Friday, February 22, 2008

HPC: Chasing the Next Big Buzz Word

Preparing groundwork for an exascale computer is the mission of the new Institute for Advanced Architectures, launched jointly at Sandia and Oak Ridge national laboratories.

An exaflop is a thousand times faster than a petaflop, itself a thousand times faster than a teraflop. Teraflop computers —the first was developed 10 years ago at Sandia — currently are the state of the art. They do trillions of calculations a second. Exaflop computers would perform a million trillion calculations per second.

The idea behind the institute —under consideration for a year and a half prior to its opening — is “to close critical gaps between theoretical peak performance and actual performance on current supercomputers,” says Sandia project lead Sudip Dosanjh. “We believe this can be done by developing novel and innovative computer architectures.”

Ultrafast supercomputers improve detection of real-world conditions by helping researchers more closely examine the interactions of larger numbers of particles over time periods divided into smaller segments.

“An exascale computer is essential to perform more accurate simulations that, in turn, support solutions for emerging science and engineering challenges in national defense, energy assurance, advanced materials, climate, and medicine,” says James Peery, director of computation, computers and math.

The institute is funded in FY08 by congressional mandate at $7.4 million. It is supported by the National Nuclear Security Administration and the Department of Energy’s Office of Science. Sandia is an NNSA laboratory.


So the race is on to start working on the exaflop system. We are also are looking at this. There's a paper that some of my coworkers did recently about a machine to do exaflop "class work" on climate simulations. The machine they came up with based on the best fit for modern tech was so ridiculously scary that it is inherently obvious that we are going to need some fundamental breakthroughs in computer tech to get there that we have a long road ahead of us. If a petaflop system is going to take over 20-30 MW now, just imagine what an exaflop will require!

Hint, it's much worse than a linear extrapolation.

Second Hint: to make it a sustained exaflop system will require probably more than 10 exaflops of peak and some very good and intelligent reworks of compiler tech.

When I can beat my coworkers into posting the paper I'll put a link here.

No comments:

Post a Comment