Topic Rybka Support & Discussion / Rybka Discussion / A Zillion processors - Rybka Cluster - Doubtful Value
If I tell my neighbor who plays chess casually, and sometimes fires up his Fritz on his netbook, that my new computer is 50 times faster than his, he might say "WOW, that chess program is going to be so much stronger than mine."
I used to think so too. That ended, however, in the late 1980's. I certainly don't think anywhere like that now. After several decades, I know that a doubling, tripling, or even much more, of speed, does not necessarily have the the impact one might think. My friend, with hits woefully underpowered netbook--hey, that is faster than the fastest power user's computer from 2001. And, even as a strong player, it's more than enough to kill me. Even R2.2 on that thing is strong enough. 2.2 on Arena is way overkill for him.
Today, on infinite mode, does an R4 go 4 ply higher when it's in the 20's given 50 times more time? I don't know, but I don't think so. There are diminishing returns, and the real value for Aquarium users is the interactive IDEA. I really doubt that a decent interactive session with 200 core R4 is going to give much more value than one with 4 or 8 cores.
And I think most strong users know that.
I used to think so too. That ended, however, in the late 1980's. I certainly don't think anywhere like that now. After several decades, I know that a doubling, tripling, or even much more, of speed, does not necessarily have the the impact one might think. My friend, with hits woefully underpowered netbook--hey, that is faster than the fastest power user's computer from 2001. And, even as a strong player, it's more than enough to kill me. Even R2.2 on that thing is strong enough. 2.2 on Arena is way overkill for him.
Today, on infinite mode, does an R4 go 4 ply higher when it's in the 20's given 50 times more time? I don't know, but I don't think so. There are diminishing returns, and the real value for Aquarium users is the interactive IDEA. I really doubt that a decent interactive session with 200 core R4 is going to give much more value than one with 4 or 8 cores.
And I think most strong users know that.
The problem with 200 cores is that you run the real risk of convincing yourself that everything is drawn and giving your opponents much more credit than they deserve. As an IDeA user who frequently plays against opponents who I suspect aren't using IDeA, I often have the feeling that I'm making my life harder than it needs to be by searching too much. I'll be focusing on several different lines or key ideas, putting all kinds of effort into it and thoroughly examining multiple options only to have my opponent make some move that I discarded long ago because it was clearly inferior.
There's a lot to be said for a certain amount of uncertainty and lack of clarity. I worry that with too many cores, it will somehow be harder to make the decisions which create the type of complications that lead to winning positions. I'm probably not explaining myself very well but hopefully that makes sense to someone!
There's a lot to be said for a certain amount of uncertainty and lack of clarity. I worry that with too many cores, it will somehow be harder to make the decisions which create the type of complications that lead to winning positions. I'm probably not explaining myself very well but hopefully that makes sense to someone!
![[us]](/mwf/v12/flags/us.png "United States")
This makes sense from a practical competitive point of view. I think you are alluding to situations where a move that is slightly less correct, but would give more subjective winning chances, against weaker hardware that would not always find the correct refutation, would not be played by a 200 core cluster, leading to more draws. Lukas mentioned that he noticed a similar issue, (When I have a chance, I'll try to find and link to it) with endgame (TB driven) play where the cluster headed for draws too quickly.
> This makes sense from a practical competitive point of view. I think you are alluding to situations where a move that is slightly less correct, but would give more subjective winning chances, against weaker hardware that would not always find the correct refutation, would not be played by a 200 core cluster, leading to more draws. Lukas mentioned that he noticed a similar issue, (When I have a chance, I'll try to find and link to it) with endgame (TB driven) play where the cluster headed for draws too quickly.
Yes, you explained it much better than I did, thank you!
Philosophically speaking, if we assume that chess is a draw, which yields better results -- a nearly perfect search or the ability to predict how an opponent will respond with near certainty? From a purely practical point of view based on my own experiences in corr chess, I'm inclined to think the latter option would lead to better results though I realize that's a very thorny problem when you start to really contemplate it deeply!
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
I.e who would give better results.... Capablanca or Kasparov? Clear, simple and possibly the most correct or complex and murky and possibly slightly worse objectively. My money would be on Kasparov (or for that matter on a sober Alekhine who I think anyone from any era would have been wary of in a match).
