IBM ASCI White
Eugene Leitl
eugene.leitl at lrz.uni-muenchen.de
Wed Jul 5 17:06:28 PDT 2000
http://www.wired.com/news/technology/0%2C1282%2C37285%2C00.html
BM's Got A Big, Bad Computer
by Leander Kahney
3:00 a.m. Jun. 29, 2000 PDT
Last Friday a large convoy of trucks left an IBM manufacturing plant
in upstate New York and headed across the country to a federal weapons
lab in Northern California.
The 28 semi-trucks were loaded with the first batch of components for
the world's largest supercomputer, a monster machine the size of two
basketball courts that draws enough electricity to power a small town.
Over the next two months IBM's ASCI White will be assembled at the
Lawrence Livermore National Laboratory, one of the U.S.'s leading
nuclear research labs.
The first convoy delivered only a quarter of the machine. Several more
will be needed to deliver the rest.
When it is up and running, ASCI White will be the most powerful
computer on the planet. It will be used to simulate nuclear test
blasts at an unprecedented level of detail and speed. One of the
simulations will run for 30 days. A Cray supercomputer built in 1995
would take 60,000 years to perform the same calculations.
ASCI White can perform a mind-boggling 12.3 trillion operations a
second, or 12.3 teraflops. It is three times faster than the previous
fastest machine, another IBM giant known as ASCI Blue, which runs at
3.8 teraflops.
"The numbers we're seeing make it by far the world's largest
supercomputer," said Jim Jardine, the ASCI White program manager at
IBM, who benchmarked the machine before it was shipped out. "It's a
fast machine."
ASCI White is so powerful, it makes Deep Blue, its famous
chess-playing cousin, look like a cheap pocket calculator. ASCI White
is 1,000 times more powerful than Deep Blue, which generated 200
million chess moves every second to famously defeat World Chess
Champion Gary Kasparov in May 1997.
ASCI White is not one computer, but a massively parallel machine made
from 512 of IBM's RS 6000 servers. Each server has 16 processors --
supercharged versions of the PowerPC chips used in Apple's Macs --
which also operate in parallel.
Total processors: 8,192.
Each server is about the size of an air conditioner. They are stacked
on top of each other in refrigerator-sized racks, which, arranged row
after row, fill a giant hall the size of two basketball courts.
"The size is unbelievable," said Debra Goldfarb, an analyst with
IDC. "Just the wiring will blow you away. It's so complex, it's
unbelievable."
Running a parallel version of IBM's AIX, Big Blue's flavor of UNIX,
the ASCI White supercomputer takes two hours to boot up.
It requires the constant attention of a small army of systems
administrators. Because of the sheer number of parts, it tends to
break down, although IBM said the machine will run 100 hours without
crashing.
"It's designed for ultra-high reliability," Jardine said. "But it's
not perfect."
ASCI White cost Lawrence Livermore $110 million but would have cost
even more if IBM hadn't used off-the-shelf parts, according to a lab
spokesman.
The machine joins a handful of giant supercomputers in the
U.S. Department of Energy's Accelerated Strategic Computing
Initiative, a multi-million-dollar effort to build supercomputers
capable of simulating nuclear test blasts.
Under terms of the 1996 Comprehensive Test Ban Treaty, the U.S. is
prohibited from test exploding weapons from its aging nuclear
stockpile. Supercomputers allow scientists to predict how volatile
materials in the warheads behave as they age and change.
ASCI White is hooked to a bank of imaging workstations that convert
data from a simulation into a visualization of the detonation, which
is projected onto a giant screen.
Weapons scientists will be able to see the actual atoms, said IBM's
Jardine.
"I've been told the physicists come away with a feeling of awe,"
Jardine said. "They say: 'That's what it looks like.' They finally get
to see the physics."
Dr. David Nowak, the ASCI program leader at Lawrence Livermore, said
ASCI White will allow scientists to design extremely complex
simulations and see the results in less time than ever before. The
length of the tests vary, running from minutes to months, but the most
complex involve thousands of variables and generate terabytes of data,
he said.
The supercomputer will also be used for non-weapons research. Twice a
year, it will be turned over to university researchers for three or
four days for work on astrophysics, materials science and drug
discovery.
Dick Sherman, president of RCI, a consortium of supercomputer vendors
and users, said it is likely ASCI White will lead to scientific
breakthroughs that weren't possible with the previous generation of
machines.
"It may help tackle problems that haven't been tackled before," he
said. "Universities are hungry for high computing access and this
would help them."
Despite its astounding speed and complexity, the current ASCI White
can't really deliver the most minute details of a nuclear blast that
scientists want. For that, the ASCI program estimates it needs a 100
teraflop machine.
AIBM said it plans to deliver an even bigger, faster version of ASCI
White by 2004.
If IBM manages it, massively parallel supercomputers will be evolving
at a rate that far outstrips Moore's Law, which dictates computing
power doubles every 18 months.
Both IBM and Lawrence Livermore say this rate of change will mean
today's hundred-million-dollar supercomputers are the widely
affordable workstation of choice within a decade.
"We can solve the problems today that a university researcher will be
able to solve in 10, 12 years from now," Nowak said. "It's a window
into the future."
More information about the Beowulf
mailing list