A Lightning Bolt from Big Blue


By Otis Port IBM has just switched on the biggest number-crunching beast in private industry. Located at Big Blue's Thomas J. Watson Research Center in Yorktown Heights, N.Y., the new computer can spit out more than 91 trillion calculations a second. Yes, you read that right: 91 trillion calculations a second, or 91 teraflops, in industry jargon.

This incredible speed ranks it as the world's second-fastest supercomputer. The only speedier machine is the mammoth, 367-teraflops system that IBM (IBM) is installing at Lawrence Livermore National Laboratory for research on nuclear weapons. Both are based on IBM's innovative BlueGene/L architecture. However, another IBM computer, now also being installed at Lawrence Livermore, will edge into the No. 2 slot by a slim margin later this year. (For a list of the world's current and upcoming speedsters, see BusinessWeek's Top 25 Supercomputers.)

LINING UP. What sort of research needs such souped-up computers? A lot of today's problems in engineering and science. For example, in chemistry and physics issues abound that would keep every computer in the world grinding away for centuries. And aerospace, automotive, and biotechnology researchers know of problems so tough that they would take decades to solve. Obviously, nobody bothers to try. Not yet, anyhow.

For its new 91-teraflops monster -- dubbed Watson Blue Gene, or BGW for short -- IBM has plenty of down-to-earth work to keep it busy. Untangling the riddle of how proteins fold to work their magic in the human body is the initial target. Two others: probing new frontiers in semiconductor physics and learning to harness nanotechnology for tomorrow's chips and materials.

IBM will also make BGW available to outside researchers. Under a novel Energy Dept. program called INCITE (in case you're wondering, the acronym was cooked up from this: Innovative & Novel Computational Impact on Theory & Experiment), researchers from industry and academia will be awarded up to 5% of BGW's capacity. INCITE now steals 10% of IBM's 10-teraflops supercomputer at the National Energy Research Scientific Computing Center (NERSC) in Berkeley, Calif., and the 5-teraflops BlueGene/L machine at Argonne National Laboratory.

LIMITLESS HORIZONS. What makes INCITE so special is that it provides enormous chunks of continuous computer time to a single researcher or team. For instance, a study of combustion in jet engines was awarded 2.5 million processor-hours; that's 16 days on NERSC's 6,656-processor system. Traditionally, such huge problems have been given short blocks of time over intermittent periods. So polishing off one run can sometimes stretch into months, with lots of thumb-twiddling in between. And then come the multiple iterations that are almost always necessary to refine previous results.

The idea behind INCITE, says Ray Orbach, head of Energy's Office of Science, is to accelerate key scientific discoveries and boost U.S. competitiveness in such industries as biotech, computing, materials, and petrochemicals. "This could mean giving somebody the whole machine for a week or even a month," says Orbach, who hatched INCITE two years ago. "That had never been done before."

Powerful machines like BGW will "help to really accelerate discoveries in science and engineering," predicts Tilak Agerwala, a vice-president at IBM Research. With its blazing speed, he adds, researchers and engineers "can now ask more questions, test more theories, try more designs, and simulate more conditions than has been possible before."

SIMULATION WIZARDY. Orbach insists that access to supercomputers "needs to be more like electricity, so people can get the time they need when they need it." This metaphor is regularly used with respect to utility computing, on-demand computing, and grid computing -- three buzzwords for the hot trend of running programs on computers scattered hither and yon across the Internet. Most major hardware vendors have set up computer farms that rent out just the amount of processing time a company or research lab needs. Even humongous programs can be run without spending megabucks to buy a supercomputer.

This will be increasingly vital to both scientific advances and industrial competitiveness. Researchers are now bumping up against tough problems so convoluted that they can be tackled only with computer simulations. That's why Orbach and other gurus insist scientific progress will no longer be determined by the usual interplay between theory and experiment. "Simulation will be the third leg of the stool," says Orbach.

While clusters of personal computers and servers are adequate for many types of simulations, others will continue to require behemoths like BlueGene. "You can do things on a single machine that you just can't do by spreading the computing all over the place," says Orbach.

