From: Robert J. Bradbury (bradbury@aeiveos.com)
Date: Wed Aug 29 2001 - 18:44:47 MDT
I believe the situation went something like this.
Cray found their market drying up and SGI bought them out
thinking they would be a good fit (turn old Cray customers
into new SGI (Origin) customers). Then I think SGI figured
out the future really wasn't in supercomputers but in bulk
Linux clusters (at least for now) and wanted to unload Cray.
TERA had been working on an advanced architecture and had
burned up a lot of cash building GaAs chips that required
very complex boards. It worked but by the time they proved
it, CMOS had caught up with GaAs speeds. So they had to
redo the architecture in CMOS to lower the production costs.
In the meantime I think they needed a revenue stream to
keep the stockholders from giving up entirely, so they
bought Cray from SGI. SGI unloads a lemon from its
perspective and TERA gets a possible future customer base.
(Replace your aging Cray's with new TERAs).
Browse on over to www.tera.com which now seems to be www.cray.com
and take a look.
I agree Seymour Cray effectively was CDC as well as Cray.
His biography was a good story from my perspective and shows
what happens when you try to skate a little to close to the
edge (pushing the technology for the Cray-3).
Now the really *interesting* thing in the Blue Gene article
is that IBM is designing the chip to be multi-threaded.
This allows the CPUs to service whatever thread has got
its data loaded from memory. It sounds precisely like the
TERA architecture. So either IBM has licensed TERA's
technology or they have reinvented something similar.
But its going to be interesting -- Blue Gene is going to
push us *2* orders of magnitude in capabilities (at least
for the applications for which it is appropriate). Thats
quite a jump from Moore's Law of 2x every 18 months or
1 order of magnitude every 5-6 years.
Signs of the Singularity? You be the judge.
The article pointed out is the degree to which they are constrained
by the limit of how much memory they can get on the chips with the CPUs.
There is a match made in heaven if you can get molecular electronics memory
denisties (much greater than current capacitors) combined with CMOS logic.
I'm a lot more optimistic for the development of molecular electronic
memories than the development of molecular electronic logic.
We shall wait and see.
Other interesting "factoids":
-- The cluster will be 40 x 40 ft.
-- Consuming a couple of megawatts of power
-- They are only running the chips @ 500 MHz
(presumably to keep the heat production down)
-- If you can get a modern-day supercomputer to stay up for 100 hours
without a hardware failure you are doing well.
-- They are going to have some way of dealing with chips
that have defective CPUs (lowering the production costs).
-- It takes 10^23 instructions to simulate 100 microseonds of
protein folding. (Yike's you would even have to "wait" for a
rod-logic nanocomputer to do the calc.)
Robert
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:10:14 MST