Author: Robert Hyatt
Date: 18:49:55 11/20/00
Go up one level in this thread
On November 20, 2000 at 17:04:51, Bob Durrett wrote: >On November 20, 2000 at 15:08:05, Robert Hyatt wrote: > >>On November 20, 2000 at 13:53:55, walter irvin wrote: >> >>>i have been thinking about multi processors and computer chess .my idea must be >>>wrong because no one uses the processors in this way .lets use the example of 32 >>>cpu machine . >>> depth cpu used >>>1 ply 4 >>>2 ply 4 = assumes that ply 1 has been played >>>3 ply 4 = assumes that ply 2 has been played >>>4 ply 4 = assumed that ply 3 has been played >>>5 ply 4 = assumes that ply 4 has been played >>>6 ply 4 = assumes that ply 5 has been played >>>7 ply 4 = assumes that ply 6 has been played >>>8 ply 4 = assumes that ply 7 has been played the cpu at the deepest >>>depth determines score >>> >>>it seems to me with the cpu's used in this manner depth of search would be much >>>deeper and could eliminate not all but alot more of the horizon effect .imagine >>>if 64 cpu's were used???surely 3000 elo could be achieved ??? like i said there >>>must be a flaw because its not done this way .anyone know why?? >> >> >>The problem is that the CPUs near the root of the tree aren't doing anything, >>while the cpus near the tips are overloaded. The tree is highly non-uniform, >>plus you can't evaluate the difference between two moves at the root until the >>searches below those two moves are finished. >> >>And then there is alpha/beta... and its serial behavior > > >Not that I really fully understood what walter irvin was proposing here, but . . >. cannot at least part of his idea be salvaged here? Is there no part of his >idea that might be used somehow? The closest that has been done is a two-tiered search as in deep blue. To do one of their 15 ply searches, they searched plies 1-5 using one processor, then they split the tree into pieces and used 30 SP processors to handle those branches and search them to depth=10 (roughly). Then they handed the resulting positions off to the special-purpose chess hardware. They were about 30% efficient, according to their publications.
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.