However, I think there is merit in the cluster for analysing very complex opening ideas especially. Force it into a murky position and let it loose. I have been analysing some rarely played andrather complex lines over the last few days where various GMs opinions on the best move and the evaluations vary and some more computer grunt would have been excellent. Nothing to stop you picking the computer's third or fourth best move if you feel it leads to an unbalanced position with chances.
However, I think there is merit in the cluster for analysing very complex opening ideas especially. Force it into a murky position and let it loose. I have been analysing some rarely played andrather complex lines over the last few days where various GMs opinions on the best move and the evaluations vary and some more computer grunt would have been excellent. Nothing to stop you picking the computer's third or fourth best move if you feel it leads to an unbalanced position with chances.
![[de]](/mwf/v12/flags/de.png "Germany")
That is certainly a feeling I got too, when I was using IDeA...
It would be amusing to find two people each spending tons of money using the cluster to fight each other in correspondence, and each missing a fairly obvious (obvious, let's say, to an OTB 2300) line of play, because both players are fighting a battle on another planet called Cluster. Amusing, but not funny.
Which leads to the next question(s):
1) How does one assure confidentiality of analysis? How does one know that an opening set of moves won't be available to someone else?
2) What happens to analysis after the lease period is over? Does the analysis get saved to a Really Big (tm) IDEA base? How would someone know that it is not happening?
Which leads to the next question(s):
1) How does one assure confidentiality of analysis? How does one know that an opening set of moves won't be available to someone else?
2) What happens to analysis after the lease period is over? Does the analysis get saved to a Really Big (tm) IDEA base? How would someone know that it is not happening?
1) How does one assure confidentiality of analysis? How does one know that an opening set of moves won't be available to someone else?
2) What happens to analysis after the lease period is over? Does the analysis get saved to a Really Big (tm) IDEA base? How would someone know that it is not happening?
These are some of my concerns about the cluster as well.
2) What happens to analysis after the lease period is over? Does the analysis get saved to a Really Big (tm) IDEA base? How would someone know that it is not happening?
These are some of my concerns about the cluster as well.
Mind you, this is not a concern of mine, as I am not anywhere near the strength that it would make a difference.
However, I would imagine that the target market might have an interest in it.
I mean, GM's don't even play their TN's in games on ICS, even when it's a speed game.
However, I would imagine that the target market might have an interest in it.
I mean, GM's don't even play their TN's in games on ICS, even when it's a speed game.
![[pl]](/mwf/v12/flags/pl.png "Poland")
The analysis (currently) isn't saved at all. Later we'll add some options for this.
Of course we give full confidentiality, this is in the announcement. It's the same as using yahoo or google email.
Vas
Of course we give full confidentiality, this is in the announcement. It's the same as using yahoo or google email.
Vas
Well I believe you in terms of confidentiality. But in the eyes of top GM users it is will not be perceived as the same as yahoo mail etc
These mail systems have a large number of users, largely anonymous, most mail being of no interest to anyone else much. It would be hard for the providers to do anything "interesting" with the contents of most emails even if they wanted to.
Your cluster will have a relatively small number of users especially to begin with. The contents of GM opening analysis will be of great interest to every other user of the cluster. The analysis would be highly exploitable. Add to that the fact that some top GMs are notoriously paranoid and you may have a trust issue.
A formal confidentiality agreement that you send on an individual basis to GMs might help the twicthier of users. I know that logically it may be no different to a general terms of usage statement, but I have been in similar situations in the commercial world before and on several occasions where general terms did not satisfy someone a personally signed individual agreement did. I am not suggesting that this is something you would do for everyone, just a few key people as you get started and build the reputation of the cluster service.
Just a thought :)
These mail systems have a large number of users, largely anonymous, most mail being of no interest to anyone else much. It would be hard for the providers to do anything "interesting" with the contents of most emails even if they wanted to.
Your cluster will have a relatively small number of users especially to begin with. The contents of GM opening analysis will be of great interest to every other user of the cluster. The analysis would be highly exploitable. Add to that the fact that some top GMs are notoriously paranoid and you may have a trust issue.
A formal confidentiality agreement that you send on an individual basis to GMs might help the twicthier of users. I know that logically it may be no different to a general terms of usage statement, but I have been in similar situations in the commercial world before and on several occasions where general terms did not satisfy someone a personally signed individual agreement did. I am not suggesting that this is something you would do for everyone, just a few key people as you get started and build the reputation of the cluster service.
