IP: Gordon Bell wins Bell vs. Hillis

[Dave Farber <farber () central cis upenn edu>]

Source is from <http://www.upside.com/resource/print/9601/teraflop.html>




Laying A Bet To Rest


By Willie Schatz, Upside


Gordon Bell claims to be the winner of the famous bet on supercomputers that
he made with Danny Hillis five years ago. But the bigger issue in the dying
supercomputer market is not who has the fastest machine, but whether anyone
will still be making Big Iron in the future.


As the bell tolls at the end of 1995, it tolls for Danny Hillis. Hillis has
been on one long, strange trip in the five years since Gordon Bell bet him a
crow-eating essay about the future of supercomputing. Hillis, the founder
and former chief scientist of parallel processing pioneer Thinking Machines
Corp. (TMC), Bedford, Mass., no longer holds that position. He has returned
to his roots at the Massachusetts Institute of Technology in Cambridge,
Mass., this time at the younger and more hip Media Lab rather than the
Artificial Intelligence Lab, dominated by Marvin Minsky, where Hillis first
cobbled together a design for a massively parallel computer.


TMC itself is just emerging from the dark side of a Chapter 11 bankruptcy
pleading. But to do so it had to die as a supercomputer maker and resurrect
itself as a software company. The twin forced march of Hillis and TMC
symbolizes the seismic shift that has rocked the traditional supercomputing
market.


Bell didn't exactly predict that shift, but he did bet on Hillis'
comeuppance. Bell, the legendary computer designer who created the PDP and
VAX series of machines at Digital Equipment Corp., Maynard, Mass., and who
is now dispensing wisdom for Microsoft Corp. from his office in Los Altos,
Calif., bet Hillis that massively parallel supercomputers might not actually
be the greatest revolution to hit the scene since the invention of rock and
roll. Specifically, in 1990, Bell bet Hillis that in the last quarter of
1995, the highest number of sustained MFLOPS (millions of floating point
operations per second) would be generated by a machine with fewer than 100
processors, rather than by a machine with many processors (more than 1,000).
The wager concerns only supercomputers, or Big Iron machines costing more
than $1 million and used for scientific purposes. That bet was chronicled in
the January 1992 issue of Upside.


The judge, jury and keepers of the bet are John Hennessey, a professor of
electrical engineering and computer science at Stanford University,
Stanford, Calif., and David Patterson, a professor and chair of the computer
science division at the University of California, Berkeley. The loser has to
'fess up his humbling defeat to the world in writing.


The prohibitive favorite when the bet was made, Hillis is now such an
underdog that Las Vegas bookies wouldn't touch him. "The only reason I might
lose is if people get a few huge machines only doing floating point
operations and lock them in a room and they just sit there and grind out
MFLOPS," says Bell. "They'd be grinding out a shitload of numbers just to
win the bet." A Cray T3E is still probably the most powerful machine not on
wheels. "The [massively parallel] stuff is a bust," says Bell. "There's no
market for it because no one can get applications for the machines."


If Bell, whose volume and frequency of pontification make the late Howard
Cosell seem mute by comparison, had said that three years ago, even his most
ardent admirers would have questioned his sanity. Until recently, many
people considered massively parallel machines to be the savior of
high-performance computing.


But the trip to the future turned out to be even stranger than Bell might
have predicted. It's not just the massively parallel machines that have
disappointed their creators and investors. We're talking upheavals in the
high-performance computing world-alias Big Iron-that no Richter scale could
measure. And the aftershocks have only just begun. "There is no future for
the Big Iron systems," declares Michael Burwen, director of the Palo Alto
Management Group, a high-performance computing market research company in
Palo Alto.


Now even Hillis is humbled. "The big surprise is [that] the supercomputing
market basically doesn't exist anymore as a definable market," he said last
October at the Media Lab's 10th anniversary party. "It's very clear that the
dinosaurs are dying."


Many of the carcasses, in fact, already line the Information Highway:
Alliant, Scientific Computing Systems, Multiflow, Floating Point Systems and
Supercomputing Systems Inc. And in just the last year or so, three more Big
Iron companies joined the list of roadkill: TMC (at least as a Big Iron
company), Kendall Square Research and Cray Computer.


