Hi All,
I just have a general question about computer chess programs like Rybka 4.1 and Houdini 1.5a.
I want to compare the ELO of them running on a modern dual core desktop (Intel dual core E8400 @ 3.0 Ghz),
versus the same chess program running on a huge 64 CPU Intel Xeon cluster (like the one which can be rented).
To use simple numbers, assuming their ELO is 3,250 on the Intel dual core E8400 @ 3.0 Ghz,
what would their ELO be on the Intel 64 CPU Xeon cluster?
Would the ELO maybe 3,400?
What I am trying to understand is that a supercomputer with 64 Intel Xeon CPU's running a chess program is only maybe 5%
higer ELO than a ordinary Intel Dual Core ordinary desktop. If this is true, it means a chess program's strength is mainly dependent on
the search and evaluation algorithms, and that exponentially faster hardware only gives a very small percentage ELO increase.
Thanks in advance for your insight into this.
Thanks,
Bill
I just have a general question about computer chess programs like Rybka 4.1 and Houdini 1.5a.
I want to compare the ELO of them running on a modern dual core desktop (Intel dual core E8400 @ 3.0 Ghz),
versus the same chess program running on a huge 64 CPU Intel Xeon cluster (like the one which can be rented).
To use simple numbers, assuming their ELO is 3,250 on the Intel dual core E8400 @ 3.0 Ghz,
what would their ELO be on the Intel 64 CPU Xeon cluster?
Would the ELO maybe 3,400?
What I am trying to understand is that a supercomputer with 64 Intel Xeon CPU's running a chess program is only maybe 5%
higer ELO than a ordinary Intel Dual Core ordinary desktop. If this is true, it means a chess program's strength is mainly dependent on
the search and evaluation algorithms, and that exponentially faster hardware only gives a very small percentage ELO increase.
Thanks in advance for your insight into this.
Thanks,
Bill
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
There is to my knowledge no data on that..... 
![[us]](/mwf/v12/flags/us.png "United States")
One thing you could try is to extrapolate a rough idea from the chess rating sites. For example CEGT does testing with different core counts of the same engine.
You could use something like Sedat's list of nodes/sec on Houdini/Rybka to generate a rough correlation between elo and nodes and then consider this when put towards larger hardware. You'd have to factor in diminishing returns with increased parallelism but at least you could get some idea.
One problem that exists though is that perhaps an increase in chess strength does not have a linear correlation with an increase in performance. It could be a punctuated equilibrium or something else that is less clear. It may be that at a certain depth in a certain subset of chess positions that an extra ply results in a large strength increase, while in another subset of positions at a certain depth an extra ply or more makes little difference.
This would mean that even with direct data we would have to take the results with a grain of salt. Yet I'd still rather have that than nothing. So far I haven't seen any such test ran but I would love to see the results.
You could use something like Sedat's list of nodes/sec on Houdini/Rybka to generate a rough correlation between elo and nodes and then consider this when put towards larger hardware. You'd have to factor in diminishing returns with increased parallelism but at least you could get some idea.
One problem that exists though is that perhaps an increase in chess strength does not have a linear correlation with an increase in performance. It could be a punctuated equilibrium or something else that is less clear. It may be that at a certain depth in a certain subset of chess positions that an extra ply results in a large strength increase, while in another subset of positions at a certain depth an extra ply or more makes little difference.
This would mean that even with direct data we would have to take the results with a grain of salt. Yet I'd still rather have that than nothing. So far I haven't seen any such test ran but I would love to see the results.
Thanks for your responses. I feel this is an important question, so that we can know whether its necessary to invest the money
in a chess playing Intel i7 six or twelve core machine, or is it enough to have an ordinary dual core PC and get almost idential ELO.
in a chess playing Intel i7 six or twelve core machine, or is it enough to have an ordinary dual core PC and get almost idential ELO.
Try to be clearer about what you want. Are you asking about a desktop c/p the Cluster or just 2 cores vs 6 or 12?
The cluster may scale very differently to a normal chess program and there's very little data on this apart from a few experts.
However, the relationship between different numbers of cores on a normal P.C and its effect of chess ratings is well known:
ROUGHLY (and anyone please correct me on this)
- for every doubling of SPEED (i.e. 2 Ghz to 4 Ghz) = +60 elo
- and for every doubling of CORES (i.e 2 cores to 4 cores both at 2 Ghz) = +40 elo
The cluster may scale very differently to a normal chess program and there's very little data on this apart from a few experts.
However, the relationship between different numbers of cores on a normal P.C and its effect of chess ratings is well known:
ROUGHLY (and anyone please correct me on this)
- for every doubling of SPEED (i.e. 2 Ghz to 4 Ghz) = +60 elo
- and for every doubling of CORES (i.e 2 cores to 4 cores both at 2 Ghz) = +40 elo
So Houdini 2.0 2CPU is 3048 elo on the CEGT 40/20 list.