Just a thought :)
I don't play a lawyer on TV, but I am one in real life.
Since Cluster is offered on the basis of "[we] make no warranties of any kind, either express or implied, including but not limited to the implied warranties of performance, merchantability, satisfaction, or fitness for a particular purpose," I can't think of much that could occur during the rental of Cluster that might be unfortunate that cannot be included within that disclaimer. That one is rather broad, and while I've seen more detailed ones in my all too long lifetime, I dare say this one is hard to fight. Just make sure, Vas, that you feature it more prominently, to help ensure that the merchantability disclaimer is enforceable. Some U.S. states get a tad ticklish about that one.
Some folks might want to know what the above says. Part of what it says is as follows: No guarantee it will do something specifically desired by the buyer. Also, no guarantee that it will reasonably conform to an ordinary buyer's expectations.
I think that just about does it.
Since Cluster is offered on the basis of "[we] make no warranties of any kind, either express or implied, including but not limited to the implied warranties of performance, merchantability, satisfaction, or fitness for a particular purpose," I can't think of much that could occur during the rental of Cluster that might be unfortunate that cannot be included within that disclaimer. That one is rather broad, and while I've seen more detailed ones in my all too long lifetime, I dare say this one is hard to fight. Just make sure, Vas, that you feature it more prominently, to help ensure that the merchantability disclaimer is enforceable. Some U.S. states get a tad ticklish about that one.
Some folks might want to know what the above says. Part of what it says is as follows: No guarantee it will do something specifically desired by the buyer. Also, no guarantee that it will reasonably conform to an ordinary buyer's expectations.
I think that just about does it.
> make sure, Vas, that you feature it more prominently, to help ensure that the merchantability disclaimer is enforceable. Some U.S. states get a tad ticklish about that one.
I'm sure Felix will have it front and center on his web site!
Vas
Regular exactly explained the difference between chess-knowledge and yahoo-mails...
Alex
Alex
> A formal confidentiality agreement that you send on an individual basis to GMs might help the twicthier of users.
Yes, we could do this if it becomes an issue. So far it hasn't been - none of our customers or our candidate customers complained about it.
BTW - the contents of google mail would have tremendous value. There would be a pretty big market for it, IMHO (and some pretty big consequences!
).Vas
How does the cluster work anyway? What is the difference between IDeA analysis?
And what's advantages of each or what's better overall (supposed engine non clustered is equal); 40 cores cluster rybka (10xquad), 40 cores IDeA (10xquad), or 40 cores (on one machine) engine ?
And what's advantages of each or what's better overall (supposed engine non clustered is equal); 40 cores cluster rybka (10xquad), 40 cores IDeA (10xquad), or 40 cores (on one machine) engine ?
#1 and #3 are very close to equal.
Vas
Vas
![[is]](/mwf/v12/flags/is.png "Iceland")
> thoroughly examining multiple options only to have my opponent make some move that I discarded long ago because it was clearly inferior.
Yes, this seems to be rather common when playing correspondence games with IDeA, but I don't complain
![[mx]](/mwf/v12/flags/mx.png "Mexico")
> I'll be focusing on several different lines or key ideas, putting all kinds of effort into it and thoroughly examining multiple options only to have my opponent make some move that I discarded long ago because it was clearly inferior.
Don't you mean "that I thought was clearly inferior, but was as good as the other key moves after IDEA focused on it"?
> Don't you mean "that I thought was clearly inferior, but was as good as the other key moves after IDEA focused on it"?
No. That's why I win two games for every game I draw.
Usually, such moves are clearly inferior but, for whatever reason, don't appear that way to the lazy correspondence player. It's rare to be surprised by a move that ends up being good for an opponent. I'm very, very thorough in my corr games, often using weeks of clock time in critical positions and taking full advantage of accumulated time. 40 days every 10 moves cumulative adds up! It's quite different than the games played here in the forums where games are over within months instead of years.
I think this question shows a common misunderstanding of how IDeA is used by correspondence players. If they run into surprises like the one you describe, they need to improve their analysis methods. In my limited experience the more common surprise is the opponent playing a suboptimal move. Also note that IDeA analysts usually combine different types of analysis (although I was practically only using IDeA), including infinite analysis, to check variations, generate alternatives etc.
