At GM, High-Performance Computing Curbs Test Costs

The automakeralong with Boeing, DreamWorks, Eli Lilly, and other companiesis using high-performance computers that run at almost 1 petaflops to improve products we use every day

Crashing cars doesn't come cheap for automakers such as General Motors. Some of the autos put through paces at GM's safety lab in Milford, Mich., cost more than $300,000 to make, though they retail for one-tenth that price.

The test vehicles are expensive because they must be specially built without the automation that makes mass production cheap.

GM is slashing the cost of testing by using high-powered computers made by International Business Machines (IBM) that let it simulate crashes that can range from a rollover off a windy mountain road in Mexico to a head-on collision on an oil-slicked highway in Los Angeles. The company has reduced the number of physical lab crashes to 300 to 400 a year, from as many as 600 three years ago, saving time and tens of millions of dollars, GM executives say.

"It increases safety and overall vehicle performance," says Ken Bonello, GM engineering group manager, in an interview in Milford, about an hour's drive from GM headquarters in Detroit. The technology at the heart of GM's cost savings—high-performance computing—is done by machines that can carry out almost 1 petaflops, or 1 quadrillion calculations per second. Supercomputers, the fastest class of high-performance computers, cost more than $500,000, according to research firm IDC in Framingham, Mass.

As of September, the fastest-known supercomputer was the Jaguar, built by Cray (CRAY), which has a peak speed of 2.33 petaflops—or more than 2 quadrillion calculations a second, according to the National Center for Computational Sciences in Oak Ridge, Tenn., where Jaguar is located. IBM is developing a supercomputer called Sequoia that will have a peak speed of 20 petaflops, or 20 quadrillion calculations every second—equal to two million laptops.

"This is a game-changing technology"

Engineers at GM use high-performance computers to simulate the new 2011 Chevrolet Cruze, while Boeing (BA) used them in developing the 787 Dreamliner. These machines help animators at DreamWorks Animation SKG (DWA) render movies such as Shrek and Kung Fu Panda, while Eli Lilly & Co. (LLY) scientists use them to research new pharmaceuticals. Chevron (CVX) used high-performance computing to do seismic imaging that led to the discovery of new reservoirs of oil in the Gulf of Mexico and Speedo International took advantage of it to model the swimsuit Michael Phelps wore at the 2008 Olympics.

"This is a game-changing technology, it cuts across all sectors," says Cynthia McIntyre, senior vice-president at the Council on Competitiveness, a nonprofit group of corporate chief executive officers, university presidents, and labor leaders. IBM, based in Armonk, N.Y., had 29.3 percent of the global market for high-performance computers, which totaled $8.64 billion last year, according to IDC. Hewlett-Packard (HPQ) was a close second, with a 28.6 percent share, followed by Dell (DELL), which had 12.7 percent. Sun Microsystems, now owned by Oracle (ORCL), and Cray each held about 4 percent of the market. The market may rise 5 percent to 7 percent this year and reach $10.5 billion by 2013, the researcher says.

GM showcased testing capabilities at its Milford site in late September. During one demonstration, red lights flashed and spotlights focused on a white Camaro, illuminating video cameras and sensors on the car's exterior and a crash-test dummy inside. Seconds later, 3,300 pounds of steel slammed into the driver's side at 30 miles an hour, emitting a loud bang with a heavy crunch of metal that startled witnesses behind plexiglass.

Some test vehicles can cost as much as $600,000, not including a test dummy that can increase expenses by more than $100,000, according to Albert Ware, director of GM's crashworthiness lab. GM complements physical crashes with virtual tests in an office building in Warren, Mich. There, engineers test everything from the impact of rollovers on drivers of various sizes to the effects of a collision on a pedestrian child.

Following each virtual crash, the company looks to make design changes. For instance, GM altered the hood of the Cruze to ameliorate the damage to a person falling on one after a collision.

vast, expensive power consumption

Virtual crash testing is no substitute for the real thing. Users of high-performance computing face other challenges, too. The software can be difficult to use, says Bill Hilf, general manager of technical computing at Microsoft (MSFT). "The complexity is high," he said. "The only way people can use this sort of capacity is if the barriers are dropped."

High-performance computers are also power-hungry. The largest supercomputers can consume from 1 megawatt to 10 Mw of energy, says David Turek, IBM's vice-president of deep computing. At a retail level, 10 Mw cost about $20 million a year, he says, adding that supercomputing facilities typically negotiate deals with suppliers.

Still, high-power machines have become more accessible to companies as computing power increases and hardware costs fall. This means that GM, which has for decades used high-performance computers in auto design and development, has recently begun using the technology in new areas such as engineering more fuel-efficient vehicles and replacing high-end photography in marketing materials with computer-generated imagery.

GM draws mainly on IBM's BlueGene high-performance computing system. For projects that require greater computing power, such as research into heightening fuel efficiency in engines or the potential of biofuels, GM works with researchers and supercomputers at Sandia National Labs in Albuquerque, N.M., and Lawrence Livermore Labs in Livermore, Calif.

Advances in computing have helped GM bring auto safety testing a long way since the 1930s, when it began crashing cars in Milford. Back then, a tester clad in leather would stand on the vehicle's running board, steer a moving car toward a large tree, and jump off just before impact, according to Ware. These days, comparable tests can be carried out over and over again at a fraction of the cost—virtually.