Restarting the supercomputer race
Silicon Graphics unveils the Chippewa Falls-built Columbia, the world's fastest supercomputer, at NASA's Ames Research Center
The super-fast world of supercomputing got much faster Tuesday.
High-performance computer maker Silicon Graphics announced it has built
the world's most powerful supercomputer at its Chippewa Falls, Wis.,
plant — a machine that will out-process one built by IBM in Rochester,
Minn., that claimed the world's-fastest title just last month.
That could be good news for businesses that continually need faster
supercomputers to design everything from a safer passenger jet to a
more aerodynamic potato chip.
Mountain View, Calif.-based Silicon Graphics officially unveiled its
latest supercomputer Tuesday at NASA's Ames Research Center, also in
Mountain View. The 10,240-processor behemoth occupies a space the size
of four basketball courts. Its 20 systems were built at Silicon
Graphics' 410-worker manufacturing facility in Chippewa Falls, a town
100 miles east of the Twin Cities known to most Minnesotans more for
its beer than its high technology.
Much of the Linux software that runs the supercomputer was developed
at Silicon Graphics' 180-person office in Eagan, spokeswoman Lisa
The new machine is named Columbia, after the ill-fated space shuttle
that crashed last year, killing all its crew. It is supposed to help
NASA's grounded space shuttle program return to flight, NASA hopes, by
next April or May. The machine's first system was dubbed Kalpana, after
Columbia crewmember Kalpana Chawla, who worked at NASA Ames before she
became an astronaut.
NASA threw its entire supercomputing budget into the project and
wanted it assembled and tested at break-neck speed this summer and fall
to make the spring flight window. Unlike traditional systems that take
years to install, Columbia was fully available to NASA scientists as it
was installed piece by piece.
"It was like building a plane while you're flying it," said Walt Brooks, chief of NASA's Advanced Supercomputing Division.
But supercomputer Columbia's selling point is its own speed, with
the ability to perform 61 teraflops — or trillion calculations per
second — at its peak. More importantly, it recorded a sustained running
speed of 42.7 teraflops.
That crushes the claims of IBM, which last month boasted it had
built the world's fastest supercomputer, named Blue Gene, at IBM's
facilities in Rochester, 75 miles south of the Twin Cities. Blue Gene
recorded a sustained performance speed of 36.01 teraflops, IBM said.
How fast is all this?
An average supercomputer can do in a day what a desktop computer
would need months to complete. But this new generation of turbo-charged
supercomputers can solve problems in that same day that ordinary
supercomputers would need months to finish, if they can solve them at
all, Silicon Graphics and NASA say.
"This is the world's most powerful supercomputer," Brooks told
cheering Silicon Graphics employees, who celebrated at a picnic outside
their Chippewa Falls plant last Thursday, washing down sloppy joes with
locally brewed Leinenkugel beer and soda pop.
Columbia was completed in 120 days on Oct. 12 by working around the clock, seven days a week.
"This is a great project," said Bill Liedl, a Chippewa Falls
engineer who estimated he contributed several hundred hours to the
project. "I mean, the fastest supercomputer in the world — who wouldn't
want to be here?"
To demonstrate its power, the Columbia supercomputer needed only two
days in an exercise to solve the mystery of the debris that broke off
the Columbia shuttle's rockets and struck its shuttle wing's edge,
leading to its crash. It recreated the accident in a three-dimensional
simulation, Brooks said.
NASA's old supercomputers took three months processing that problem, he said.
The new machine opens up the possibility of diagnosing and fixing
shuttle problems in flight, giving more accurate hurricane forecasts
days earlier or even jumpstarting manned flights to Mars, some experts
There is a little oneupmanship going on among computer makers. Both
Silicon Graphics and IBM were gunning for Japan's NEC Corp., whose
Yokohama-built Earth Simulator has held the title of world's fastest
supercomputer for nearly three years, with a sustained speed of 35.86
Last week, NEC announced it would start shipping in December a new machine called SX-8, with a peak speed of 65 teraflops.
IBM also has said that when it finishes installing a Minnesota-built
Blue Gene system at the nuclear weapons-oriented Lawrence Livermore
National Laboratory in Livermore, Calif., next year, its machine will
have 360 teraflops of power.
Silicon Graphics officials dismissed the claims of IBM and NEC,
saying it alone has a machine up and running that can claim the title
of world's fastest supercomputer today.
"We believe supercomputing is not about delivering a stunt or a
piece of media fodder," said David Parry, Silicon Graphics senior vice
president and general manager for servers and platforms. "It's about
building new tools to do substantive science."
The jockeying by Silicon Graphics, IBM and other supercomputer
makers like Seattle-based Cray Research, which was once owned by
Silicon Graphics, means the American supercomputing industry is
reviving after playing dead for the past few years, said Addison Snell,
research director of high-performance computing at IDC, a Massachusetts
firm that measures technology markets.
The $6 billion supercomputer market is small compared to the $46.1
billion business server market, but demand for supercomputers is
growing steadily — nudged in part by a White House task force that saw
Earth Simulator as a computing "Sputnik," challenging U.S.
supercomputing supremacy, Snell said.
The growth is not in high-end research with outfits like NASA but in
slightly less powerful machines used by businesses to design cars and
airplanes, find and drill for oil and gas or create movie special
effects, Snell said. Silicon Graphics touts its new Altix systems in
Columbia as easily convertible to such uses.
At Cincinnati-based Procter & Gamble, for instance, director of
modeling and simulations Tom Lange used Silicon Graphics machines a few
years ago to build a better Pringles potato chip.
The chips are made at high speed, and he needed to run complex
aerodynamics studies to figure out how to keep them from flying off the
conveyor belts, Lange said.
This year, supercomputer simulations knocked months off designing
prototypes for the new all-plastic "AromaSeal" canisters for Folgers
coffee, he said.
"We never could have done it without a supercomputer," Lange said.
"It does not eliminate all prototyping. It just means you eliminate all
the dumb ones that wouldn't work."
The demand for innovation means he needs new supercomputers every
three to five years when the old ones become obsolete. "We have an
insatiable need for speed," Lange said.