Can this be effectively used for MC or EGTB on the fly

[M ANSARI]

I guess some of you have heard of Tilera. Apparently they have a product that has up to 100 processors per CPU and you can insert it into existing PCI 2.0 slots. It is integer based and according to the website can run 100 instances of independent OS sessions or of course one MP OS. The price of each card is somewhere between $400 to $1000, and I think there are some motherboards that have 8 PCI 2.0 slots. Imagine 8 of these cards on one motherboard.

I am not really thinking if these are good for chess as they run at only 1.5 Ghz ... but they must be like a perfect platform to run highly parallelized Monte Carlo code. I think present MC code is very primitive in that it is using the existing engine executable (in case of Rybka at least) to generate the games rather than an optimized engine designed for MC. So far it seems MC implementation has been almost an afterthought rather a concentrated effort and I think there might really be something there for chess. Of course when you think of Monte Carlo then "on the fly" EGTB generation is also possible. Do you think this card (or a set of them) can be usefully configured as an endgame module add on card? For analysis it would be really great tool, but I think it might also be useful to an existing engine setup in game play as well.

http://www.theinquirer.net/inquirer/new ... -core-chip

http://www.slashgear.com/tilera-announc ... r-2661746/

[Gian-Carlo Pascutto]

See this thread about the nonsense of pulling MC into such a discussion:

http://talkchess.com/forum/viewtopic.php?t=36787

The hardware would be interesting for a normal chessprogram, if you could actually get it. But I certainly didn't manage to buy a board... This was discussed in the Rybka forum too.

[M ANSARI]

I am surprised you were not able to purchase a board. I managed to talk to some of the Tilera people some time back, and was able to get some contacts. I didn't follow it up, but felt that they were for real. They seem to have several good solid clients and I am pretty sure they are not bogus.

You don't think this would be useful for MC or EGTB generation? 100 cores per PCI 2.0 slot seem like a tremendous processing capability, and MC and EGTB generation seem like they would be very efficiently parallelized.

[Gian-Carlo Pascutto]

You don't think this would be useful for MC or EGTB generation? 100 cores per PCI 2.0 slot seem like a tremendous processing capability, and MC and EGTB generation seem like they would be very efficiently parallelized.

As for the illusion that MC suddenly makes parallel scaling easier, again, read the thread I linked.

[bob]

You don't think this would be useful for MC or EGTB generation? 100 cores per PCI 2.0 slot seem like a tremendous processing capability, and MC and EGTB generation seem like they would be very efficiently parallelized.

As for the illusion that MC suddenly makes parallel scaling easier, again, read the thread I linked.

It does make it _easier_. But it also is much less effective than traditional parallel search. When the word "scaling" enters the discussion, things often take unexpected twists and turns.

If by "easier" you meant "easier to get decent speedup" then I certainly agree, but with the word "scaling" involved, that has several meanings..

[Gian-Carlo Pascutto]

The nuances are already explored in the linked thread, no need to do so again here, unless there's some actual new info.

[bob]

The nuances are already explored in the linked thread, no need to do so again here, unless there's some actual new info.

I was just commenting on the context in this thread. MC certainly scales well and easily. This is one of several programming assignments I give in a parallel programming course I teach (not chess, but MC in general) Whether it produces decent results or not is a completely different issue, applied to chess. I'm not much of a fan of statistically sampling many very short games to predict which move is best. One has to hope that quantity (number of games) offsets lack of quality (very shallow searches will miss deep tactics).