little.
I see the ramp up curve is way different from one engine to another, even among different versions of some engines. That warm-up phase might be something any tester should stay away.
Right, especially MP engines. Good observation, the engines need at least several seconds in order to kick-in. I am testing now SF 121014 against SF 5 on 4 cores at 5 sec/move, for now the score in 60 games is 34:26 for SF 121014. Later today I will post the final result after 100 games. Pretty painful to test, at your 15 sec/move the test would have taken 2-3 days.
I played 16000 bullet games and switched to a 30 seconds experiment. Elo differences between two TC were remarkable, up to +/-150 ELO.
Then i tried 15 secs to save time and checked elo effect. Almost nothing noticeable. 10-20 ELOs for each engine. They simply keep the same balance and distribution, be it 15 or 30 seconds.
Several comments of the community speak about "short" or "long" time controls. Sincerely i'm sceptical about that relative naming. How long is long and how short is short?
What matters to me is rather nodes/move. This is a criteria independent of the hardware, in line with CCRL's scaling factor. Same logic. They wanted to pin the nodes/move in relation to the reference cpu.
Therefore, even if i spend months to play games at rapid chess time control on Android, it means nothing more than bullet games one would play on TCEC hardware. Because SF reaches 16000-17000K on Martin's monster PC, while we're limited to 700-800K on a Galaxy Note 2.
I simply focus on the ratio between.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 7:28 am
by Gurcan Uckardes
After your latest result, I start to think i686 and arm7 may not be the same. Well i intereupt Rapidroid today to take a look at the 15 sec duel on Exynos.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 12:40 pm
by stavros
yes i bet this is the issue.:i686 vs arm,anyway great blog:)

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 12:47 pm
by Gurcan Uckardes
There arw two droidfish 1.55, one present at gplay and another development version. They extract two binaries very different in size on arm7. I will check my i686 versions to reveal the truth. Meanwhile i started sF5 vs 121014 gplay version on my note 2 today :lol:

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 1:16 pm
by stavros
omg, yes 2 different sizes! from googleplay 3.289 mb, and from http://web.comhem.se/petero2home/droidfish/ 3.382 mb!
now which is better :?

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 2:21 pm
by Laskos
Gurcan Uckardes wrote:There arw two droidfish 1.55, one present at gplay and another development version. They extract two binaries very different in size on arm7. I will check my i686 versions to reveal the truth. Meanwhile i started sF5 vs 121014 gplay version on my note 2 today :lol:
I have used and use Stockfish 121014 dev't version for armv7 from your blog. Now I got the final result:
Stockfish 121014 vs Stockfish 5
Exynos 4 cores, 5 sec/move, 100 games.
56.5:43.5

Curious what you will get at 15 sec/move on Exynos 4 cores.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 3:01 pm
by Gurcan Uckardes
I'm preparing a summary of all available versions, arm7, i686 and their knps benchmark values.
In parallel to your results, i have so far:
SF5 4.5 - SF121014 6.5 after 11 out of 64 games.
I've started with TCEC6 superfinal openings.

What i wanna discover now is how SF5 crushed 121014 on i686.

Re: Komodo and SF: small android tourney

Posted: Thu Jan 08, 2015 3:37 pm
by stavros
Laskos wrote:
Gurcan Uckardes wrote:There arw two droidfish 1.55, one present at gplay and another development version. They extract two binaries very different in size on arm7. I will check my i686 versions to reveal the truth. Meanwhile i started sF5 vs 121014 gplay version on my note 2 today :lol:
I have used and use Stockfish 121014 dev't version for armv7 from your blog. Now I got the final result:
Stockfish 121014 vs Stockfish 5
Exynos 4 cores, 5 sec/move, 100 games.
56.5:43.5

Curious what you will get at 15 sec/move on Exynos 4 cores.
why not the google play version? whats the difference from dv version?