The high-end market, in fact, never ended up as anything more than a tiny
blip on serious financial radar screens. The Smaby Group, Minneapolis,
estimates last year's entire market for high-performance
(scientific/technical/engineering) computing at $2.05 billion. The company
projects that in 1999 the market will barely creep over $3 billion. But the
true, high-end supercomputer segment of that market is beginning to crumble
away. Smaby Group projects that the top end of that market-machines costing
more than $5 million apiece-will decline by 6 percent in that time.
Similarly, Chris Willard, manager, high-performance technology, in the
Mountain View, Calif., office of Framingham, Mass.-based International Data
Corp., predicts that by 1999 the revenue for the traditional supercomputer
market will drop to $767 million from the 1994 total of $877 million, a
negative 2.6 percent compound annual growth rate.


Market watchers, however, are not yet ready to bury the entire Big Iron
market. Willard concurs that the high-end market "is not really dead, but
it's certainly not growing."


But if it's not a dinosaur, it's at least a white rhinoceros at the top of
the Endangered Species List. That's a huge contrast from the days when
Hillis and Bell were betting on which type of supercomputer would get
bragging rights to the title of fastest in the world. After all the hype and
promise, most of the racehorses collapsed before reaching the finish line,
and Hillis and Bell have found themselves betting over tombstones. "From
that changed market standpoint, I think we both lost," says Hillis. "The bet
may fall in the dead zone."


The Bigger They Are...


It was one hell of a ride, though. A few short years ago, supercomputers and
minisupercomputers were as hot as the World Wide Web is today. Venture
capitalists were pouring in money and new companies were sprouting up every
few weeks.


But today, Cray Research Inc., Eagan, Minn., still towers over the
traditional supercomputing business. And perhaps it should. Cray Research
has defined the excitement over supercomputing since it was founded in 1972.
As soon as the first Big Iron box sprang from Seymour Cray's head, it was
crystal clear that the thing had attitude-simultaneously mysterious,
ethereal and fascinating. There was a unique majesty to the word
supercomputer. It crunched numbers and solved problems that previously had
been the stuff of dreams.


The genius of supercomputers, Seymour Cray, changed the world-and made one
hell of a lot of money doing it-as the master of the bipolar-logic
superprocessor. When the supercomputing universe centered around
single-processor systems, he who made the fastest CPU ruled. In that domain,
Seymour Cray was unbeatable.


But in recent years, even the mighty Seymour has begun to look outmoded.
While most people recognized that killer microprocessors are taking over the
field, Cray refused to yield. He tenaciously-some say stubbornly-clung to
his mission of making the fastest, most powerful superprocessor. In 1989,
the pioneer left the company he founded to start another venture, Cray
Computer Corp., to create even more powerful superprocessors from gallium
arsenide.


But the superprocessor crusade, which had worked so well for so long, had
become an anachronism. In 1990, Lawrence Livermore Laboratory agreed to pay
$42 million to be the proving ground for an eight-processor Cray-2 and
Seymour's newest creation, a 16-processor Cray-3. But the Cray-3 was a
stillborn prototype.


Because of Seymour's reputation, the mere possibility of completing the
Cray-4 kept the company alive long after less tolerant-or
hero-worshipping-creditors would have ripped out its life support systems.
Death came last April, not with a bang but a whimper when financing finally
dried up.


Steve Chen, another supercomputer genius from Cray Research, also ran away
to fight another day. Reveling in his X-MP and Y-MP glory, he formed
Supercomputer Systems Inc. (SSI) after convincing IBM that he had the cure
for its supercomputing sickness. After four years and possibly as much as
$250 million from Big Blue, however, Chen proved incapable of walking the
walk. He allegedly produced a prototype based on superprocessors, but he
never sold a machine. So a sadder but wiser IBM, now a self-made
supercomputing heavyweight, pulled the plug in 1992.


Unlike his mentor, however, Chen recently demonstrated that he may not still
be crazy after all these years. Funded entirely by MCSB Systems PTE Ltd., a
Singapore-based technology conglomerate, Chen Systems Corp. (formerly
SuperComputers International) in Eau Claire, Wis., began beta testing its
Pentium Pro-based Chen 1000 server line last April. Its eight-processor
machine was released on September 18, mostly to yawns. The company at the
time claimed it had 20 orders for the new machine, and if so, that's an
impressive debut. But it's hardly a supercomputer. As one source who
declined to be identified asked, "What's the big deal about another Intel
microprocessor-based machine?"