Houdini 2.0 4CPU is 3096 elo. So that's +48 elo doubling the cores, pretty close to 40.
Although the CPUs are not the same, with the same scaling if we were to play on an SGI Altix UV 100 with 256 cores and Windows Server 2008, at +40 per doubling and starting at 3048 elo we could get up to +280 elo, or 3328 elo :-P
Houdini 2.0 4CPU is 3096 elo. So that's +48 elo doubling the cores, pretty close to 40.
Although the CPUs are not the same, with the same scaling if we were to play on an SGI Altix UV 100 with 256 cores and Windows Server 2008, at +40 per doubling and starting at 3048 elo we could get up to +280 elo, or 3328 elo :-P
![[de]](/mwf/v12/flags/de.png "Germany")
Try my Excel file:
http://www.sendspace.com/file/cci8w8
In case the cluster has 3 GHz too and based on the facts that doubling the cores adds 1.4x on nodes/s and doubling the nodes adds 70 Elo points, it would have Elo 3420.
Which doesn't make the chess engine smarter strategy-wise. Humans could still win against 4000 Elo engines in closed positions.
http://www.sendspace.com/file/cci8w8
In case the cluster has 3 GHz too and based on the facts that doubling the cores adds 1.4x on nodes/s and doubling the nodes adds 70 Elo points, it would have Elo 3420.
Which doesn't make the chess engine smarter strategy-wise. Humans could still win against 4000 Elo engines in closed positions.
![[ca]](/mwf/v12/flags/ca.png "Canada")
An old supercomputer I have an account on has 60 cores available and when I tried Crafty on all cores it was a lot better than my crappy X2 CPU could ever muster.
Sooner or later I will have a better CPU
Sooner or later I will have a better CPU
>on a modern dual core desktop (Intel dual core E8400 @ 3.0 Ghz)
That's certainly no modern computer.
The cluster consists of 12 and 8 core computers. If we compare that E8400 to a 5 x 12 core cluster we can calculate like this:
12 cores compared to 2 cores is effectively ~4 times faster for Rybka. Nehalem / Westmere is 30% faster than Core 2 - so we get a factor of 5.2 for one 12 core. Doubling of speed adds 50 - 70 Elo points. A factor of 5.2 is ~2.4 doublings, so going form E8400 to 12 core Westmere adds ~ 120 - 165 Elo points. A cluster of 5 of these computers is roughly 100 Elo points stronger, so you'll get something in the range of + 240 Elo points. The cluster software is stronger than Rybka 4 which adds even more Elo.
Thanks for the info. So if the chess program remains the same on both the ordinary computer and clustered super computer,
we could expect to see about a 240 ELO point increase on the cluster. This becomes more interesting if we look at the ELO
increase in percentage terms: (240 ELO increase on cluster) / (3250 ELO on E8400 dual core) = 7% ELO increase on cluster.
So chess program strength is almost entirely based on its search and evaluation algorithms.
Running it on exponentially faster 60+ core cluster Super Computers only increases playing stress by a small 7% margin.
we could expect to see about a 240 ELO point increase on the cluster. This becomes more interesting if we look at the ELO
increase in percentage terms: (240 ELO increase on cluster) / (3250 ELO on E8400 dual core) = 7% ELO increase on cluster.
So chess program strength is almost entirely based on its search and evaluation algorithms.
Running it on exponentially faster 60+ core cluster Super Computers only increases playing stress by a small 7% margin.
![[se]](/mwf/v12/flags/se.png "Sweden")
>"So chess program strength is almost entirely based on its search and evaluation algorithms.">
Correct, I have never seen any radical departure in style of play regardless of core increase.
Using a percentage on Elo points doesn't work. A chess entity which is 240 Elo points stronger gets roughly 80% of the points in a match. That's far more than an increase of 7%.
I'm wondering at what point this will begin to plateau, as the number of draws increases and the number of positions played incorrectly decreases so dramatically.
I mean, aren't we there already? Say if Rybka is 3000 elo on one list, and the cluster makes 3240 elo. Is there such thing as 3480 elo that would pick up 80% of the points against the cluster? How far does it go? Is there a tried and true 4000 elo?
I mean, aren't we there already? Say if Rybka is 3000 elo on one list, and the cluster makes 3240 elo. Is there such thing as 3480 elo that would pick up 80% of the points against the cluster? How far does it go? Is there a tried and true 4000 elo?
>So chess program strength is almost entirely based on its search and evaluation algorithms.
What else would it be based on besides search and evaluation? You mean software vs. hardware?