THE MISSION: "OUT-COMPUTE." In materials science and climate modeling, he notes, something can happen in one part of the model that has long-range effects on other parts of the model. If those distant segments of the model have been parceled out to physically remote computers, Orbach explains, "the other computers have to wait for data to come in, and the simulation bogs down."

Even minor hiccups in simulations could be detrimental to competitiveness. Bob Bishop, chairman of Silicon Graphics (SGI), summed up the situation nicely is his oft-quoted maxim: "To out-compete in the 21st century, U.S. industry must out-compute." So there's no end in sight to ever-faster supercomputers.

TOP 25 SUPERCOMPUTERS

Speedy computers have become indispensable to new breakthroughs in science, engineering, and business. A worldwide race is under way to build ever-faster machines for tackling ever-tougher problems. To keep tabs on this competition, BusinessWeek Online will regularly update this list of the top teracrunchers

Rank

Location and User

Computer (vendor)

Peak Speed teraflops*

**

U.S.: Lawrence Livermore Lab

BlueGene/L (IBM)

367

1

U.S.: Lawrence Livermore Lab

BlueGene/L (IBM)

183.5

**

U.S.: Lawrence Livermore Lab

ASCI Purple (IBM)

93.4

2

U.S.: IBM Watson Research

BGW (IBM)

91

3

Japan: Riken

Molecular Dynamics Machine

78

**

Germany: Leibniz Computing Center

HLRB-II (SGI)

69

4

Japan: University of Tokyo

Grape-6 (self-made)

64

5

Japan: Undisclosed

(Hitachi)

62

6

U.S.: NASA Ames

Columbia (SGI)

61

**

France: Nuclear Power Agency

Tera10 (Bull)

60

7

Netherlands: Astron

(IBM)

43

8

U.S.: Sandia National Labs

Red Storm (Cray)

41.5

9

Japan: Earth Simulator Center

Earth Simulator (NEC)

41

10

Spain: Barcelona Super Center

MareNostrum (IBM)

40

**

U.S.: Oak Ridge National Lab

(Cray)

40

11

U.S.: Lawrence Livermore Lab

Thunder (California Digital)

23

12

Switzerland: Ecole Polytechnique

Blue Brain (IBM)

23

13

China: Meteorlogical Admin.

(IBM)

22

**

U.S.: Army Corps of Engineers

(Cray)

21

14

U.S.: Los Alamos Lab

ASCI Q (Hewlett-Packard)

20.5

15

U.S.: Virginia Tech

System X (Apple Computer)

20

16

U.S.: Naval Oceanographic Office

Blue Wave (IBM)

20

17

UK: European Center for Medium-Range Weather Forecasts (ECMWF)

(IBM)

16.5

18

UK: ECMWF

(IBM)

16.5

19

U.S.: IBM

BlueGene/L Prototype (IBM)

16.4

**

Korea: Meteorological Admin.

(Cray)

16

20

U.S.: Nat'l Center for Supercomputing Applications (NCSA) Tungsten (Dell)

(Dell)

15

21

U.S.: Army Research Lab

John vonNeumann (Linux Networx)

14

22

U.S.: Wright-Patterson AFB

Eagle (SGI)

13

**

Japan: Atomic Energy Research

(SGI)

13

23

Japan: Riken

Riken Cluster (Fujitsu)

12.5

24

U.S.: Lawrence Livermore Lab

ASCI White (IBM)

12

25

Japan: National Aerospace Lab

Primepower (Fujitsu)

12

25

U.S.: Pacific Northwest Lab

Mpp2 (HP)

12

*A teraflops is a trillion (tera) floating-point operations per second (flops), or calculations where the decimal point isn't fixed.

**System has been ordered but isn't yet fully installed.

DATA: www.Top500.org and company reports

Port is a senior writer for BusinessWeek in New York


Race, Class, and the Future of Ferguson
LIMITED-TIME OFFER SUBSCRIBE NOW

(enter your email)
(enter up to 5 email addresses, separated by commas)

Max 250 characters

 
blog comments powered by Disqus