Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Distributed Computing / GPUs
- - By RangeFinder (**) Date 2009-11-10 03:57
I wonder if it would be possible or interesting to have a distributed computing option in Rybka one day? 

Almost like playing a video game over the Internet you would have the option to either “host” or “join” an analysis session.  The client that was hosting the session would dictate the  game or position that was being analyzed.  Helper clients would receive chunks of information to analyze:  specific lines, variations, or moves.  The host client might even make intelligent decisions about what to “outsource” depending on how many helper clients were connected and their speeds. 

On a small scale, I could see someone using several home computers to more thoroughly analyze a game overnight.  On a grander scale, I could imagine a site like TWIC or ChessOK hosting a session to analyze an important tournament game that was currently in progress.  Everyone following the game live could join the session and offer their computing power. 

Or even better maybe the Rybka Cluster could involve into Distributed Rybka where all the Rybka fans help out during the next computer championship.   ^.^

On a similar note, Folding@Home hit 8.13 PFLOPS as of Nov 1, 2009.  There aren’t enough players to do something like that with chess.  However, one of the things that has made the Folding@Home project so successful is the use of PS3s and GPUs.  Is there any possibility of a future version of Rybka using GPUs? 

GPU-power and/or distributed computing would offer a huge leap in analysis speed.
Parent - By M ANSARI (*****) Date 2009-11-10 07:22 Edited 2009-11-10 07:25
I think this "sort of" has been tried with a Toga version, and I am not so sure how good the implementation was.  I don't think it would be too successful for a chess engine since an engine would depend a lot on comparing output and then of figuring out which output is more accurate (depending on the hardware used).  For a good algo to work for a chess engine, I would think you would need an extremely fast network or even better a shared memory system.  On non shared systems you would at least have to have pretty much equivalent hardware as an evaluation is only as strong as it's weakest link.  The latencies involved in a distributed network would be so great as to make this feat quickly reach a point of diminishing returns where if you put it on a graph, the performance would end up actually being worse than better.  However DC might have a very important application when it comes to chess openings and chess endings.  You would basically be doing a massive Monte Carlo database without the need of having strong hardware to do many many small tasks.  Ofcourse this would create a giant database that would have to be somewhere, but with todays extremely cheap large HDD it would not be too expensive I would think.  This definetely will be an application to watch in the future.
Parent - By albitex (***) Date 2009-11-11 21:16 Edited 2009-11-11 21:28
I have read really few on chess programs written in CUDA. (CUDA: environment of development GPU NVidia).
I believe is an interesting road for chessmen. But at least be necessary try.
" the planners,  are the very bad popularizers. "
Up Topic Rybka Support & Discussion / Rybka Discussion / Distributed Computing / GPUs

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill