Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: OT Supercomputer Red Storm

Author: Robert Hyatt

Date: 06:54:47 11/10/03

Go up one level in this thread


On November 09, 2003 at 07:52:54, Vincent Lejeune wrote:

>
>A paper about the new supercomputer (peak at 40 Tflops) "Red Storm", end
>building in August 2004
>
>http://www.lanl.gov/orgs/ccn/salishan2003/pdf/camp.pdf
>
>Bob could give a critical view of this new supercomputer, please ? :)
>Are all the informations in this doc really accurate ?
>
>All this paper seems to be about too good :)

This is just another cluster.  A very big one, with a better
interconnection scheme (3d mesh rather than the usual ethernet
or hyper-cube approach to things.  But it is based on traditional
computers with local memory, connected in a very fast way.  No
shared memory, everything is local to each processor.

It's interesting.  But not for chess, particularly.  It is a
pure message-passing architecture based on the classic MPI
solution.  Using that many processors effectively has never been
done by a chess program...

It is also really tailored to particular types of matrix
math because of the 3-d mesh interconnect.  IE interior nodes are
connected to 6 nodes, up, down, left, right, front, back.  You can
get data to/from those six nodes quickly.  If you have to go farther,
you have to go through those as the first hop.  For a 1000 node machine,
you have a worst-case transit time of 18 hops for one corner to talk
to the opposite corner (in 3d space).  This makes mapping the algorithm
to the architecture critical, to minimize these "hops"...



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.