DARPA ponders taking supercomputing to the extreme

The Defense Department wants to take supercomputing to the next level by funding the development of a new breed of supercomputers that will be smarter and faster and yet smaller and require much less power than today's massive machines.

DOD officials believe such computers will be necessary to make sense of the avalanche of data that will gush forth from tomorrow's network-tethered sensor systems. Current computer systems will not be able to handle the load.

The Defense Advanced Research Projects Agency is considering the possibility of starting a new program to guide and fund development of such systems, tentatively named the Ubiquitous High Performance Computing (UHPC) program. Last week, DARPA's Information Processing Techniques Office issued a request for information on a proposed program structure and goals.

"The UHPC program is seeking solutions that will explore the technologies and architectures required to enable the development of revolutionary computing architectures and systems and overcome 'business as usual' advances," the RFI states. "This can only be achieved via dedicated investment, hardware/software co-design, integrated design techniques and continuous innovation."

How would such a revolutionary system operate? For one, it would use far less power than today's systems. The agency would like such a system to be able to execute 50 billion floating-point operations/sec per watt of power. The RFI explains that each floating-point operation in that scenario could run at under 20 picojoules per operation, a small margin of the thousands or even tens of thousands of picojoules now required to carry out such an operation.

The new system would also have much smarter software. Programmers would not have to worry about the underlying hardware, which would make writing programs much easier. The operating system and runtime solutions will have to “behave like a self-aware system that 'learns' to address a particular problem by building self-performance models, responding to user goals, and adapting to changing goals, resources, models, operating conditions and even to failures," the RFI states.

DARPA envisions the program being carried out in five phases. The first phase will fund the development of conceptual designs. In the second phase, an execution model will be detailed, along with metrics to gauge the success of a system built from the model. Phase three will involve a full-scale simulation of the system. In the fourth phase, the winners of the UHPC awards will deliver systems based on the specifications. Finally, the fifth phase will involve modifying and refining the designs.

The deadline for responding to the RFI is July 27.