Unable to match Cray at the high end, most potential rivals went low. They
gradually realized that more microprocessors meant more power to the people.
They coupled a few, then tens, then dozens, then hundreds in a single
system. They didn't fare any better, however.


No company rose higher quicker and fell lower faster than the classic of
parallel processing, Thinking Machines. TMC at the beginning of the decade
had sales of $65 million and thoroughly dominated the burgeoning parallel
processing market it had created. The future seemed limitless.


Hillis had his own worshippers, including Steven Squires, director of the
Computer Systems Technology Office at the Advanced Research Projects Agency
(ARPA). Through targeted grants and sweetheart deals, ARPA came very close
to violating the spirit, if not the letter, of government contracting
regulations, enabling Thinking Machines to lead its charmed life.


Since its May 1983 founding on little more than a wing and a prayer, TMC
never had to worry about business plans, competitive strategies or the color
of its balance sheet. Every time the company leaked, ARPA patched the hole.


But the days of the future passed in a nanosecond. With the Big Red Menace
broken and no longer the justification for an unlimited military budget,
ARPA pulled the plug.


The company's Connection Machine couldn't cut it in the real world. TMC was
also wracked by management turmoil. In August 1994 it sought shelter in the
Chapter 11 bankruptcy womb.


Kendall Square Research, another one-time high-flier in the massively
parallel world, even drew Bell as investor and consultant. Founder Henry
Burkhardt was a terrific technologist and a terrible accountant. After
revealing accounting irregularities in 1995, it also filed for Chapter 11.


The fate of the minisupers paralleled that of their big brothers. Convex
Computer Corp. was started in 1982 as a venture capital-financed
mini-supercomputer alternative to the expensive, massive Big Iron boxes that
dominated the high-performance computing market. The plan was to build
"affordable" supercomputers-in other words, smaller, cheaper and more
efficient machines.


Everything proceeded smoothly for almost a decade. Convex went public in
1986 and its stock traded at more than $20 by 1990. It had an impressive
string of consecutively profitable quarters. The company was all the rage on
Wall Street.


But when the Cold War went down, it also brought Convex with it. The Defense
Department stopped writing checks as freely as it did in 1983, when defense
official Richard Perle told a Congressional hearing that an Apple II was
capable of starting World War III. Convex's profits turned to losses-$140
million since 1993, including $15.3 million in the first half of this year.
Money went out faster than it came in; revenue plummeted to $144.2 million
last year from $231.8 million in 1992. "I hate to use the term, but this
paradigm shift is changing everything that has been the norm for the last 10
years," says Steve Wallach, co-founder and senior vice president of
technology at Richardson, Tex.-based Convex, now the Convex Technology
Center of Hewlett-Packard Co., which bought the company last September for
$150 million.


Back From the Dead


What happened to the market? The real problem was that the market never
really happened. The breathless anticipation of the military turned into a
panting enthusiasm from entrepreneurs and venture capitalists, and none of
it was really deserved. Thinking Machines and others discovered that
corporations are not as free-spending as the military. Says Bell, "TMC got
in so much trouble because in the beginning they extrapolated, from a few
sales directly related to government placements, that there was a market for
huge processing machines. There never was a real market there."


For a while, the military enthusiasm seemed infectious. There was a certain
amount of prestige to owning a supercomputer. Apple Computer Inc. even
bought a Cray for its research efforts. But as the chill of the Cold War
thawed, million-dollar supercomputers became as popular as $500 toilet
seats. Business process reengineering, a buzzword that mostly means cutting
expenses and people, accelerated the trend. The CIOs and the MISers began
buying solution cycles, not just MFLOPS. Users stopped genuflecting at the
high-end altar and started asking what the machines could actually do for
them.


"The ultimate in hot-vector technology were the machines from [the extinct]
Cray Computer Corp.," says Stephen Brobst, managing partner at market
researcher Strategic Technologies and Systems, Cambridge, Mass. "They were
phenomenally fast but they weren't economically feasible." Adds market
researcher Burwen, "There won't be any vector machines sold in another five
years. There's no future for those systems."


