Author: Uri Blass
Date: 16:36:00 01/20/06
Go up one level in this thread
On January 20, 2006 at 17:58:42, Dann Corbit wrote: >On January 20, 2006 at 17:27:10, Uri Blass wrote: > >>On January 20, 2006 at 17:07:09, Dann Corbit wrote: >> >>>Consider Deep Shredder or Deep Fritz on a 4 way box with dual core CPUs >>>8x CPU speed would mean 150 Elo (with some SMP loss). >>> >>>That is a system runnable today. >>> >>>Now, imagine an 8 way box with 4 cores each (probably next year) >>>That would be another 100 Elo. >>> >>>If we lost one whole doubling from SMP loss, it would still be +200 Elo. >>> >>>But I think that on a single CPU, Rybka is probably the strongest. >> >>I think that there is a diminishing returns so I am not sure if your estimate is >>correct. > >I allowed for a 50% SMP loss for the +200 Elo estimate. >I think that even with 32 cores, that should be achievable. >{16x speedup for 32 cores} > >But maybe not. > >I did see an experiment where linear speedup was achieved for a large number of >processors. > >In order to accomplish it, the tree was simply expanded and processors were >given leaf nodes of the expanded tree to work on. > >And so, with 20 CPUs at the root position, you would give each CPU one of the 20 >possible first moves to work on. And if you had 400 CPUs, you could expand 2 >levels. > >It is also what the chessbrain project does, according to my understanding. >Of course, they have a huge network latency to overcome, so it is not like SMP >at all. That is why I did not consider network based solutions for the >strongest possible chess engine. Note that I think about diminishing return from being twice faster and not only about diminishing returns from more processors. Uri
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.