Re: State of the Art - Gflops/m^3

From: Brian Atkins (
Date: Mon Jul 02 2001 - 16:04:59 MDT

Dani Eder wrote:
> --- Brian Atkins <> wrote:
> > P.S. there is an article on Blue Gene in the July
> > Wired.. they talk about
> > the spinoffs which will likely be multi-teraop
> > supercomputers that will
> > take up only perhaps a single rack of floorspace,
> > require less than 1/20th
> > the electrical power of existing supercomputers, and
> > cost less than one
> > million bucks... perhaps closer to $100k. Oh yeah
> > this is all scheduled
> > to be around 2004.
> This year's state of the art would be a rack full of
> Dual Athlon systems. By this fall they will be
> available in a 1U form factor at 1.4Ghz. Each
> CPU produces 2 Gflops. A dual-CPU system is then
> 4 Gflops, and you can pack about 32 1U systems in
> a rack, for 128 Gflops in a 1 rack space. Initially
> these systems will go for around $3K each, so you
> can figure about $100K for a rack full.
> If you are going for maximum processing capacity
> per dollar rather than per cubic meter, you can
> already build a cluster out of desktop PC parts
> for around $500-1000 per node, giving you $250-500/
> Gflop, or $250K-500K/Teraflop. The major variable
> is how much RAM and hard disk per node you buy.

So if we guess that we get two doublings in performance in these
commodity chips, then we get into the range of 4tflops for $250k-
500k by 2004 at the earliest. Not taking into account that a beowulf
may not be able to achieve the same real world performance of a
blue gene design due to slower inter-node links. What's interesting
to me is that this means the beowulf price advantage may disappear
after 2004... it looks like by then you can buy a blue gene spinoff
for the same price, which may have better performance, and may be
more easily supportable due to its self-healing nature, and also will
likely take up less space and power. Will supercomputers continue to
outstrip the advances in "consumer level" systems?

Brian Atkins
Director, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT