Body of Secrets: Anatomy of the Ultra-Secret National Security Agency (97 page)

Not to be
outdone, Los Alamos National Laboratories, by stringing together an array of
supercomputers and associated networks, was able to perform more computing work
in a twenty-four-hour period than had been done by all of humanity before the
year 1962. And that estimate was considered conservative by other researchers,
who suggested that a date in the late 1970s might be more accurate.

The speed
of electrons, however, was not NSA's most immediate problem; the agency was
also worried about the speed of the Japanese. Japan was the only other nation
aggressively pursuing supercomputer development. In the summer of 1988, a
gathering of leading computer science experts, among them NSA's director of
supercomputer research, met to assess Japan's progress in supercomputers. If
they felt confident when they walked into the meeting, they were more than a
little nervous when they left. Starting only six years earlier, Japan had
already matched or surpassed the United States in a field the United States
invented and had been advancing for twenty years.

The main
problem for the American supercomputer industry was dependence on Japanese
computer companies—their arch-competitors in a cutthroat business—for critical
parts, such as computer chips, for their machines. This was a result of the
gradual abandonment of semiconductor manufacturing in the United States during
the mid to late 1980s. In 1986, for example, NSA was virtually dependent on a
Japanese company, Kyocera, for critical components that went into 171 of its
196 different computer chips, according to the minutes of a Department of
Defense study group. When, without warning, Kyocera stopped making a component
known as a ceramic package, used in a key chip, NSA began to shudder.

In a worst-case
scenario, Japanese computer manufacturers could slow down or cut off the supply
of essential components to their American supercomputer competitors—and NSA.
This fear led the panel to conclude that within a few years, "U.S. firms
would be most fortunate if they found themselves only a generation or so
behind."

As a
result of such worries, NSA, with the help of National Semiconductor, built its
own $85 million microelectronics production and laboratory plant, known as the
Special Processing Laboratory. Located in Crypto City, the ultra-modern,
windowless, 60,000-square-foot building first began producing chips in 1991.
Today it employs several hundred people. The building contains 20,000 square
feet of "class 10" clean rooms—rooms whose air is 10,000 times
cleaner than normal air. The water must also be ultra-pure because the
particles in the water can destroy a transistor.

Building
its own plant also solved another problem for NSA: the need for supersecrecy in
producing highly customized parts for use in the agency's unique codebreaking
machines. These components, "applications specific integrated
circuits" (ASICs), are often the "brain" of a codebreaking
system, thus making outside procurement "a nightmare," said one
NSAer. With the ability to squeeze 1 million or more transistors on a single
piece of silicon, designers can now build entire algorithms on a chip—a
complete crypto system on a piece of material many times smaller than a dime.
For such a chip to fall into the wrong hands would be disastrous.

So NSA
added another new feature: a secret self-destruct mechanism. Developed by
Lawrence Livermore and Sandia National Laboratories, NSA's chips are shielded
by special self-destructing coatings. "If a hostile agent tries to take
off the lid," said one knowledgeable source, "the coating literally
rips out the top [circuit] layer."

Six months
after the 1988 computer science panel meeting, fear over Japan's rapid push
into the supercomputer industry once again surfaced. On December 6, 1988,
Japan's Fujitsu—a key supplier of critical chips to Cray—announced a major new
advance: a blisteringly fast computer with a theoretical top speed of 4 billion
operations per second. This equaled and perhaps beat Cray's top-of-the-line
machine, the Y-MP, which had been on the market for less than a year. The
problem for NSA was that the Japanese company could easily sell the superfast
computer to other nations, which might then use it to develop encryption
systems far too fast for NSA's codebreaking computers to conquer.

But while
Japanese companies were catching up and maybe even passing their American
competitors in speed, the U.S. supercomputer industry was far ahead in both
software development and the use of parallel processing. As fast as the Fujitsu
computer was, it had only two processors. Cray and ETA had both developed
machines with eight processors—eight brains, in a sense—which could
simultaneously attack separate parts of a problem.