I have yet to face real trouble against an IDEA opponent, the only time I lost against one (turbojuice) was when I tried to spice the game with an exchange sacrifice blunder that was heavily punished, but with moves like that he could have beaten me even if he didn't use IDEA. IDEA users usually make it very clear that they're at the edge of technology by using it, so one usually knows when facing one.
Now, Moz is talking about something else entirely, several years per game? I can see that as radically different from the Blitz Correspondence games that I play (I could even call the games I play Donal Duck time controls! ), my mind boggles just trying to think about an analysis method by an IDEA user applied for over a week on a single game position...
I recall 64x told me IDEA was super effective when used interactively to guide it and then letting it run for 30 hours, or something, I thought a honor system challenge of 1 hour per move could be in order: IDEA user Vs. non-IDEA user, no pondering, on each player's time to move they have 1 hour to analyze and decide on a move, and after it is over they make that move. They can start that hour at any time (no time limit). Could be interesting.
Now, Moz is talking about something else entirely, several years per game?
I can see that as radically different from the Blitz Correspondence games that I play (I could even call the games I play Donal Duck time controls! ), my mind boggles just trying to think about an analysis method by an IDEA user applied for over a week on a single game position...I recall 64x told me IDEA was super effective when used interactively to guide it and then letting it run for 30 hours, or something, I thought a honor system challenge of 1 hour per move could be in order: IDEA user Vs. non-IDEA user, no pondering, on each player's time to move they have 1 hour to analyze and decide on a move, and after it is over they make that move. They can start that hour at any time (no time limit). Could be interesting.
![[ar]](/mwf/v12/flags/ar.png "Argentina")
You can play blitz in playchess and see if the players "IDeA" that say they have good or bad results.
I challenge you to play a game advanced 30 minutes.
http://www.ficgs.com/user_page.php?page=entry_list&list=ficgs_chess_advanced
and we will see if it is only "IDeA" and technology. We can play in FICGS.
Regards
Pato Donald
I challenge you to play a game advanced 30 minutes.
http://www.ficgs.com/user_page.php?page=entry_list&list=ficgs_chess_advanced
and we will see if it is only "IDeA" and technology. We can play in FICGS.
Regards
Pato Donald
hehehe. hi ruben. i can feel that dual xeon's strength. hehehe. wink
Not what I had in mind, with your huge hardware, how much could you do in 30 minutes? The equivalent of 6 hours in my rig even if we did exactly the same interactions? Or more? (or 12 hours since you could leave IDEA running while it's my time to move, thus doubling your time).
I don't know. I think the most important factor influencing ELO (besides the engine itself) is still calculating power. You may get it by speed and/or by more cores. Of course Fritz 6 on my old notebook kills me as reliable as Lukas' cluster monster would - but that's not the question. I'm playing CC, and in CC even the slightest unaccuracy might be enough to decide the game. So, being able to look ahead one or two plies deeper may be all you need, and thus there is no doubt about the potential use of the cluster for me. There are several considerations, though, that speak against the cluster, it's pricing scheme being only one of them. So I guess I rather would accept Lukas as the world's strongest CC playing entity than buying cluster time (even with a much better pricing scheme)...
The main announcement page for the cluster rental - http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=20058;pg=1 - states:
Finally, the Rybka Cluster has a highly developed and fully configurable notion of draw avoidance. This is useful in computer chess tournaments, to make sure that the cluster's ultra-deep searches don't result in overly cautious moves leading to too many draws. Our clients have also found the feature useful for preparing opening variations which may be objectively equal but which yield good practical winning chances.
I am noticing now that the uci "contempt" option that was available in Rybka 3 is not available in Rybka 4. For the promised Rybka 4 patch, can we get all of these options, or at least the contempt option back.
Finally, the Rybka Cluster has a highly developed and fully configurable notion of draw avoidance. This is useful in computer chess tournaments, to make sure that the cluster's ultra-deep searches don't result in overly cautious moves leading to too many draws. Our clients have also found the feature useful for preparing opening variations which may be objectively equal but which yield good practical winning chances.
I am noticing now that the uci "contempt" option that was available in Rybka 3 is not available in Rybka 4. For the promised Rybka 4 patch, can we get all of these options, or at least the contempt option back.
![[fi]](/mwf/v12/flags/fi.png "Finland")
The piece values are the new Contempt.
Topic Rybka Support & Discussion / Rybka Discussion / A Zillion processors - Rybka Cluster - Doubtful Value
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill