the benefit of 4 cores compared to 1 core (CEGT 40/4)

Discussion of computer chess matches and engine tournaments.

Moderator: Ras

ThatsIt
Posts: 992
Joined: Thu Mar 09, 2006 2:11 pm

the benefit of 4 cores compared to 1 core (CEGT 40/4)

Post by ThatsIt »

Hi to all !

Some stats "4CPU ---> 1CPU".

Code: Select all

Komodo 5.0 x64 1CPU        ELO 3001 /  3800
Komodo 5.1 x64 4CPU        ELO 3117 /  1500 + 116 **

Houdini 3.0 x64 1CPU       ELO 3082 /  3200
Houdini 3.0 x64 4CPU       ELO 3183 /  3400 + 101

Shredder 12 x64 1CPU       ELO 2800 / 11038 (list reference)
Shredder 12 x64 4CPU       ELO 2899 /  4900 +  99


some notable engines:
Critter 1.4 x64 1CPU       ELO 2978 /  2400
Critter 1.4 x64 4CPU       ELO 3068 /  3600 +  90

Equinox 1.70 x64 1CPU      ELO 2955 /  1700
Equinox 1.70 x64 4CPU      ELO 3050 /  2700 +  95

Junior 12 x64 1CPU         ELO 2688 /  3100
Junior 12 x64 4CPU         ELO 2847 /  3600 + 159 (!)

Rybka 4.0 x64 1CPU         ELO 2966 /  7842
Rybka 4.0 x64 4CPU         ELO 3055 /  4900 +  89

Stockfish 2.2.2 x64 1CPU   ELO 2961 /  2700
Stockfish 2.2.2 x64 4CPU   ELO 3085 /  1600 + 124
** Maybe this number should be reduced a bit (by 5-10 points)
because Version 5.0 is not totally equal with version 5.1 ?!

Best wishes,
G.S.
User avatar
Ajedrecista
Posts: 2188
Joined: Wed Jul 13, 2011 9:04 pm
Location: Madrid, Spain.

Re: The benefit of 4 cores compared to 1 core (CEGT 40/4).

Post by Ajedrecista »

Hello Gerhard:
ThatsIt wrote:Hi to all !

Some stats "4CPU ---> 1CPU".

Code: Select all

Komodo 5.0 x64 1CPU        ELO 3001 /  3800
Komodo 5.1 x64 4CPU        ELO 3117 /  1500 + 116 **

Houdini 3.0 x64 1CPU       ELO 3082 /  3200
Houdini 3.0 x64 4CPU       ELO 3183 /  3400 + 101

Shredder 12 x64 1CPU       ELO 2800 / 11038 (list reference)
Shredder 12 x64 4CPU       ELO 2899 /  4900 +  99


some notable engines:
Critter 1.4 x64 1CPU       ELO 2978 /  2400
Critter 1.4 x64 4CPU       ELO 3068 /  3600 +  90

Equinox 1.70 x64 1CPU      ELO 2955 /  1700
Equinox 1.70 x64 4CPU      ELO 3050 /  2700 +  95

Junior 12 x64 1CPU         ELO 2688 /  3100
Junior 12 x64 4CPU         ELO 2847 /  3600 + 159 (!)

Rybka 4.0 x64 1CPU         ELO 2966 /  7842
Rybka 4.0 x64 4CPU         ELO 3055 /  4900 +  89

Stockfish 2.2.2 x64 1CPU   ELO 2961 /  2700
Stockfish 2.2.2 x64 4CPU   ELO 3085 /  1600 + 124
** Maybe this number should be reduced a bit (by 5-10 points)
because Version 5.0 is not totally equal with version 5.1 ?!

Best wishes,
G.S.
Thanks for the info. Let me include error bars with the method of SRSS (Square Root of Squares Sum). If an engine A has a rating of r_A ± e_A and other engine B has a rating of r_B ± e_B, and in the case of r_A > r_B, then the rating difference will be r_A - r_B ± sqrt[(e_A)² + (e_B)²]. If I find an engine with rating between [r - x, r + y] (x =/ y) then I take the average error bar (x + y)/2... luckily, I did not face those cases. I give the results of square roots rounded up to 0.1 Elo:

Code: Select all

----------------------------
4 CPU vs. 1 CPU (CEGT 40/4).
----------------------------

Komodo 5.0 and Komodo 5.1:
(Komodo 5.1 x64 4CPU) - (Komodo 5.0 x64 1CPU) = 3117 - 3001 ± sqrt[(15)² + 9²] ~ +116 ± 17.5 Elo.