At first it seemed as though massively parallel systems might in fact win
out as much more cost-effective than the vector machines. However, it was
much harder to create software for them.


The software steadily improved but not as fast as the microprocessors
themselves. Suddenly, real-world problems that people thought only Big Iron
could solve-with either vector or massively parallel machines-were being
tackled by machines with fewer than 100 processors, and getting the job
done. The knotty problem of making software run on a thousand
microprocessors simultaneously is often simply not worth tackling.
"Microprocessors are getting so powerful that even a small number of them
can handle problems that traditional supercomputers used to do," says
Burwen.


In a recent study, the Palo Alto Management Group predicted that the market
for parallel processing systems would increase at better than a 40 percent
annual rate to about $14.3 billion in 1999. Approximately three-quarters of
that amount will come from commercial applications such as on-line
transaction processing (OLTP), decision support systems (DSS) and
multimedia. Science and engineering will still be important markets but will
lose ground by century's end.


The market will not be fed by Big Iron machines. "There will be no
1,000-processor machines sold this quarter," predicts UC Berkeley's
Patterson. "Even machines costing a few million dollars are very unlikely to
have more than 64 processors. There's no demand for the traditional big
supercomputers anymore."


Companies are surviving by adopting the new religion that says smaller can
be better. Even Seymour Cray's original company, Cray Research, once
synonymous with Big Iron, has gotten the drift. Its share of the traditional
supercomputer market is actually rising as competitors die off,


but the market is shrinking. A few years ago Cray realized that it could no
longer live on supercomputers alone. In 1991 it created Cray Research
Superservers after purchasing selected assets of Floating Point Systems.


Last February, Superservers merged with two other groups and became the
Business Systems Division. That entity's sole offering is the CS6400
enterprise database server, a multiprocessor system based on the 85MHz
SuperSPARC II microprocessor from Mountain View, Calif.-based Sun
Microsystems Inc. Cray recently demonstrated a 48-processor machine cranking
away at a 1.6 terabyte database. Cray claims the CS6400 has penetrated more
customer sites in a shorter time than any competing equipment. Cray's
commercial customer base has grown so that it currently accounts for more
than half of all Cray sales revenue.


"Even major parts of Cray believe [the traditional supercomputer market] is
dead," says consultant Brobst. "But they've done a terrific job engineering
the 6400 and moving into the commercial market. Only Cray could have done
that."


Thinking Machines has dropped out of the Iron business altogether. In
October 1994 it reorganized as a software-only company and promptly went on
to four consecutive profitable quarters. It apparently has learned its
lesson well.


"The market spoke loud and clear, and told this company that building some
of the world's fastest computers was not, by itself, enough to sustain
growth," TMC president and CEO Robert Doretti admitted recently. "We took a
hard look at our core competencies and quickly realized that we had
substantial expertise in the software that harnesses the power of
multiprocessor computers. We believe there's a huge untapped market for this
capability, and this is our strategic focus going forward."


Staying Alive


There are, however, naysayers to all this negativity. "It's a big
misperception that the high-end supercomputer market is dying," contends
Cray Research President and COO Bob Ewald. Just last November, Cray sold a
top-of-the-line 32-processor T90 supercomputer to Nippon Telegraph and
Telephone Corp. (NTT). According to Cray Chairman and CEO J. Phillip Samper,
demand at the end of the third quarter for the T90 line represented 45
percent of the company's backlog. Cray also claims to have $90 million in
orders for its T3E supercomputer, which isn't scheduled for delivery until
the first quarter of 1996.


And what does it say about the condition of a market when Tera Computer,
which since its founding eight years ago has done absolutely nothing but
lose money-a mere $9.5 million-goes public and oversubscribes the offering?
"Big Iron is not stone-cold dead," contends Patrick Grady, senior vice
president, corporate finance, at H.J. Meyers & Co., the Rochester,
N.Y.-based underwriter of Tera's IPO. "It's contracting for the companies
still able to play in it, but there's still plenty of business there. People
are waiting for a new approach. Vector processing machines are long in the
tooth. Parallel processing machines are good but not exceptional. The
high-performance world needs a major breakthrough in programming." He
believes that Tera is the company to do exactly that.


