By Don Clark
U.S. officials revealed plans Friday to spend $425 million on
advanced supercomputer technology, the latest sign of the
government's determination to leapfrog China in a field often
linked to national security and economic competitiveness.
The Energy Department said it would install two International
Business Machines Corp. systems valued at $325 million at Lawrence
Livermore National Laboratory and Oak Ridge National Laboratory.
The project, called Coral, also includes Argonne National
Laboratory. The machines, which also include technology from chip
maker Nvidia Corp., will carry out calculations five to seven times
faster than the most advanced U.S. systems now in use, the agency
said.
The department will spend another $100 million to further
develop "extreme scale" supercomputing technologies as part of a
program titled FastForward 2. Argonne is expected to pick a
supercomputer under the Coral program later, the agency said.
Supercomputers, room-sized systems that comprise thousands of
microprocessor chips, perform tasks that include simulating nuclear
explosions, cracking encryption codes, projecting climate trends
and locating oil deposits. China's 2013 success in building a
system that topped a closely watched ranking of computer
performance--interrupting years of U.S. dominance--prompted calls
by U.S. scientists for greater government support.
Energy Secretary Ernest Moniz, who announced the projects Friday
at an event in Washington, D.C., said in prepared remarks that they
would foster "transformational advancements in basic science,
national defense, environmental and energy research."
Installation work on the two supercomputers is expected to begin
in 2017. Beyond their sheer scale, the new machines will exploit
new technologies in the race to solve tough scientific problems
more quickly.
IBM, for example, will employ what it calls a data-centric
design to reduce the need to shuttle massive amounts of data within
the supercomputers. Nvidia, meanwhile, will contribute chips to
handle number-crunching as well as a new technology called NVLink
designed to transfer data between them at unusually high speed.
These new approaches are designed to reduce energy consumption
and address other emerging challenges, executives of the two
companies said.
Starting in the 1990s, computer makers started assembling
supercomputers by stringing together the kinds of processor chips
and other components used in personal computers. In recent years,
that technology has been augmented by chips from Nvidia and others
that evolved from technology used to render graphics in
videogames.
But such machines consume huge amounts of electricity. Improving
performance by simply stacking more of those components together
wouldn't be practical, industry executives and computer scientists
say.
Moreover, computer designers are grappling with an explosion in
the amount of data that users want to sift through, said Dave
Turek, an IBM vice president in charge of high-performance
computing efforts. Much of the computation carried out in efforts
to locate underground oil deposits, for example, is in preparing
data for analysis rather than the analysis itself, he said.
Without an expansion in computing capacity, the data-handling
challenge appears likely to come to a head in about three years,
Mr. Turek said, based on conversations IBM had with supercomputer
users.
"They were telling us, they saw a real problem around 2017," he
said.
To solve it, IBM added computing power to components--including
data storage and networking devices--that usually carry a light
burden of calculation. That way, the supercomputer can delegate
heavier computing tasks to those components and reduce the amount
of data that must be transferred and processed by central
processing units, Mr. Turek said.
The systems to be installed at Lawrence Livermore and Oak Ridge
will use future versions of IBM Power chip line to handle basic
computation chores. Other calculations are handled by a new version
of Nvidia's graphic chips, code-named Volta, said Sumit Gupta, an
Nvidia executive in charge of its accelerated computing efforts.
Mellanox Technologies Ltd. will supply other communications
technology.
The news from the Energy Department comes a few days before the
release of a twice-yearly ranking of the 500 largest
supercomputers, dubbed the Top500. It isn't clear whether the new
systems will surpass China's Tianhe-2 machine, which first reached
the top of the list in June 2013 and since has remained there.
Cray Inc. previously was selected to install supercomputers at
Lawrence Berkeley National Laboratory and Los Alamos National
Laboratory that are considered contenders for the top ranks of the
Top500 list. The company uses Intel Corp. chips in some of its
latest machines.
Horst Simon, deputy director of the Lawrence Berkeley lab and
head of its supercomputer efforts, said many U.S. scientists would
have preferred that the Energy Department had launched its latest
expenditures sooner.
"This investment is late, but it's still a significant
investment," he said.
Write to Don Clark at don.clark@wsj.com
Subscribe to WSJ: http://online.wsj.com?mod=djnwires