HPC

DARPA: Can we have a one-cabinet petaflop supercomputer?

With 'self-aware OS' so anyone can program it, please

By Lewis Page • Get more from this author

Posted in HPC, 26th June 2009 13:26 GMT

Free whitepaper – Operationalizing Information Security:

Famous US military crazytech agency DARPA has issued a challenge to the IT community: do you think it's possible to build a petaflop supercomputer that fits in a single air-cooled 19 inch cabinet and requires no more than 57 kilowatts of power? One which requires no special programming skills to use?

The challenge is laid out in a request for information issued by DARPA this week. The Pentagon edge-bleed bureau says it is "looking for feedback from the computing community" on its rough plans for a Ubiquitous High Performance Computing programme, in which machines with the same performance as today's most titanic supercomputers would become commonplace equipment - at least in the US defence department.

Present-day petaflop supercomputers weigh hundreds of tonnes and consume megawatts of power. Worse still, according to DARPA, their complicated massively parallel architecture calls for programmers to have a deep knowledge of the specific machine they are writing for in order to use its vast resources efficiently.

The military boffins say none of that is acceptable. They want a petaflop machine that fits in a single ExtremeScale cabinet and can be powered by a fairly normal 80-horsepower generator (that includes cooling, by the way). Furthermore, the UHPC supercomputer-in-a-box is to be programmable by ordinary coders who don't have a deep understanding of its hardware.

This last bit is to be achieved, perhaps, by a notional "self aware operating system" which will determine the best way to use the massive resources available to it without the need for human programmers to intervene.

Full details on DARPA's concept for the notional UHPC programme can be read in pdf here. The agency says that if feedback from the computing community is positive, it will move forward with a formal announcement inviting proposals for funding.

Naturally, as with all DARPA projects, it's relatively unlikely that this one will reach fruition. Petaflop computers will probably remain huge, power-hungry and difficult to program for the moment.

But if UHPC succeeds, we can expect some interesting developments. US boffins in charge of existing petaflop machines have already said that their equipment "throws open the door to eventually achieving human-like cognitive performance in electronic computers". As an example, they consider that such computers could be able to process 2D visuals in the same way the human brain does to achive such tasks as driving a car. To date, robotic vehicles can't develop a useful picture of road traffic merely from video - they generally use 3D-mapping laser radar or such techniques, so as to make the computing problems practically feasible.

A UHPC super-cabinet, though, might be able to navigate a relatively normal road vehicle about with nothing more in the way of sensors than cameras. Other previously humans-only tasks might be automated fairly simply: at long long last, useful robot butlers and the like might be on the cards.

Then there are the artificial-intelligence implications. The existing Roadrunner petaflop machine appears to be equal in media-yardstick terms to either 45 or four-and-a-half Reg standard mouse brains, depending on allowance for the speed of the simulated murine intellect. This equals one visual cortex, a major chunk of a single human brain. (Evidently these mice aren't Douglas Adams ones, as they have between one and ten per cent of a human's brainpower.)

"Because there are about a quadrillion synapses in the human brain, human cognition is a petaflop computational problem," US nuke boffins at the Los Alamos lab stated last year.

If DARPA get their way, this sort of computing power will now be available in a single cabinet. Any decent-size data centre will in future be more intelligent than its human admin. ®

Free whitepaper – Enabling Datacenter and Cloud Service Management for Mid-Tier Enterprises