Newsgroups: comp.misc
Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!think.com!snorkelwacker.mit.edu!bloom-picayune.mit.edu!athena.mit.edu!abennett
From: abennett@athena.mit.edu (Andrew Bennett)
Subject: Re: 5 billion computations per second!!!
Message-ID: <1991Mar26.160731.7527@athena.mit.edu>
Sender: news@athena.mit.edu (News system)
Organization: Massachusetts Institute of Technology
References: <2480@umriscc.isc.umr.edu> <1991Mar26.015138.28898@csrd.uiuc.edu>
Date: Tue, 26 Mar 91 16:07:31 GMT
Lines: 44

In article <1991Mar26.015138.28898@csrd.uiuc.edu>, eijkhout@s41.csrd.uiuc.edu (Victor Eijkhout) writes:
|> 
|> Thinking Machines in Boston make the Connection Machine,
|> which is (they claim) a 64K processor machine.
|> However, every processor is a bit-processor, and there are
|> 16 of those on a single custom chip, so it is closer to
|> the truth to say it's a 4K processor machine, with a 12-dimensional
|> hypercube architecture. And their latest model has
|> a floating point accelerator per node so you have 4K
|> fp chips. Now, if you take an 'algorithm' that has hardly
|> any communication (vector plus vector, or so)
|> 5Gflop is easy to believe. I seem to remember from 
|> a few years back that they can do inner products
|> at 20Gflop, and slightly less trivial algorithms such
|> as a Conjugate Gradient method at 5Gflop. (flop is
|> floating point op per sec.).
|> 
|> Victor.

Also, you can only use a Connection Machine (CM) on a limited (but important)
class of problems; those that can use fine-grain parallelism.  i.e. if you can
break your simulation into lots & lots of little routines (fluid flow and finite
element analysis are the classic examples), then the CM is perfect for you.

If, on the other hand, your job is a complex series of calculations on a limited
number of loops (physics has lots of these kind of problems), then the CM won't
be very useful to you.

An interesting fact about the Los Alamos tests:  There were three tests run
on the supercomputers, but thinking Machines only competed in one of them.  It's
kind of like Edwin Moses (the hurdler), entering the Decathalon and then only
competing in the hurdles part.  Hardly indicative of his overall abilities...

Hedge your bets, buy a CM _and_ a Cray!

-Drew

--
Andrew Bennett             MIT Network Services                 abennett@mit.edu
MIT Room 11-124H                                   abennett%athena@mitvma.bitnet
77 Massachusetts Ave.                                      Phone: (617) 253-7174
Cambridge, MA  02139                                <Standard disclaimers apply>
"In the Morning, Laughing Happy Fish Heads,
                                          In the Evening, Floating in the Soup!"