In my opinion hardware is the surest way to improve playing strength. It is much more difficult to find 200 elo in software than in hardware!
![[tr]](/mwf/v12/flags/tr.png "Turkey")
Just my 2 cents over this issue,
In case of doubling the processor speed (depending on conditions-engine,time control,book,hashtables vs), we can get extra approx.40-80 Elo points
For example i have already a such measuring table...
I mean, Rybka 4.1 6 core (under SCCT Auto232 conditions) performed 65 Elo better than Rybka 4.1 4 core
http://www.sedatcanbaz.com/chess/ratings/scct-auto232/
Note:i7 980X / i7 970 @4.33GHz machines are approx.2 times faster than i7 920 @3.0GHz/QX9650 @3.66GHz
Best,
Sedat
In case of doubling the processor speed (depending on conditions-engine,time control,book,hashtables vs), we can get extra approx.40-80 Elo points
For example i have already a such measuring table...
I mean, Rybka 4.1 6 core (under SCCT Auto232 conditions) performed 65 Elo better than Rybka 4.1 4 core
http://www.sedatcanbaz.com/chess/ratings/scct-auto232/
Note:i7 980X / i7 970 @4.33GHz machines are approx.2 times faster than i7 920 @3.0GHz/QX9650 @3.66GHz
Rank Name Elo + - games score oppo. draws
6 Deep Rybka 4.1 x64 6c 3358 15 15 1363 59% 3305 54%
15 Deep Rybka 4.1 x64 4c 3293 13 14 1602 47% 3314 56%
Rank Name Elo + - games score oppo. draws
12 Fire 2.2 xTreme x64 6c 3311 23 22 530 52% 3297 63%
18 Fire 2.2 xTreme x64 4c 3275 16 16 1165 41% 3328 59% Best,
Sedat
from everything I have read it seems as if an engine running on an i5 cpu (ordinary, commonplace computer) is more or less just as strong as it running on a super computer.
the problem I have with this is that I frequently see people post games on the internet and say that they have beaten houdini and rybka. how can all these random chess enthusiasts beat 3000 elo engines, but super-gms can't? doesn't make much sense to me.
the problem I have with this is that I frequently see people post games on the internet and say that they have beaten houdini and rybka. how can all these random chess enthusiasts beat 3000 elo engines, but super-gms can't? doesn't make much sense to me.
![[mx]](/mwf/v12/flags/mx.png "Mexico")
Are those the same dudes that sell penis enlargement pills?
since the measure is the cm between the 'top' and the 'bottom',benzodiazepines pills enlarges penis without pain...
is easy to prove it:
1)measure the length of the penis
2)give ot the guy 40mg of oral benzodiazepines pills
3)after it works,and the guy is unconcious,take a scalpel
4)amputate the penis at the half of his length
5)take the amputated part,and go to the next room
6)measure now the distance between the bottom of the penis,and his head
7)compare both measures
8)if the second measure is greater than the first,voilá!you enlarged the penis with pills (with a bit help of the scalpel )
after that,you can wake up the guy and tell him his penis was enlarged
p.s.please,take a pic of his face at the moment he hear that from you,and upload to facebook
is easy to prove it:
1)measure the length of the penis
2)give ot the guy 40mg of oral benzodiazepines pills
3)after it works,and the guy is unconcious,take a scalpel
4)amputate the penis at the half of his length
5)take the amputated part,and go to the next room
6)measure now the distance between the bottom of the penis,and his head
7)compare both measures
8)if the second measure is greater than the first,voilá!you enlarged the penis with pills
(with a bit help of the scalpel )after that,you can wake up the guy and tell him his penis was enlarged
p.s.please,take a pic of his face at the moment he hear that from you,and upload to facebook
I don't have time for these jokes right now....I have to pick up a 2 million dollar check sent to me by a high ranking Nigerian government official.
im sorry,but i dont think that an scenario like you describe,picking up a 2 million dollat check,can affect your sense of humour in a negative way like that
es que siempre me llegan estos emails avisandome que alguien en Nigeria necesita mi ayuda y que me va a pagar tantos milliones de dolares por mi ayuda, es casi tan increible como un jugador regular facilmente ganando a Rybka
i forgot put the smiley in my latest message
a mi tambien me llegan,y no solo eso,sino que entro en todas las paginas que envian spam y mensajes de esos,y los colecciono
tengo en una cuenta mas de 15.000 de esos mensajes...eso si,el que pica en ellos,es que es idiota;parace que los haya traducido un mono retrasado mental
a mi tambien me llegan,y no solo eso,sino que entro en todas las paginas que envian spam y mensajes de esos,y los colecciono
tengo en una cuenta mas de 15.000 de esos mensajes...eso si,el que pica en ellos,es que es idiota;parace que los haya traducido un mono retrasado mental
> i forgot put the smiley in my latest message
que bueno, temi que te habia ofendido mi amigo.