In fact, the technical supercomputing market could continue to stumble along
for several more years. Bell paints a picture with several types of
surviving supercomputers, including Cray-style evolutionary supercomputers
with multiple vector processors; multicomputers formed from
microprocessor-based workstations connected via high-bandwidth, low-latency
switches; and "multis," or multiple microprocessors connected to large
caches that access a common memory via a common bus.


Bell also believes that a few trends could keep the momentum for such
machines going for a long time. For one, the government's "buy U.S." policy
is still alive, although not as visible as it was a few years ago. For
another, Cray, IBM and Silicon Graphics Inc. have large installed, loyal
customer bases for their supercomputers and super servers.


In fact, Intel Corp. announced the sale of a massively parallel machine just
a few months ago. The Department of Energy's ASCI (Advanced Scientific
Computing Initiative) program is paying $46 million for a 9,000-processor
machine, which it claims is the first teraflop box. It's scheduled to be
installed at Sandia National Laboratory in New Mexico in 1996.


But don't look to it as a model for future supercomputer sales. "That's a
one-shot, boutique computer deal," Patterson scoffs. "It's a research
vehicle [to help develop a nuclear testing visualization program]. It's not
going to affect the high-performance market."


Mostly, the supercomputer business has become a market for niche players.
"Forget general purpose players," says John Harte, president of MasPar
Computers, Sunnyvale, Calif. "In today's market you'd better have a
specialty where you own a significant part of the niche you're playing in.
You can't maintain a competitive advantage unless you've got a specialized
enough architecture that the large players won't adopt." Some examples of
such niches: Data mining (extracting info from huge databases) and gene
sequencing.


In other words, supercomputing has returned to where it started: a very
specialized niche market. "We're not strictly back to the future, but we're
pretty close," says a White House technology policy advisor who asked not to
be identified. "Big Parallel Iron, and maybe all Big Iron, is reverting to
the niche market it was 20 years ago," when it was a small market dominated
by nuclear weapons, weather forecasting and a few industries, such as
petroleum and aerospace, that needed major horsepower.


"Now we've come completely around," the adviser says. "Big Iron supports a
few niche markets, and the rest of the world-because of the amazing increase
in computer power and distributed computing, now has on its desk machines
more powerful than Big Iron was 20 years ago. You use software tools and
couple these together over corporate LANs and public data networks and it
beats the shit out of Big Iron in cost effectiveness."


But those niches may no longer be the breeding ground for entrepreneurs.
"All the upstarts that entered the supercomputer business are either dead or
part of another company," market researcher George Smaby says. And those
that aren't-MasPar, NCube (which is nominally independent but would
disappear if Oracle Corp. President Larry Ellison decided he had better
things to do with his fortune), Meiko and Tera-seem one phone call away from
being eaten.


John Toole, director of the National Coordinating Office for High
Performance Computing and Communications, says that the huge amount of
capital required to play the high-performance game will prevent any startups
from joining the roster. And Hillis adds that technology is changing so
rapidly that buyers cut risk wherever they can.


"I'm not suggesting that there won't be startups," IBM's Barnes says. "What
people are really saying is that for a hardware vendor to get into the
high-end space without fantastic technology is almost impossible. The
capital costs are too huge."


For proof, just look at the Cemetery of the Innovators. "Everyone talks
about how great the high-performance industry has been for innovation,"
Barnes says. "But I'm not sure there ever was a time like that. The
half-dozen examples that we all talk about-Kendall Square, Thinking
Machines, Cray Computer-never really made it. Pieces of their technology may
show up in other products, but the companies aren't around."


So if all the innovators and tweakers and tinkerers and experimenters and
revolutionaries have turned to dust, who's still standing when the
high-performance sun sets?


Does anyone want to bet on the last "Big Iron company?" After all, there are
still 30 white rhinos left.


Willie Schatz heads the Schatz Group, a Washington, D.C., firm specializing
in technology, policy and communications. He is editor of HPCC Week. His
last article for Upside was "Cutting the Gordian Knot" in January 1995.


