Energy will boost supercomputer speed using latest version of HPPI switches

The Energy Department’s Accelerated Strategic Computing Initiative will turn to
HPPI switches with an aggregate switching capacity of 512 Gbps to crank up a supercomputer
to 3 trillion floating-point operations per second.

Silicon Graphics Inc. is constructing the 3-teraFLOPS machine for Los Alamos National
Laboratory in New Mexico. The switches, from ODS Networks Inc. of Richardson, Texas, are
based on the American National Standards Institute’s High Performance Parallel
Interface 6400 standard.

ODS Networks’ Gigabyte System Network 6400 switch operates eight times faster than
previous HPPI switches, said John Gibbon, vice president of engineering for ODS’
Essential subsidiary. Each of the 32 ports can transfer 6,400 Mbps.

“It was designed so everything could fit” on one application-specific
integrated circuit, Gibbon said. The switch has only four levels of priority to reduce
latency, he said.

DOE’s ASCI supercomputing project, which will model and simulate the nation’s
nuclear arsenal, is the first application for the switch. Silicon Graphics’ machine
likely will go online late this year.

IBM Corp. is scheduled to have another

3-teraFLOPS computer operating at Lawrence Livermore National Laboratory in California
about the same time, said Gil Weigand, deputy assistant secretary of Energy for strategic
computing and simulation.

“I’m fully confident we’ll bring the system up” on schedule,
Weigand said.

The ASCI timetable calls for tripling supercomputer speed every two years toward a goal
of 100 teraFLOPS by 2004. By then, computer simulations will have become the primary tool
for evaluating the nation’s nuclear stockpile.

The size and complexity of the simulations call for a computer capable of executing 100
trillion operations per second to get results within a reasonable time, which Weigand said
is four or five days.

ASCI grew out of the Comprehensive Nuclear Test Ban Treaty, signed by President Clinton
in 1993. Weigand said ASCI is one of the most aggressive scientific research programs ever
attempted by government.

“It’s a computer scientist’s dream come true,” he said. “No
one has ever done this before.”

Assembling the fast parallel supercomputers is only one element of the program, which
includes development of application code for the simulations and for interpreting and
verifying results. But the computer is the most visible aspect. Weigand said commercial
technology was moving toward 100 teraFLOPS, and ASCI will push it there 10 years early.

No one has ever before assembled the thousands of processors necessary for teraFLOPS
performance or written code for them. The first single-teraFLOPS computer, from IBM, is
now operating at Sandia National Laboratories in New Mexico.

“That computer has a queue on it today,” Weigand said, at a “low level
of production. There is no such thing as a high level of production” because bugs are
still being worked out. Use is limited to a few large, important problems.

One of the difficulties with the existing high-speed computer is that current HPPI
switches, which can handle 800 Mbps per port, are not fast enough.

“We’ve never been able to get enough bandwidth, and the latency is higher
than we’d like,” Weigand said.

Software algorithms can make up for some of the latency, but the faster switches should
provide a better solution, he said.

In addition to IBM and SGI’s Cray Research Inc. subsidiary, Intel Corp. also is
working on an ASCI teraFLOPS computer. Weigand said ASCI will take advantage of
off-the-shelf technology while pushing industry to develop more rapidly.  

About the Author

William Jackson is a Maryland-based freelance writer.

Stay Connected

Sign up for our newsletter.

I agree to this site's Privacy Policy.