Imagine 1,000 trillion ops per second

[Eugene.Leitl at lrz.uni-muenchen.de]

(((sorry for the HTML garbage earlier, blame zdnet)))

http://www.zdnet.com/filters/printerfriendly/0,6061,2676421-2,00.html


magine 1,000 trillion ops per second
By Stephen Shankland and Erich Luening, Special to ZDNet
January 19, 2001 12:25 PM PT
URL: 

Compaq Computer is teaming with nuclear research facility Sandia National
Laboratories and biotechnology company Celera Genomics to build what is
anticipated to be one of the fastest supercomputers in the world. 

The deal, announced Friday at a Department of Energy news conference in
Washington, will result in a computer that can perform 100 trillion calculations
per second--100 "teraflops"--by 2004, Bill Blake, vice president of
high-performance technical computing at Compaq, said in an interview. A
second phase, which will be open to other bidders besides Compaq, will result in
a "petaflop" machine 10 times faster, the same speed as the Blue Gene machine
under design at IBM. 

The machine will benefit from computer hardware and software research from
all three partners, Blake said. The computer will be used to extract medically
useful information from raw databases of genetic information, but Sandia
officials said the advances also will benefit the nation's nuclear weapons
program. 

The deal is one of a series of major contracts for Compaq, whose Unix
computers using the Alpha chip are selling well with technical customers but
not as well with commercial buyers who prefer computers from Sun
Microsystems, Hewlett-Packard and IBM. 

Other major Compaq victories include a $36 million partnership with the
Pittsburgh Supercomputing Center to build a 2,728-CPU machine for academic
use and ASCI Q, a $200 million nuclear weapons simulation machine with
12,000 CPUs to be built at Los Alamos National Laboratory. 

Supercomputer competition
Compaq must reckon with competitor Sun, which is trying to convert its
business success into supercomputer success, and IBM, which has jumped into
the top spot in supercomputer popularity and speed rankings and has the
current fastest machine, a $110 million, 8,192-CPU machine that can perform
12.3 teraflops, about one-eighth the projected speed of the upcoming Compaq
machine. 

Houston-based Compaq is also developing a supercomputer for the Department
of Energy's National Nuclear Security Administration, which is expected to be
able to perform 30 trillion operations per second once delivered in 2002. 

Rockville, Md.-based Celera for years competed with the federally funded
Human Genome Project to compile a map of human genetic information. The
two efforts joined forces at the last minute, and now Celera and other
companies are trying to benefit from the research. 

The new supercomputer will help in further genomics research and the study of
the structure, function and interactions of proteins in the cells of humans and
other organisms, according to a joint press release. 

Celera will provide applications that life science researchers can use in their
work, such as further mapping of the human genome and other biological
research. 

Although terms of the agreement were not disclosed, the three companies plan
to spend millions of dollars, employ hundreds of workers, and develop new
technology as part of the effort, scheduled to be completed in 2004. 

At the press conference Friday, government officials, researchers and industry
executives hailed the supercomputer as the next step in the Human Genome
Project. 

Advancing human health
With years of experience building computer models of nuclear explosions,
Sandia will assist Celera in building software to analyze biological data. 

The supercomputer will help "to advance the knowledge from the human
genome to improve human health," said Sandia President Paul Robinson.
"Nothing beats the complexity of the Human Genome Project and the
opportunities ahead. We look forward to working in these fields." 

For Compaq, the deal is an example of its ongoing push to add servers and other
high-end computing products to its successful PC business. 

"Compaq has been reforming itself into providing computer and Internet
technology to a whole span of different fields," said Compaq Vice President Bill
Blake. "Part of this move from the desktop to the supercomputer has been with
these partners." 

The new supercomputer likely will have 10,000 to 20,000 CPUs, Blake said.
That number is sufficiently large that the system designers must grapple with
reliability problems not seen in typical business machines, even those with a few
dozen CPUs. 

"You start out on the basis that something will always be broken in the system
somewhere," Blake said. Then designers build monitoring and job scheduling
features that make sure jobs are reassigned to functioning CPUs if one fails. 

The Compaq machine will take advantage of future Compaq Alpha chip designs
that build networking abilities straight into the chip itself, Blake said in the
interview. This feature will allow more CPUs to communicate directly with each
other or to route shared information along from one chip to another as fast as
possible. 

"We're optimizing a special form of network between all the Alphas," Blake
said. 

This architecture varies considerably from that used by IBM, which prefers to
have a smaller number of nodes, each with multiple CPUs and connected to
central high-speed switches.