pero fijate que deberian estan ganando algo de eso, si no, no seguirian haciendolo.
no,no me ofencio,es que se me olvido poner el smiley...quise decir que si yo fuese a cobrar esos 2 millones,estaria contento y una cara asi
bueno,siempre hay algun estupido que pica en esas cosas,y no lo entiendo...
a mi el otro dia me llego un email de parte de banesto,que es un banco de aqui en España,diciendo que habia problemas en mi cuenta,que necesitaba entrar y dar unos datos...
hasta ahi,aunque huele a timo,pareceria cas correcto,la pagina estaba muy bien imitada,etc...
solo que no tengo cuenta en ese banco
bueno,siempre hay algun estupido que pica en esas cosas,y no lo entiendo...
a mi el otro dia me llego un email de parte de banesto,que es un banco de aqui en España,diciendo que habia problemas en mi cuenta,que necesitaba entrar y dar unos datos...
hasta ahi,aunque huele a timo,pareceria cas correcto,la pagina estaba muy bien imitada,etc...
solo que no tengo cuenta en ese banco
> solo que no tengo cuenta en ese banco
hay gente que es increiblemente tonta,la verdad...
yo creo que algunos se merecen que los estafen,por avariciosos;quieren conseguir el dinero de una 'herencia' de 15.000.000 dolares a cambio de 1500,y por caraduras todavia les pasa poco
yo creo que algunos se merecen que los estafen,por avariciosos;quieren conseguir el dinero de una 'herencia' de 15.000.000 dolares a cambio de 1500,y por caraduras todavia les pasa poco
> from everything I have read it seems as if an engine running on an i5 cpu (ordinary, commonplace computer) is more or less just as strong as it running on a super computer.
Wrong.
well that is what I thought, that the engines running on i5's would be significantly weaker. but if you read this thread it seems like bumping up the cores to 16 will only increase the engine's strength by a couple hundred points, maybe 300-400? no?
> r. but if you read this thread it seems like bumping up the cores to 16 will only increase the engine's strength by a couple hundred points, maybe 300-400? no?
So you don't think a 300 or 400 Elo gain is significant ?
well I mean if it's a difference between 3000 and 3400...how significant is this if you are a human?
in any case, so you are saying that this is essentially correct that rybka or houdini running on an i5 is around 3000 elo?
I have a little trouble believing this given what I have seen and what players have told me. one FM stated that beating rybka (on an ordinary computer) isn't all that difficult. is everyone lying? I would think if rybka/houdini were around 3000 elo running on an i5, then no one would be able to defeat it. if someone did beat it they would invariably have to be a super gm.
in any case, so you are saying that this is essentially correct that rybka or houdini running on an i5 is around 3000 elo?
I have a little trouble believing this given what I have seen and what players have told me. one FM stated that beating rybka (on an ordinary computer) isn't all that difficult. is everyone lying? I would think if rybka/houdini were around 3000 elo running on an i5, then no one would be able to defeat it. if someone did beat it they would invariably have to be a super gm.
> in any case, so you are saying that this is essentially correct that rybka or houdini running on an i5 is around 3000 elo?
No, where did I say that ?
> one FM stated that beating rybka (on an ordinary computer) isn't all that difficult.
Ask Kramnik how difficult it was to beat Deep Fritz.
A 300 or 400 Elo gain would not be significant enough to allow me to beat Carlsen.
>from everything I have read it seems as if an engine running on an i5 cpu (ordinary, commonplace computer) is more or less just as strong as it running on a super computer.
What are you reading? A guide to an alternate universe where everything is backwards?
what? everything written on this thread clearly gives the impression that the difference between the elo of an engine running on something like a i5 or i7 and the same engine running on 16 cores is a matter of at most a few hundred elo points. so this would mean if houdini is 3300 or whatever on 16 cores it would still be 3000 or around on an i5 or i7. no?
so, how are all these chess enthusiasts beating a 3000 elo engine when grandmasters cannot? and I have also heard FM's claim that beating rybka and houdini on their laptop or home computer isn't very difficult. so it would seem to me that either houdini or rybka running on an i5 or i7 has to be significantly weaker than on the advanced hardware used in computer chess tournaments or something else is going on.
so, how are all these chess enthusiasts beating a 3000 elo engine when grandmasters cannot? and I have also heard FM's claim that beating rybka and houdini on their laptop or home computer isn't very difficult. so it would seem to me that either houdini or rybka running on an i5 or i7 has to be significantly weaker than on the advanced hardware used in computer chess tournaments or something else is going on.
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill