[108290] in Cypherpunks
The Super-Duper Hypercomputer
daemon@ATHENA.MIT.EDU (jei@zor.hut.fi)
Thu Feb 11 09:14:05 1999
From: jei@zor.hut.fi
Date: Thu, 11 Feb 1999 15:51:49 +0200 (EET)
To: cypherpunks@toad.com
In-Reply-To: <199902110458.FAA21811@replay.com>
Reply-To: jei@zor.hut.fi
http://www.wired.com/news/print_version/technology/story/17862.html?wnpg=3D=
all
The Super-Duper Hypercomputer
by Christopher Jones=20
3:00 a.m. 11.Feb.99.PST A small Utah-based company made its first
technology announcement this week, and what an announcement it was:
Star Bridge systems claims that it has developed the fastest, most
flexible computer architecture in the world, which will be used as the
basis for everything from Web servers to toasters in the future.
Star Bridge recently completed tests on its new "hypercomputer," a
supercomputer it claims is faster and more versatile than any others
on the market. Called Hal, the hypercomputer tested at a rate of 12.84
trillion calculations per second (TeraOPs) =96- equivalent to 60,000
times the speed of a 350-megahertz PC. Even better, Hal is the size of
a normal PC, weighs less than 150 pounds, and plugs into a normal wall
outlet.
"This system is not only smaller, but also a lot cheaper and runs on
less power [than other supercomputers]," said Brent Ward, executive
vice president of Star Bridge. "It applies across the realm of
information technology and electronics. Most devices with a chip in
them will be using our technology to do things faster in a smaller
space -=96 from toasters to VCRs to automobiles."
By comparison, IBM's Blue Pacific, one of the fastest supercomputers
built to date, performs about 3.9 trillion calculations per second and
is used for nuclear-weapons simulations at the Lawrence Livermore Lab
in California.
"It's like a B Western," said David Schwoegler, a spokesman at the
Lawrence Livermore Laboratory. "When you get a reputation for being
fast, everyone wants to take a shot at you."
Several observers noted that Star Bridge measured its Hal computer
using different benchmarks than the IBM computer. And, as yet, Hal
hasn't been tested with real live applications.
Gregory Benson, an assistant professor of computer science at the
University of San Francisco, explained that Star Bridge appears to
have estimated its performance figures based on a 4 bit integer add
operation, which is a much less intensive calculation than a 64 bit
floating point operation.
Supercomputers are normally measured in terms of FLOPS, or floating
point operations per second. Thus, the Star Bridge measurement of 12.8
teraOPs is not directly comparable to the IBM-ASCI Blue-Pacific
measurement of 1.2 teraFLOPS, he said.
"I wouldn't be swayed until I see an application running on it,"
Benson added.
Though its claims have been met with skepticism, the company's
founder, Kent Gilson, said he has spent 15 years hacking out code for
such a "re-configurable" computing design and has developed products
in the past that incorporate it. In the 1980s, he created a
128-channel add-in board for sound mixing on PCs, and since about
1987, has been working on the software design for Hal.
Hal's fundamental architecture is based on Field-Programmable Gate
Array (FPGA) chips, essentially blank pieces of silicon that can be
reprogrammed an infinite number of times. Xilinx Corp. invented FPGAs
about 15 years ago and has provided Star Bridge with free chips and
technical support to build its Hal machines.
"There are lots of research picking up steam now on FPGAs," said Mike
Seither, a spokesman for Xilinx. "A number of companies have
commercial applications now, taking advantage of them. IBM is using
FPGAs in switches for ATM networks."
Since the standard for creating ATM is in flux, he said, IBM created
FPGA circuits that can be reconfigured remotely when the standard is
hashed out.
The Hal computer has 280 Xilinx chips on 36 proprietary integrated
circuit boards. Star Bridge has also developed its own software
system, Viva, which is an operating system with programming tools to
develop specific applications. The Hal will cost about US$26 million,
and the company has 120 programmers to build the customized
applications for each machine it sells.
One thing that makes his system so fast is the hierarchical
arrangement of the FPGAs, which provides more control over the
information-carrying components of the system.
"You can configure the memory any way you want. Each FPGA has a
distinct memory environment, but you can have those memories any way
you want," he said. "You could have one master processor or one part
of one processor accessing any other part. You have fine-grain control
down to the bit level."
Gilson said his company has deals in the works with several telecom
and pharmaceutical manufacturers and is interested in developing
rendering systems for Hollywood and geophysical visualization systems
for oil companies. The first application under design is a pattern
matching algorithm for gene searching and predictive biological
systems engineering, he said.
Undaunted by the likes of IBM and Cray, Gilson said he expects to take
the world by storm when the system is demonstrated, running real
applications, later this year.
"It will be some time before people will believe this is real. But in
six to eight months, it will be an even bigger story, when the
applications roll out," he said. "It will take this economic process
before it will happen."
