Author: Russell Reagan
Date: 18:59:25 05/13/04
Go up one level in this thread
On May 13, 2004 at 20:35:39, Robert Hyatt wrote: >On May 13, 2004 at 20:02:16, Russell Reagan wrote: > >>On May 13, 2004 at 18:44:50, martin fierz wrote: >> >>>while this is all correct, remember that doubling the number of processors very >>>clearly has diminishing returns :-) >>> >>>while going from 1->4 is a 3.1 speed increase (i think we can trust this magic >>>number by now...), going from 4->16 will be a much smaller improvement. and of >>>the commercials, at least fritz is also capable of running on a 4-way box. don't >>>know too much about the others, but i guess that most can run on at least a >>>dual. >> >>On a NUMA machine, Crafty does much better than 3.1x on a quad. Here is one post >>by Bob where he gives 3.9x for a quad. I seem to recall seeing the number 3.98x >>for a quad, and reading Eugene saying that Crafty scaled almost linearly on a >>NUMA machine, but I couldn't find any posts indicating that, so I might be >>wrong. >> >>http://chessprogramming.org/cccsearch/ccc.php?art_id=345901 > >Wrong kind of speedup. > >You can measure raw NPS improvement, which is what your link is about; > >You can measure time-to-solution speedup. Which is what the 3.1 is about that >we have discussed here... That makes sense. I guess it isn't hard to get a linear speedup if you only measure nodes per second (mostly uneccessary, duplicated work?). Is time-to-solution the same thing as time-to-depth?
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.