To Seymour
Cray, sixteen brains were better than eight, and for several years he had been
trying to prove it by building a sixteen-processor CRAY-3. It was an expensive
and time-consuming endeavor— too much so, it turned out, for Cray Research, the
company he had founded but no longer owned. In May 1989, the two split. Seymour
Cray took 200 employees and $100 million and moved to Colorado Springs to found
Cray Computer, Inc., as a wholly owned subsidiary of Cray Research. Eventually,
it was planned, Cray Computer would become independent.

Like a
race-car driver with his foot stuck to the accelerator, Cray continued to push
for more and more speed; he hoped to break sixteen billion operations a second.
The secret would be to make the hundreds of thousands of chips that would
constitute the soul of the new computer not out of conventional silicon but out
of a radical new material: gallium arsenide. Although it was more difficult and
costly to work with, electrons could travel up to ten times as fast through the
new compound as through silicon.

But as
"the Hermit of Chippewa Falls," as Cray was affectionately known,
quietly pushed ahead in his new laboratory in Colorado Springs, the world
around him began shifting and turning. The Cold War had ended and weapons
designers were no longer shopping for supercomputers. The fat Reagan years of
Star Wars were giving way to the Clinton era of cutbacks and deficit reduction.
And industry was turning away from the diamond-encrusted CRAYs, made of a small
number of superpowerful processors, and toward less pricey massively parallel
computers made up of thousands of inexpensive microprocessors. The enormously
expensive, hand-built Formula One racers were being forced off the track by
cheap stock cars packed with store-bought superchargers and sixteen-barrel carburetors.

At ETA
Systems, which had pushed the supercomputer speed envelope with its ETA 10, 800
employees showed up for work on a spring Monday in 1989 to find the doors
locked shut. The company had developed a super debt of $400 million.

Four years
later, Steve Chen folded up his new company, Supercomputer Systems, when IBM
finally cut off funding for his SS-1. Partly funded by NSA, Chen had spent half
a decade attempting to build a computer a hundred times faster than anything on
earth. But in the end, the innovations were overtaken by excessive costs and
endless missed deadlines. A few months after the company's doors closed, one of
its former engineers driving past a farm spotted a strange but familiar column
of metal. A closer look confirmed his worst fears: it was the outer frame for
the SS-1, and it had been sold for scrap.

In 1991,
Thinking Machines Corporation delivered to NSA its first massively parallel
computer—the Connection Machine CM-5, which the agency named Frostberg. Used
until 1997, the futuristic black cube with long panels of blinking red lights
looked like something left over from a
Star Wars
set. Just two years
after the $25 million machine was installed, NSA doubled its size by adding 256
additional processor units. This allowed Frostberg to take a job and break it
into 512 pieces and work on each piece simultaneously, at 65.5 billion
operations a second. Equally impressive was the Frostberg's memory, capable of
storing up to 500 billion words.

By the
time the CRAY-3 at last made its debut in 1993—clocking in at roughly 4 billion
operations a second—there were no takers. Nearly out of money, the company
spent a year looking for customers and finally landed a deal with its old
partner, NSA. In August of 1994, the agency awarded Cray $4.2 million to build
a highly specialized version of the CRAY-3 for signal processing and pattern
recognition—in other words, eavesdropping and codebreaking. Named the
CRAY-3/Super Scalable System, the machine would become the brain of what has
been dubbed "the world's ultimate spying machine." It would link two
supercomputer processors with a massively parallel array of chips containing
more than half a million inexpensive processors designed by NSA's Supercomputer
Research Center.

But while
hoping for Cray to succeed, NSA scientists were also working in-house on new
ideas. One was a processor called Splash 2, which, when attached to a
general-purpose computing platform, was able to accelerate the machine's
performance to super-Cray levels at only a fraction of the Cray cost.

As Seymour
Cray struggled to complete his CRAY-3, he was also in a race with his old
parent company, Cray Research, which was building a successor to its Y-MP
called the C-90. The company was also near completion on a powerhouse known as
the T-90, which would operate at up to 60 billion operations per second.
Meanwhile, Seymour Cray hoped to leapfrog his competitors once again with his
CRAY-4, due out in 1996.

