Probabilistic computing to cut energy costs

Thermodynamics, the science which makes car engines more efficient, is likely

to come in handy in designing efficient computers. According to a research in

probabilistic computing, trading off accuracy for energy saving is a viable

option and Krishna V Palem from Georgia Institute of Technology, shows the way.

He is driving a research project funded by the US defense department in embed-ded

systems computing.

In Hyderabad, to showcase the probabilistic computing model to the delegates

at the High Performance Computing Conference, Palem said, "Power is the

biggest squeeze when high computing speed is the objective. And when the

question is how to make the battery last longer in embedded computing devices,

chemistry is a costly option and detrimental to the ecosystem as well."

According to Palem, energy consumption is dependent on the speed and accuracy

you want. If you model computing so that it meanders instead of taking a

straight line approach and can control the meandering behavior of computing to

the desired level, it can save energy at negligible cost.

Drawing an analogy with drunken behavior, where a person high on alcohol will

require more energy to walk in a straight line than meander, Palem said he was

inspired by one of the comments made by Richard Fineman, a Physicists, in a

series of lectures at the University of Washington. While many will find the

idea of choosing little bit of imperfection when you could actually have faster

and accurate computing alright, marketing the concept of making your computers

weak is likely to face a mindset resistance. However, Palem believes that this

may not be the issue because the mathematical model he has developed talks of

strong results from weak computing. "When one can guarantee that it will

not go off precision more than one in a million times, it is not so

unpalatable," he said.

According to Palem’s model, the trade-off is miniscule and in embedded

computing applications such as graphics and speech recognition, it does not

impact the final output so the potential for commercialization is huge. When the

control of the trade off is in the hands of the user, one can have the level of

quality one needs and save on energy.

"How to get more from the same combination of dollar investment is one

of the big issues in the industry today," said Palem and "If one can

settle for a little wobbly behavior from the chip it does wonders on the energy

front," he added.

Nandita Singh


(CyberMedia News Service)