- ----------------------------------------------------------------------------


$40 Million Vaporware,Anyone? How's this for a hot prospect?


"The Company is a development stage enterprise that had an accumulated loss
of approximately $9.6 million as of June 30, 1995. The Company has
experienced net losses in each year of operations and expects to incur
substantial further losses while it builds its MTA (Multithreaded
Architecture System) system prototype and commences production, and possibly
thereafter. The Company has had no revenue or earnings and does not expect
to recognize revenue from the sale of its MTA system sooner than the second
half of 1996, if ever."


Sounds like the kind of statement that just builds confidence, doesn't it?


Well, that's the wording from the September 25 IPO statement of
Seattle-based Tera Computer Co. The company went public in order to raise
money for its nonexistent supercomputer, whose selling price it expects to
set between $5 million and $40 million.


Since its December 1987 inception, Tera has spent $27 million to develop the
MTA system. More than $18 million of that was a gift from the Advanced
Research Projects Agency (ARPA). Still, the prospectus notes that the MTA
system "has been subject only to computer simulation and the Company has not
yet built its initial prototype."


So who would buy stock in such a company? Tera put 850,000 units up for
sale, each consisting of two shares and one warrant. The stock price was $12
per unit. The warrant entitles the holder to purchase, at any time over a
five-year period starting September 25, 1995, one share of common stock at
$7.20 per share through March 24, 1998, and at $8.40 per share through
September 24, 2000, when the warrants expire.


A lot of people bought. The offering was a monster success, raking in $8.55
million ($9.9 million if the underwriter's over-allotment is exercised).


"We didn't have much difficulty selling it because we've got a very exciting
and contrarian story to tell," says Jim Rottsolk, Tera's president, CEO and
CFO. "It's unusual to go public when you have no revenue. But it's the
promise of the product that interests investors and the government."


Indeed, the government seems especially interested. Aside from helping to
fund the company's research, ARPA is first in line to buy its products-once
they're done. Last January the agency signed a contract with Tera that
provides ARPA with options to purchase up to $20 million of MTA systems for
early evaluation over the next three years. Tera is also negotiating a
contract with ARPA to jointly fund the development of certain components of
its next-generation MTA system.


In addition, the San Diego Supercomputer Center (SDSC) has submitted a
proposal to ARPA to purchase Tera's first production MTA system. ARPA has
told Tera that it plans to exercise an option under the January 1995
contract to buy an MTA system to place at SDSC. (Calls to ARPA officials
seeking comment about this expenditure of taxpayers' funds were not
returned.)


This kind of backing indicates there's a new guru in town. "This is
absolutely startling," says Bob Stern, a Washington-based IT consultant.
"There seems to be a serious national investment in Burt Smith [Tera's
founder, chairman and chief scientist]. Seymour Cray was in the same
position six months ago-he needed $25 million but couldn't get it."


Stern also notes that this seems to be government supercomputer business as
usual. "This obviously shows that the days of the 'state computer' [Gordon
Bell's term for companies kept alive by the federal government] aren't dead.
It's amazing that Smith pulled this off."


Get real, counters Patrick W. Grady, senior vice president in the San
Francisco office of Rochester, N.Y.-based H.J. Meyers & Co. Inc., the
offering's underwriter. Grady claims that he has been approached by many of
Tera's competitors several times and hasn't given them a damn thing. "I'm
completely unconvinced that Cray Research holds the future of
high-performance computing," he adds.


Tera's management insists it does have the answer. According to Vice
President Gerald Loe, customers are desperately seeking general purpose
scalable parallel machines with large-scale memories. Tera promises to
create such machines because it claims it can solve the memory latency
problem that slows down other architectures. Grady contends that cure
extends all the way to the desktop.


"We know the machine's not built," Rottsolk concedes. "We know no one knows
if the hardware works. We know that everyone could lose all their money. But
we strongly believe that we're going to release our prototype in the first
quarter of 1996 and deliver it by June 30.


"Right now, though, you've got to take it on faith because the product
doesn't exist." Amen.-W.S.


----------------------------------------------------------------------------


Associated Links:


   * Cray Research is not conceding the end of the supercomputer business,
     however. It recently announced the Cray T3E Scalable Processing System.