Intelligence agency has a cold plan for faster, cheaper supercomputing
- By Kevin McCaney
- Dec 04, 2014
The Intelligence Advanced Research Projects Activity has awarded the initial contracts in a five-year program that could up the ante in the field of high-performance—and low power—computing.
IARPA, the Intelligence Community’s research division, will work with IBM, Raytheon and Northrop Grumman in developing the first phase of its Cryogenic Computing Complexity, or C3, program, which aims to build a superconducting computer that can operate at exascale capacity—about 40 times faster than today’s fastest supercomputers—while requiring much less power than today’s machines..
For the intelligence community, exascale computing would be valuable in code-breaking and sifting through the big data generated by its cloud-based Intelligence Community Information Technology Enterprise. But should the C3 program realize a supercomputing breakthrough, the impact could be felt at many other agencies that deal with increasingly large data sets.
Supercomputer makers have made progress in recent years in increasing processing power while requiring relatively small amounts of energy, by designing low-power architectures and mixing graphical processors in with standard processors. But they still draw about 10 megawatts of power for 20 petaflops, or 20 quadrillion computations per second. Today’s fastest supercomputer, China’s Tianhe-2, has a theoretical peak of 55 petaflops. Getting to an exaflop, or a thousand petaflops, under the current model would result in some serious power and cooling requirements.
Current supercomputing architectures, which use CMOS, or complementary metal-oxide-semiconductor, switching devices and metal interconnects, “appear to have no path to be able to increase energy efficiency fast enough to keep up with increasing demands for computation,” according to IARPA. That’s why the agency is investigating superconductors, which operate at near absolute zero (minus 273 degrees Celsius) to eliminate resistance and power dissipation among circuits and transistors, thus resulting in lower power use and less heat generation. Superconducting switches use Josephson junctions, which work extremely fast while dissipating little energy.
IARPA estimates that operating in this cold, cryogenic environment could allow a supercomputer to produce 100 petaflops of performance for about 200 kilowatts.
IARPA isn’t expecting to build a superconducting supercomputer overnight—in fact, the C3 program is designed only to build a small scale model, to see if this approach can work.
The first phase of the program, expected to last three years, will focus on developing the technologies required for a superconducting processor. That work will focus on developing cryogenic memory with improved capacity and energy efficiency, and advanced superconducting circuits integrated with memory and other components. The second phase will focus on incorporating those technologies into a small prototype.
As with any radical new approach, there’s no guarantee the C3 program will succeed. But superconducting supercomputing, along with attempts to develop a quantum computer, could be the way to get beyond CMOS-based computing, which could be hitting its limits.
Kevin McCaney is a former editor of Defense Systems and GCN.