By the
fall of 1994, work on the CRAY-4 was going surprisingly well. Cray Computer in
Colorado Springs was predicting a completion date in early 1995 with a machine
with twice the power of the CRAY-3 at one-fifth the cost. There was even talk
of a CRAY-5 and CRAY-6 before the planned retirement of Seymour Cray. Which was
why the yellow tape came as such a shock. When employees came to work on the
morning of March 24, 1995, they were first confused to see the yellow police
tape sealing the doors. But when they saw the white flag that had been run up
the flagpole, they did not need a supercomputer to conclude that the end had
finally arrived. The man with the unlimited ideas that reached to the stars had
tumbled to the bottom of his finite bank account.

Ever
optimistic, Seymour Cray pulled together a few of his most loyal followers,
scraped together some money from their own bank accounts, and formed SRC
(Seymour Roger Cray) Computers. Cray felt almost liberated at this chance to
"start from a clean sheet of paper." It was also, he believed, a
chance to finally break the speed barrier by building the first teraflop
supercomputer, capable of a trillion mathematical operations a second—12,000
times more than his CRAY-1.

But the
enemy had landed. In the spring of 1996, even the U.S. government had turned
its back on all the Cray companies and awarded a $35 million contract to the
Japanese computer giant NEC for its 128-processor SX-4 supercomputer. The SX-4
would go to the National Center for Atmospheric Research. The agency was
worried because meteorological centers in Australia, Canada, England, and
elsewhere were installing systems that by January 1998 would be capable of
between 20 and 80 billion operations a second. And Cray Research, the agency
concluded, was just not producing computers fast enough. "Simply
put," said William Buzbee, head of the weather center, "Cray Research
lost this procurement because their offer had unacceptable technical
risk."

Others,
too, knew that despite the never-say-die bravado and the endless promises of
illions of flops, the luster was at last disappearing from Cray's blinding
star. "The rules changed when it became clear that Cray Computer Corp.
wasn't going to make it," said John Mashey, director of systems technology
at Silicon Graphics. "It's like watching your favorite quarterback, who
won the Super Bowl many times. But it's not 1976 anymore—his knees are gone and
those three-hundred-pound defensive tackles are fierce. While he keeps getting
up, it's agonizing to watch and you really wish he could have quit on a
high."

A few
months later, while returning from a brief trip to a software store, Cray was
seriously injured when his black Grand Cherokee was struck by another car and
rolled over three times. Two weeks later, on October 5, 1996, the shy maverick
who hand-built the fastest machines on earth, with the meticulous care and fine
craftsmanship of a Swiss watchmaker, died, never having regained consciousness.
His ashes were scattered among the cragged peaks and somber valleys of the
Colorado mountains. They had served as his inspiration, and as silent
comforters, during his last years. "In the days before PCs brought
megaflops to the masses," said one computer expert, "Cray was the
computer industry's closest equivalent to a rock star."

Sadly,
only months before Cray's death, the daring company he had given birth to in
Chippewa Falls, Wisconsin, decades earlier, also died. Following the worst
financial year of its life, in which it was forced to cut nearly a quarter of
its employees, and facing an uncertain future, Cray Research called it quits.
It was acquired by Silicon Graphics, Incorporated—later known simply as SGI—a
Mountain View, California, manufacturer of high-performance workstations, the
sort of machines that became Cray's greatest competitor. "Cray represents
the last of the 1980s pure plays in the supercomputer market," one market
analyst said wistfully. "There are no other major players left standing
from the supercomputer battles of the 1980s and 1990s."

In fact,
there was one. The shakeout and the death of Seymour Cray left a single
independent to fight the army of "killer micros," the massively
parallel microprocessors that turned the budget-draining, high-performance
supercomputer into an endangered species. The large, rumpled man with the Don
Quixote dream was Burton Smith, whose company, Tera Computer, stunned many in
the field by building a machine that in 1997 set a world speed record for
sorting integers. Burton's idea was to increase speed by decreasing the waiting
time it took for processors to be sent new data on which to work. This, Burton
believed, would overcome the Achilles' heel of powerful computing—the gap
between a computer's short-term theoretical "peak" speed and its
long-term "sustained" speed.