Goodbye Talkchess

[Albert Silver]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

[bob]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

Floating point or not is not the primary issue. Graphics devices have poor memory latency, they are designed to stream lots of data (high bandwidth) thru, because the primary function is to manipulate pixels as quickly as possible to update a graphical display. Chess engines don't step thru memory sequentially and run into ugly latency issues. Also the concept of a "cache" is foreign.

[Albert Silver]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

Floating point or not is not the primary issue. Graphics devices have poor memory latency, they are designed to stream lots of data (high bandwidth) thru, because the primary function is to manipulate pixels as quickly as possible to update a graphical display. Chess engines don't step thru memory sequentially and run into ugly latency issues. Also the concept of a "cache" is foreign.

Is there a way to combine the strengths of a CPU and a GPU in such a way that the gains outweigh the losses?

[Dann Corbit]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

Floating point or not is not the primary issue. Graphics devices have poor memory latency, they are designed to stream lots of data (high bandwidth) thru, because the primary function is to manipulate pixels as quickly as possible to update a graphical display. Chess engines don't step thru memory sequentially and run into ugly latency issues. Also the concept of a "cache" is foreign.

Is there a way to combine the strengths of a CPU and a GPU in such a way that the gains outweigh the losses?

There are several efforts at chess on a GPU farm, but there are many big problems. Memory transfer is one, and complete lack of recursion is another.

[Sam Hull]

Point was, the old board treated everybody the same. If you could see anything, you could see everything...

Point is, the board works now just like it did then for anyone who wants to legitimately sign up. These forums exist as a service for members and there is no obligation whatsoever to provide/expose any contributed material to the general public. The fact that parts of the board can now be viewed by non-members is simply a bonus.

Hint: read, parse, understand. You will _never_ be able to produce _any_ quote that said warez links are OK. True Warez-type links. True copied/stolen/cracked software. I challenge you to find _any_ such quote. My only comments have been related to Fruit/Rybka/Robo, which is quite a different matter altogether.

Clue: if you believe some links should be deleted and others shouldn't, somebody is going to have to decide which are which; i.e., moderators. That's exactly what I stated above if you read, parse, and understand it.

Having grown up in the US, having lived here for all of my 62 years, from having my grand father fight in WW1, by father in WW2, all in the name of freedom and the US constitution, yes, I believe in "Innocent until proven guilty." Those proven guilty have met with swift action over the years with respect to clones. Those unproven were left alone. Whether it be from claims made at an ICGA event, an ACM event, or here on CCC.

Disallowing download links to questionable software isn't unpatriotic and it isn't equivalent to a criminal conviction. Airport security does not allow items on airplanes that could be used as a weapon; does that imply that anyone who has to surrender his pocketknife is a felon? No.

... We originally _did_ disallow any links to IP* and family if you look back. And waited on "reasonable proof". _NONE_ was ever offered. After several months, we decided that "innocent until proven guilty" was the rational course. I think the decision was right then, and it is still right today. Not what has been done with the blatant attempt to hide the discussions since it seems obvious they can not be completely quashed.

In other words, your mod team made the call on whether or not to allow links, and you initially deleted questionable links. I assume at the time you didn't think deleting those links was "stupid." The current team is doing the same thing - you just disagree with some of the calls, just as there has been disagreement with some of yours.

The notion that discussions available to every single member of this board are being "hidden" is silly. Was CCC "hiding" everything on the board for the first nine years of its existence?


-Sam-

[Dr.Wael Deeb]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

Floating point or not is not the primary issue. Graphics devices have poor memory latency, they are designed to stream lots of data (high bandwidth) thru, because the primary function is to manipulate pixels as quickly as possible to update a graphical display. Chess engines don't step thru memory sequentially and run into ugly latency issues. Also the concept of a "cache" is foreign.

Is there a way to combine the strengths of a CPU and a GPU in such a way that the gains outweigh the losses?

I wonder why this great idea hasn't benn realized yet....taking under consideration the powerfull GPUs that are in the market nowadays....but it looks like there are reasons for the lack of interest from the programers side....
Dr.D

[Thomas Mayer]

Hi Bob,

And so far, all we have for IP* is that the code _appears_ to be reverse-engineered. From what is unknown.

we definitely have a bit more then that, e.g. Gian-Carlo presented a lot more observations. Also we have the sayings of Don Daily & Larry Kaufmann - which you called real guys in one of your postings. They both clearly pointed out that they have seen enough to call it a Rybka derivative. With not accepting their foundings you clearly call them liars. Is that your real opinion about Larry & Don ?

Greets, Thomas

[bob]

Fabian's first iteration of Fruit that I saw was not particularly strong. Elo was a little over 2000. He obviously made rapid progress, though.

His eval function used floating point at one time. You won't see many like that.

Hmmm..... Does that mean it would be better suited to run on a GPU? I have seen many claim that while powerful, GPUs are not well-suited for chess engines, which is why no one really tries to make an uber-engine on it.

Floating point or not is not the primary issue. Graphics devices have poor memory latency, they are designed to stream lots of data (high bandwidth) thru, because the primary function is to manipulate pixels as quickly as possible to update a graphical display. Chess engines don't step thru memory sequentially and run into ugly latency issues. Also the concept of a "cache" is foreign.

Is there a way to combine the strengths of a CPU and a GPU in such a way that the gains outweigh the losses?

Everything we do is a compromise of sorts. For example, with multiple CPUs, we still have one path to memory and contention becomes an issue. AMD uses a NUMA architecture to provide more paths, at a cost of more latency. GPUs are slowly becoming more general purpose in nature, because of their power. So anything is possible. And the ultimate might one day be a "cluster on a single board" where we might have the equivalent of the cluster I test on but on a single board anyone can use, rather than needing 30 tons of AC to keep it cool and a huge room to hold the thing.

But their primary impetus is graphics, which is all about taking a large array of pixels and modifying them as quickly as possible so that one can draw smooth images that move and look natural. So sucking in 24 bits per pixel, modifying the entire thing in some form, and then spitting it back out into video/main memory so that it can be displayed, is what this is all about. High bandwidth, latency irrelevant, parallel computation important so that multiple pixels can be computed at once to improve the frame rate.

[Gian-Carlo Pascutto]

No idea, since I have not participated in one in a long while. I suppose someone _could_ if they saw fit, as it is a real issue.

Then perhaps that would be the way to settle this long festering issue once and for all?

How so? Do you think Vas would release his source to them? ICGA is not going to go through the effort of reverse-engineering to prove/disprove this.

When you join the tournament, you submit to the rules, which is that you must be able to provide sources to the TD.

At least in the past, the ICGA applied them (Graz, 2003). I am not entirely confident they would do so for Rybka, but let's still assume the rules are valid for everyone.

Assuming then that Rybka was derived from Fruit (I am operating in a pure what-if scenario here, I do not want to make any claim for or against this), the question is still if it would be recognizable after x years of fulltime development.

If you believe it is, you should complain the next time Rybka and Crafty are together in a tournament.

[Harvey Williamson]

the question is still if it would be recognizable after x years of fulltime development..

I think you are right it could be very hard to tell now. If there is a case to answer the challenge should have been made when Rybka 1 was released.