Houdini 3:
(Houdini 3.0 x64 4CPU) - (Houdini 3.0 x64 1CPU) = 3183 - 3082 ± sqrt[(12)² + (10)²] ~ +101 ± 15.6 Elo.

Shredder 12:
(Shredder 12 x64 4CPU) - (Shredder 12 x64 1CPU) = 2899 - 2800 ± sqrt(8² + 5²) ~ +99 ± 9.4 Elo.

Critter 1.4:
(Critter 1.4 x64 4CPU) - (Critter 1.4 x64 1CPU) = 3068 - 2978 ± sqrt[9² + (11)²] ~ +90 ± 14.2 Elo.

Equinox 1.70:
(Equinox 1.70 x64 4CPU) - (Equinox 1.70 x64 1CPU) = 3050 - 2955 ± sqrt[(10)² + (12)²] ~ +95 ± 15.6 Elo.

Junior 12:
(Junior 12 x64 4CPU) - (Junior 12 x64 1CPU) = 2847 - 2688 ± sqrt[9² + (11)²] ~ +159 ± 14.2 Elo.

Rybka 4:
(Rybka 4.0 x64 4CPU) - (Rybka 4.0 x64 1CPU) = 3055 - 2966 ± sqrt(8² + 6²) = +89 ± 10 Elo.

Stockfish 2.2.2:
(Stockfish 2.2.2 x64 4CPU) - (Stockfish 2.2.2 x64 1CPU) = 3085 - 2961 ± sqrt[(13)² + (10)²] ~ +124 ± 16.4 Elo.
I did the math with a Casio calculator, so it may contain errors. As a bonus, Zappa Mexico II is here:

Code: Select all

----------------------------
4 CPU vs. 1 CPU (CEGT 40/4).
----------------------------

Zappa Mexico II:

(Zappa Mexico II x64 4CPU) - (Zappa Mexico II x64 1CPU) = 2837 - 2694 ± sqrt(6² + 7²) ~ +143 ± 9.2 Elo.
Thank you very much for all the effort of CEGT team!

Regards from Spain.

Ajedrecista.
Modern Times
Posts: 3803
Joined: Thu Jun 07, 2012 11:02 pm

Re: the benefit of 4 cores compared to 1 core (CEGT 40/4)

Post by Modern Times »

Thanks. +100 Elo is about the norm from my experience. Some engines do better, other don't quite manage that.
beram
Posts: 1187
Joined: Wed Jan 06, 2010 3:11 pm

Re: the benefit of 4 cores compared to 1 core (CEGT 40/4)

Post by beram »

ThatsIt wrote:Hi to all !

Some stats "4CPU ---> 1CPU".

Code: Select all

Komodo 5.0 x64 1CPU        ELO 3001 /  3800
Komodo 5.1 x64 4CPU        ELO 3117 /  1500 + 116 **

Houdini 3.0 x64 1CPU       ELO 3082 /  3200
Houdini 3.0 x64 4CPU       ELO 3183 /  3400 + 101

Shredder 12 x64 1CPU       ELO 2800 / 11038 (list reference)
Shredder 12 x64 4CPU       ELO 2899 /  4900 +  99


some notable engines:
Critter 1.4 x64 1CPU       ELO 2978 /  2400
Critter 1.4 x64 4CPU       ELO 3068 /  3600 +  90

Equinox 1.70 x64 1CPU      ELO 2955 /  1700
Equinox 1.70 x64 4CPU      ELO 3050 /  2700 +  95

Junior 12 x64 1CPU         ELO 2688 /  3100
Junior 12 x64 4CPU         ELO 2847 /  3600 + 159 (!)

Rybka 4.0 x64 1CPU         ELO 2966 /  7842
Rybka 4.0 x64 4CPU         ELO 3055 /  4900 +  89

Stockfish 2.2.2 x64 1CPU   ELO 2961 /  2700
Stockfish 2.2.2 x64 4CPU   ELO 3085 /  1600 + 124
** Maybe this number should be reduced a bit (by 5-10 points)
because Version 5.0 is not totally equal with version 5.1 ?!


Best wishes,
G.S.
*** It should probably be reduced by 20-30 points, as because in the IPON list Komodo 5.1 1CPU is 26 ELO better than Komodo 5.0 (15 ELO behind Komodo CCT)