/-------------------- advertisement -----------------------\
Explore more of Starbucks at Starbucks.com.
http://www.starbucks.com/default.asp?ci=1015
\----------------------------------------------------------/
Supercomputing's New Idea Is Old One
August 4, 2003
By JOHN MARKOFF
After a period of neglect, the intellectual legacy of
Seymour Cray, the father of the modern supercomputer, is
being revived.
The scientists in government, industry and academia who are
engaged in the race to build the world's fastest computing
machines are now turning their attention once again to Mr.
Cray's elegant approach to building ultra-fast computers.
When Mr. Cray died after a car accident in 1996, the
one-of-a-kind machines that embodied his computing
philosophy had gone out of fashion, largely replaced by
designs based on thousands of connected microprocessors
that are inexpensive and mass produced.
Mr. Cray's custom machines are known as vector
supercomputers and have special hardware that is intended
to handle the long strings of numbers in complex scientific
computing problems. The machines are highly regarded for a
design that balanced computing speed and the ability to
transfer data extremely rapidly within the computer while
the calculation is taking place.
This design philosophy is being revitalized by Burton J.
Smith, a founder and the chief scientist of the
Seattle-based Tera, which bought the original Cray Research
in 2000. In the three years since the acquisition, Mr.
Smith has been seen in the industry as the most prominent
champion of Mr. Cray's approach.
"The pendulum swung way too far," Mr. Smith said of the
research and attention given to the microprocessor approach
to supercomputing.
For the first 25 years of supercomputing, Mr. Cray's
machines held the title of "faster computer in the world."
But beginning in the late 1980's, the entire computing
world became subject to what became known as the "attack of
the killer micro."
Instead of specialized hardware, along the lines of Mr.
Cray's designs, these rival computers rely on
microprocessors - with hundreds, and then thousands, and
then tens of thousands of microprocessors lashed together
to make ever-faster supercomputers.
Known as "massively parallel processors," or M.P.P.
machines, these computers rapidly took over the
supercomputer market, gaining popularity in research and
weapons laboratories.
Looking back, the arrival of M.P.P. machines was a product
of government policy.
Up until the 1980's, the United States government viewed
supercomputing as part of its technological competition
with the Soviet Union and Japan, heavily subsidizing
research and development. But with the end of the cold war,
federal support evaporated because of a lack of strategic
urgency and a belief that M.P.P. machines proved that
supercomputing could prosper with components derived from
the consumer computing industry.
Today, the M.P.P. designs are proving to have limits, a
fact made obvious last year when a Japanese supercomputer
built in the Cray tradition by NEC Electronics Corporation
won the computing speed crown.
Now there is a emerging consensus among American computer
designers that the Lego-block approach of chaining together
microprocessors fails when it comes to certain classes of
computing problems.
"There's a buzz in the air," said Jack Dongarra, a
University of Tennessee computer scientist who tracks the
world's fastest 500 computers. Scientists and computer
designers have been trying to convince the government
agencies of the need to invest in designs in the tradition
of Mr. Cray, he said. "We're crossing our fingers."
According to a number of scientists, the real issue,
outlined in a National Science Foundation committee report
this year, is too little investment in advanced computing
technologies. A similar report from the Defense Department
and a report soon to be released by the National Academy of
Science echo that theme.
In hearings before the House Science Committee last month,
supercomputing experts put forward a case for a significant
increase in federal spending on supercomputing, citing
their relevance to counterterrorism projects and scientific
research.
This refocus is underscored by supercomputer experts who
said that JASON, an elite group of scientists who perform
studies for the Pentagon, is investigating shortcomings in
the military-financed Advanced Strategic Computing
Initiative, which relies heavily on massively parallel
computers.
There is a growing concern that the billion-dollar program,
which is intended to monitor the nation's nuclear stockpile
and simulate weapons explosions, has been a disappointment
because the M.P.P. computers used by the project have
proven inefficient and difficult to program.
The Cray approach is now seen as a natural alternative.
Cray Inc., the name Tera took after purchasing Cray
Research, last month reported strong revenue and profits
for its second quarter of 2003. Revenue for the quarter was
$61.8 million compared with $38.6 million for the same
quarter last year. Net income was $7.9 million, or 10 cents
a share, up from $1.2 million, or 2 cents a share, reported
in the second quarter of 2002.
