Startup Turns To Specialized Processors For Low-Cost Supercomputing Power

PeakStream, a Silicon valley startup, has released beta software designed to boost the performance of scientific and financial applications up to 20-fold by letting them tap graphics chips and other specialized processors.

The technology is designed to feed the need for supercomputing in industries including oil and gas exploration, investment banking, and defense. Benefits could trickle down to mainstream computing as Advanced Micro Devices, Intel, and Sun Microsystems release chips with multiple processing cores that could benefit from similar techniques.

PeakStream, with 35 employees, was founded by Pat Hanrahan, a Stanford University computer science professor, and Matthew Papakipos, a former technical executive at Nvidia, and is backed by $17 million in venture capital from Kleiner Perkins Caufield & Byers, Sequoia Capital, and Foundation Capital. The company released an evaluation version of its software, which lets standard C and C++ programs enlist the aid of graphics chips to execute complex mathematical functions at blazing speeds. PeakStream's development tools and run-time environment can execute portions of high-performance computing apps on FireGL and Radeon graphics chips from ATI Technologies, which AMD is acquiring.

Programmers use a PeakStream API that lets applications call a virtual machine within a run-time environment on each machine in a computing cluster. The API specifies what math functions to execute, and the virtual machine calls a library of pre-compiled numerical software. PeakStream's compiler schedules those jobs to the ATI chips.

The idea is to let customers take advantage of specialized silicon without slowing down software development work, says PeakStream CEO Neil Knox. "There's a tremendous need for more and more computing power," Knox says. The company plans to release future versions of the software for IBM's Cell chip, as well as Intel and AMD multicore processors, which make programming more difficult.

The technology could capitalize on scientific users' demand for more computing power, says Margaret Lewis, director of commercial ISV marketing at AMD. "You have to recompile your applications and use their APIs," she says. "But for companies willing do that, you might have some takers."

The $9 billion high-performance computing market is growing at 20% a year as companies apply supercomputing power to new problems. A massive oil strike by Chevron and others in the Gulf of Mexico last month was enabled in part by computerized seismic imaging technology.