Thanks to the hard work of some very busy electrical engineers at
the University of Kentucky, supercomputing can now be considered a
The machine they made, dubbed KASY0 -- Kentucky Asymmetric Zero -- not only outworks its predecessor, it's cheaper too.
In general, the cost of a supercomputer depends on its size and power.
For instance, a decade ago, it would have cost about $1 million per unit of performance.
"Then, a supercomputer was very fast and very expensive," said Hank
Dietz, 43, professor of electrical and computer engineering and James
F. Hardymon chair in networking at UK.
So how fast is this new machine?
"It's damn fast," Dietz said.
And how expensive is it?
It's damn cheap.
The machine Dietz and other university researchers built in August
cost about $39,500 for more than 471.5 billion floating point
operations per second, or less than $84 per GFLOPS (calculated in 32
bit floating point). Don't be scared, GFLOPS are just the computer's
way of representing the number of calculations per second it can do.
The team made history with KASY0's predecessor in May 2000 when they
achieved $640 per GFLOPS, smashing the $1000 per GFLOPS barrier.
Remember, by comparison a decade ago the price was about $1 million per
Helping drive down the cost of KASY0 is the use of standard personal computer parts and hardware.
From the outside, KASY0 looks like 128 of your typical PCs sitting
next to one another on a series of shelves in a room. But, on the
inside of each, you have a very streamlined piece of equipment that
looks as though something's missing.
"You don't need the whole PC there. You won't find a video card, a
mouse or hard drives," Dietz said. "Everyone knows that PCs are
relatively cheap and fast. The trick is to use that."
While supercomputer components have become faster and cheaper by
using standard personal computer parts, an efficient network needed to
accommodate them remains expensive. The accomplishment of the UK team
was to bring down the networking cost without sacrificing speed.
"What makes what they are doing interesting is that lots of people
are putting together PC-based supercomputers, but not a lot of people
are handling the latency numbers," said Steve Conway, spokesman for the
supercomputer manufacturer Cray Inc.
Latency is important to supercomputing because it refers to the amount of time it takes the computer to figure out a problem.
"It's great to see folks there making strides with latency and
that's what's going to be more important to supercomputer progress than
raw processing power," Conway said.
What exactly does a supercomputer compute with all this speed and power?
"The weather forecast you see at night is done by a supercomputer.
Every car is now designed by supercomputers. New drugs," Conway said.
"Where they really get to be important is for the problems people want
to do and can't."
Cray builds systems that range from $1 million to $10 million for the most complex problems.
"Computer simulations are cheaper than physical experiments," said
Thomas Hauser, assistant professor of mechanical engineering at Utah
State University. Hauser was on the research team during his time at UK
that developed an earlier supercomputer.
"It makes it possible to simulate interesting physics that we don't
know already or understand completely. If you would do simulations on
your personal computer, it would take years to complete, versus weeks,"
"We're making the cost of supercomputers that much less so that more
engineers are able to have it," said Tim Mattox, 31, research assistant
and electrical and computer engineering doctoral student. Mattox is the
primary visionary of the KASY0's network.
The team's unofficial policy is to share the technology, Mattox
said. The public and engineers alike can visit their Web site to
compute problems or configure a supercomputer.
"It's nice to have a theoretical research breakthrough," Mattox
said. "But it's more wonderful that someone can immediately use it."