Page 1 of 2
How many elo do ponder equals to
Posted: Sat Jun 22, 2019 2:21 pm
by Nordlandia
Engine A ponder while engine B doesn't. Advantage +/- 50 elo. Doubling is worth more. Now the question is if it's likely engine A equalize engine B in fast time formats or who got the upper hand. The effectiveness of ponder is probably higher in shorter time controls than in longer.
Do you put your bet on 4-core PB=on versus 8-core PB=off in 1 minute bullet games?
Hardware is assumed to be equal.
I'm referring to to this old thread.
http://www.talkchess.com/forum3/viewtop ... =7&t=48864
Re: How many elo do ponder equals to
Posted: Sat Jun 22, 2019 2:44 pm
by zullil
Nordlandia wrote: ↑Sat Jun 22, 2019 2:21 pm
Engine A ponder while engine B doesn't. Advantage +/- 50 elo. Doubling is worth more. Now the question is if it's likely engine A equalize engine B in fast time formats or who got the upper hand. The effectiveness of ponder is probably higher in shorter time controls than in longer.
Do you put your bet on 4-core PB=on versus 8-core PB=off in 1 minute bullet games?
Hardware is assumed to be equal.
I'm referring to to this old thread.
http://www.talkchess.com/forum3/viewtop ... =7&t=48864
If A and B are the
same engine, it seems that A's pondering will very often be successful in pre-searching the move eventually made by B. If so, it's roughly like giving A twice the search time that B has. So I'm voting for A.
Re: How many elo do ponder equals to
Posted: Sat Jun 22, 2019 3:11 pm
by syzygy
Nordlandia wrote: ↑Sat Jun 22, 2019 2:21 pm
Hardware is assumed to be equal.
And I suppose you will be running the engines on separate machines (or the pondering engine will compete for resources with the non-pondering engine when it is the latter's turn).
Re: How many elo do ponder equals to
Posted: Sat Jun 22, 2019 3:26 pm
by Nordlandia
syzygy wrote: ↑Sat Jun 22, 2019 3:11 pm
Nordlandia wrote: ↑Sat Jun 22, 2019 2:21 pm
Hardware is assumed to be equal.
And I suppose you will be running the engines on separate machines (or the pondering engine will compete for resources with the non-pondering engine when it is the latter's turn).
Imagine two machines with same hardware linked together. Same engine on both machines.
Re: How many elo do ponder equals to
Posted: Sat Jun 22, 2019 11:50 pm
by MikeB
I can actually test this on my 12 core machine - I believe a decisive advantage will got the 8 core on the simple basis everty search counts - whereas pondering will have some wasted searches , maybe not many with like engines , but enough to be the difference maker.
Re: How many elo do ponder equals to
Posted: Sat Jun 22, 2019 11:58 pm
by MikeB
MikeB wrote: ↑Sat Jun 22, 2019 11:50 pm
I can actually test this on my 12 core machine - I believe a decisive advantage will got the 8 core on the simple basis everty search counts - whereas pondering will have some wasted searches , maybe not many with like engines , but enough to be the difference maker.
Will modify the threads to be 6 and 3, same concept, so as not to use more than 12 threads at any onetime.
Re: How many elo do ponder equals to
Posted: Sun Jun 23, 2019 12:27 am
by MikeB
MikeB wrote: ↑Sat Jun 22, 2019 11:58 pm
MikeB wrote: ↑Sat Jun 22, 2019 11:50 pm
I can actually test this on my 12 core machine - I believe a decisive advantage will got the 8 core on the simple basis everty search counts - whereas pondering will have some wasted searches , maybe not many with like engines , but enough to be the difference maker.
Will modify the threads to be 6 and 3, same concept, so as not to use more than 12 threads at any onetime.
Everyone interested can follow along here - this is just a 100 game match using 50 positions.
Initial results:
Code: Select all
100 games were planned
Match started: 06/22/19 : 18:18:34
Target completion: 06/22/19 : 21:38:34
Base: 60 seconds
Increment: 0 seconds
found 1 hosts to use
1 of 100 games completed...
Time control: 60 seconds + 0 seconds
Threads = 12
Hash = 256
Date: 06/22/19 : 18:20:34
rm: pgn.txt: No such file or directory
1 game(s) loaded
Rank Name Rating Δ + - # Σ Σ% W L D W% =% OppR
---------------------------------------------------------------------------------------------------------
1 McCain-X4-6CPU 3107 0.0 7 7 1 1.0 100.0 1 0 0 100.0 0.0 3093
2 McCain-X4-3CPU 3093 13.9 7 7 1 0.0 0.0 0 1 0 0.0 0.0 3107
---------------------------------------------------------------------------------------------------------
1 game(s) loaded
2 of 100 games completed...
Time control: 60 seconds + 0 seconds
Threads = 12
Hash = 256
Date: 06/22/19 : 18:22:34
2 game(s) loaded
Rank Name Rating Δ + - # Σ Σ% W L D W% =% OppR
---------------------------------------------------------------------------------------------------------
1 McCain-X4-6CPU 3106 0.0 6 6 2 1.5 75.0 1 0 1 50.0 50.0 3094
2 McCain-X4-3CPU 3094 11.6 6 6 2 0.5 25.0 0 1 1 0.0 50.0 3106
---------------------------------------------------------------------------------------------------------
2 game(s) loaded
Results - updated every two minutes:
https://www.dropbox.com/s/xlz8d9rt5c1sj ... a.txt?dl=0
PGN Files - also updated every two minutes:
https://www.dropbox.com/s/e808r1e9o253i2x/all.txt?dl=0
Re: How many elo do ponder equals to
Posted: Sun Jun 23, 2019 12:57 am
by MikeB
zullil wrote: ↑Sat Jun 22, 2019 2:44 pm
Nordlandia wrote: ↑Sat Jun 22, 2019 2:21 pm
Engine A ponder while engine B doesn't. Advantage +/- 50 elo. Doubling is worth more. Now the question is if it's likely engine A equalize engine B in fast time formats or who got the upper hand. The effectiveness of ponder is probably higher in shorter time controls than in longer.
Do you put your bet on 4-core PB=on versus 8-core PB=off in 1 minute bullet games?
Hardware is assumed to be equal.
I'm referring to to this old thread.
http://www.talkchess.com/forum3/viewtop ... =7&t=48864
If A and B are the
same engine, it seems that A's pondering will very often be successful in pre-searching the move eventually made by B. If so, it's roughly like giving A twice the search time that B has. So I'm voting for A.
You are forgetting B is using twice as many cores. So A has to ponder CORRECTLY , just to stay even in productive CPU time . Every ponder that was incorrect , was inefficient use of CPU time.
Re: How many elo do ponder equals to
Posted: Sun Jun 23, 2019 8:04 am
by Nordlandia
MikeB: how did you let 3-core utilize ponder?
I don't know of GUI that allow one side use ponder and the other not. Like a sort of handicap.
One solution is to disable ponder in the stockfish source code if that is a possible.
Re: How many elo do ponder equals to
Posted: Sun Jun 23, 2019 8:22 am
by hgm
WinBoard offers a work-around for this kind of things: the commands 'new' and 'computer', sent at the start of each game, are actually configurable strings on a per-engine basis, through the options /firstInitString and /firstComputerString. You could include commands to switch ponder on or off in those, to overrule any earlier commands for that normally sent by the GUI. I think especially the computre string is sent after anything else, at game start.