Researchers Ponder The Next Step In Supercomputing: The Exaflop

Sandia and Oak Ridge national labs aim to bust the million trillion calculations-per-second barrier

Time was, a teraflop (that's one trillion, or 1012 floating point operations per second) was just a dream. But the supercomputer ASCI Red nabbed that prize in 1996. Since then, it's been the grueling, relentless march to a petaflop--that's 1015 flops for those keeping count--a goal achieved by the Riken MDGrape-3 computer in 2006 (some dispute this claim, as the machine is so specialized it can't properly run the benchmark software. For them, we present the latest iteration of IBM's Blue Gene/P, which is purportedly capable of a petaflop as well). Now it's the exaflop, 10^18 flops, that puts a twinkle in the eyes of supercomputer aficionados.

The Institute for Advanced Architectures, a joint project between Sandia and Oak Ridge national labs, has started preparing the foundations upon which to build the world's first exaflop computer, and they have $7.4 million from the government to get things going. One major challenge is data management -- it's relatively easy to build a faster processor, but it is not so easy to get data to hungry processors in a timely fashion. Another is the software architecture, and how to most efficiently distribute the processing to each of the cores. Finally, with current technologies, the IAA calculates that an exaflop computer would need tens of megawatts of power to operate, costing the lucky owner some tens of millions of bucks a year in energy bills. Researchers with the IAA hope to bring down the energy requirements drastically to keep the computer relatively economical, at least when it comes to the Con Ed bill.

No word on a timeline (it only took a decade, give or take, to get from a teraflop to a petaflop), and who knows what kinds of hurdles will materialize once research gets going, but, hey, what's another 10 years to wait for better weather reports?