🚀 go-pugleaf

In article <01bc7800$5b1b9a40$651c...@venger-no-fear.intervoice.com>,
"Michael E. Chester II" <mche...@utdallas.edu> wrote:

> Pete Wilson <pwi...@dial.pipex.com> wrote in article
> <33A10989...@dial.pipex.com>...
> > Andy wrote:
> > > 200+ PPros in parallel would not make the fastest computer.  Cray
> > > makes supercomputers using 512 500MHz Alphas in parallel.  To make an
> > > Intel-based machine as powerful as this, you'd need to parallel 2048
> > > PPros (large-scale parallelling is done in powers of 2).  This begs
> > > the question: why?  The Alpha solution would be more efficient and
> > > less expensive to implement.
> > > 
> > > I could see a pretty powerful computer being put together with PPros,
> > > but though I can't vouch for whether or not it has been done on this
> > > scale, I could think of better ways of getting that kind of power.
> 
> The Sandia Labs machine is now online. All 9000 or so PPro's...


Just to clairify:

1)The Sandia machine is the world's first general purpose Teraflops
machine. (There was another machine called GRAPE-4 in Japan that was
application specific). The Intel Red uses 9200 200Mhz Pentium Pros, and it
is currently the worlds fastest machine.

2)Cray does not sell any machines using the 500Mhz Alpha. They do sell the
T3E-900 with 450Mhz 21164A Alphas. Some of these machines reach 1024
processors.

3)Not totally sure if you can say Cray still exists. Cray Research and SGI
merged under the SGI logo i thought.

4)Cray/SGI and IBM are installing machines faster than the Intel Red by the
end of next year. The IBM uses 4000 processors based on the POWER or
PowerPC series of CPUs. I think it will be the 630 which is a PowerPC
follow on to the current P2SC chips from IBM. The SGI/Cray will use 3000
processors, most likely a very high speed version of the MIPS R10000 or
R12000.

5)There is no reason for MPP machines to be power of two, even though it is
by far the most common. There was a recent discussion about this in
comp.parallel or comp.sys.super, where we talked about MPP machines with
prime number of nodes.

6)To answer the question posed "Why not Alphas", that is because Intel had
the best low cost implementation. No one said they had to actually make
money on a highly visable machine like the Intel ASCI Red. I might not
think very highly of Intel or the PPro MPP machine, but i do know that they
won the competition over the other bidders. But choice not to follow up
with larger machines, and so IBM and SGI/Cray moved in with the 3 teraflops
machines (ASCI Blue Mountain and ASCI Blue Pacific).

-r.