Author: Robert Hyatt
Date: 08:08:54 07/21/00
Go up one level in this thread
On July 21, 2000 at 07:47:23, Chris Carson wrote: >http://www.research.ibm.com/deepblue/meet/html/d.3.3a.html#same > >IBM never said 1B max. Show me the link or are you quoting the >1B NPS averge that was the goal in 1990 (see Scientific America) >and never reached in 1996 or 1997 or do you have your facts wrong again? > >Best Regards, >Chris Carson Read carefully: Hsu has stated, in public forums, that he had two versions of the chip they used in the 1997 match. Early batches had a cross-talk problem and could only run at 20mhz, which at 10 clocks per node turns into 2M nodes per second. Later versions of the chip ran at 24mhz, or 2.4M nodes per second. This is in his IEEE Micro article from a couple of years back. He said they had a 50-50 mix of these. So we go with an average of 2.2M nodes per chip. That math bother you? N*2.0 + N*2.4 / 2*N = 2.2 Next, from the IBM web site, and articles and talks by Hsu and Campbell, we know they used 480 chess processors distributed over 32 nodes on the IBM SP machine they used for a host. More math: 480 * 2.2M > 1000M Do you want to dispute that? Hsu said that in the typical case, he could drive the chips at about 70% duty cycle. .7 * 1000M = 700M. Still there? In his thesis, and in things he has written and also presented publicly since the match, he said his parallel search was about 30% efficient. That is if he uses 100 processors, he runs about 30 times faster. .30 * 700M is about 200M. That is a _very_ conservative number when compared to how others report their raw NPS. If you do your homework, you can arrive at those same numbers on your own. Or you can say "wrong again" and stop there. The numbers are _real_. Like 'em or not...
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.