The company's stock, which had been as low as $3.05 in the
last 12 months, ended last week at $11.21.
Cray's revival was helped last month when the company
became one of three computer makers, along with I.B.M. and
Sun Microsystems, chosen by the Pentagon's Defense Advanced
Research Projects Agency to develop prototypes of next
generation of supercomputers that can reach peak speeds of
a petaflop - quadrillion mathematical operations per second
- by the end of this decade.
Cray had been in decline since the late 1980's. The company
was purchased by SGI in 1996 for $740 million, a move that
some Cray designers have referred to as "the occupation."
SGI put an end to the Cray vector supercomputer line, but
was forced to resurrect it several years later when the
National Security Agency indicated a need for the
traditional machines. The agency financed the development
of a new class of supercomputer called the X1, which the
company introduced last year.
In 2000, the company was sold again, this time to Tera, a
company founded by Mr. Smith and his partner James
Rottsolk. While Mr. Cray was known as a hardware-packaging
genius, Mr. Smith, a 63-year-old math expert, has a
reputation as a remarkably inventive computer designer.
He was a pioneer of the concept of multithreading, a
technique that is widely used today in modern
microprocessors to add parallelism and to make more
efficient use of hardware resources.
At the company, Mr. Smith's vision is matched with the
pragmatism of the chief architect of the Cray X1, Steve
Scott, say scientists who have worked with both men.
"At Cray, Burton serves the role of the idea person," said
Bill Dally, a Stanford computer scientist who has been a
consultant to Cray over many years. "He's always coming up
with very creative, very bold ideas."
Mr. Dally said that the two men work well together, with
Mr. Scott being more practical and down to earth. "It's a
good match," he said.
A crucial part of Mr. Smith's success has been his ability
to create a company that can remain viable in a business
where product design and development can take decades. "I
have this idea of the manifest destiny of computers," he
said in explaining why he is motivated to persist in the
business. They are the only tools, he said, that "truly
leverage the power of the human mind."
When Mr. Smith and Mr. Rottsolk founded their company in
1987, they chose Seattle because it was a place that did
not have the intense hothouse culture of Silicon Valley.
The company's headquarters, where 120 of Cray's 900
employees work, is in Pioneer Square, in the heart of the
city's historic district.
One indication of a shift away from M.P.P. design is that
Thomas Sterling, a computer scientist at the California
Institute of Technology, and Mr. Smith's consultant in the
design of the new supercomputer, was the creator of a
particular class of inexpensive massively parallel
computers known as Beowulf clusters that can be assembled
from off-the-shelf personal computers.
Mr. Sterling acknowledged that while those cheap machines
solve many problems there are computing challenges that
will require fundamental new designs like the new Cray
machine, called Cascade, that may reach a petaflop sometime
near the end of this decade.
It will take that kind of computing speed, for example, to
simulate all the effects on an airplane wing as it moves
through the atmosphere: the flapping of the wing,
turbulence and the changing temperature of the wing. Mr.
Sterling said it was not yet proved whether Mr. Smith's new
machine would be able to meet this kind of challenge.
"He is simply one of the brightest guys about all aspects
of computing that I've met," Mr. Sterling said. "The
problem here is that Burton's legacy will be determined by
his impact."
http://www.nytimes.com/2003/08/04/technology/04SUPE.html?ex=1061027832&ei=1&en=5967fa5e276b2bd6
---------------------------------
Get Home Delivery of The New York Times Newspaper. Imagine
reading The New York Times any time & anywhere you like!
Leisurely catch up on events & expand your horizons. Enjoy
now for 50% off Home Delivery! Click here:
http://www.nytimes.com/ads/nytcirc/index.html
HOW TO ADVERTISE
---------------------------------
For information on advertising in e-mail newsletters
or other creative advertising opportunities with The
New York Times on the Web, please contact
[EMAIL PROTECTED] or visit our online media
kit at http://www.nytimes.com/adinfo
For general information about NYTimes.com, write to
[EMAIL PROTECTED]
Copyright 2003 The New York Times Company
Tired of spam? Get advanced junk mail protection with MSN 8. -------------------------------------------- This service is hosted on the Infocom network http://www.infocom.co.ug
