Author: Robert Hyatt
Date: 14:18:34 12/22/99
Go up one level in this thread
On December 22, 1999 at 15:10:51, Dann Corbit wrote: >On December 22, 1999 at 14:05:42, Robert Hyatt wrote: >[snip[ >>100 times as many is not easy. It is not easy to understand, but the SP and >>the chess processors fit together very nicely. If you make the SP 10x faster >>without changing the chess processors, you go _no_ deeper. Because the SP will >>overrun the chess processors and they have to back off one ply to keep up. Net >>gain: 0. If you make the SP 10x faster, you can add 10x more chess processors >>and _really_ pick up the pace 10x. The SP will be going 1 ply deeper, the >>chess processors can stay at the current 4 ply search, because there are 10x >>as many of them to take the 10x increased data from the SP. > >From: >http://www.rs6000.ibm.com/hardware/largescale/SP/index.html > >You can now get SP's with Spec_Int base rate of 908 and Spec_fp base rate of >1760. I suspect that they can push special machines to a higher plateau. I >don't know the throughput of the machine that was used, but I suspect that new >machines can do much better. > >By the SP going 10x faster do you mean total machine throughput, or does the >clock rate have to increase by that much? You can do it either way. 10x as many chess processors require 10x as many positions to search to the normal (4 ply) depth. 10x the number of SP nodes, or 32 nodes each 10x faster than before would do the trick. Probably the latter would be more preferable for efficiency, but that would mean pretty dense processor boards for the chess processors...
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.