Author: Robert Hyatt
Date: 14:23:13 05/14/04
Go up one level in this thread
On May 14, 2004 at 15:51:20, Frank E. Oldham wrote: >On May 14, 2004 at 08:23:41, Robert Hyatt wrote: > >>On May 13, 2004 at 23:58:43, Vincent Diepeveen wrote: >> >>>On May 13, 2004 at 12:46:29, Robert Hyatt wrote: >>> >>>>On May 13, 2004 at 07:47:32, Vincent Diepeveen wrote: >>>> >>>>Also you do know this is a cluster? Not a NUMA box? >>>> >>>>Of course you did. >>>> >>>>I'm sure your program does _great_ on a message-passing cluster... >>> >>>May i remind you your message passing cluster you get there has a faster one way >>>pingpong latency than origin3800 at 512 processors.... >> >>May I remind you that NUMA and clusters are _two_ different things??? >> >>Or is it pointless??? >> >>Clusters have _zero_ shared memory. >> >> > >Apparently this machine will be based on the Cray X1 architecture -- this is >NUMA and supports a single-system-image (as well as various MPI types). It also >uses traditional vector processors. > >Frank It doesn't look that way to me. This appears to be a "son of red storm" which is opteron-based and not related to the X1 whatsoever. IE they have already built a 10K node opteron box to deliver something around 50 tflops. This was old news earlier this year. In this case it is a "NUMA-extreme" system because it is connected in a 3d mesh. It'll work well for problems that have 3d data sets that fit. It'll die for others like chess without a _lot_ of very specific tuning. Way beyond normal NUMA-type tuning I'll add. I can't see the X1 extending that far. The biggest configuration I have heard of has 4096 cray vector processors. Vector processing is really memory intensive and that's the "meat and potatoes" of the X1 architecture.. some cool stuff with cache and address translation...
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.