It’s normally the perennial goal of chipmakers to ensure precision in their microprocessors. But a while ago researchers figured out that, in some cases, allowing for a little uncertainty could actually help achieve better performance.
In a paper presented at the International Solid State Circuits Conference in San Francisco yesterday, researchers from Rice University described a prototype encryption chip based on this idea, and it is 30 percent more energy-efficient and seven times faster than today’s best technology. The chip employs a new type of logic that comes out of research in an emerging field called probabilistic CMOS (PCMOS), which was highlighted by Technology Review as one of the top 10 emerging technologies of 2008.
As the transistors on a chip shrink, the electrical “noise” produced by electrons flowing through them also increases. Instead of trying to squelch this noise by increasing the voltage, the basic idea of PCMOS is to lower the voltage and simply account for the increase in noise instead. This means that PCMOS circuits sometimes arrive at an incorrect result when performing a calculation. But, because engineers assign a probability that the result will stray from a given value, this uncertainty can be accounted for. The approach could work in chips that don’t require high precision, like audio and visual processors and those that make use of random results, such as encryption chips.
The new chip is specifically designed to implement cryptographic algorithms–the type of calculations used to protect personal information when making a banking transaction over the web. But Krishna Palem, a professor of electrical engineering at Rice and a pioneer the field, says that PCMOS could be used to improve the energy efficiency of a variety of chips.
As we talked on the phone, Palem noted that there were anywhere between four and 10 computers between us, all built to do a perfect job. But if they make a mistake from time to time, they can still function, and we can still hear each other, because our brain ignores the more subtle errors. Palem argues that the benefits of more energy-efficient portable devices outweigh these imperceptible errors. “You give up much less than what you get back,” he says.
He adds that his team is on track to develop probabilistic chips for video and audio compression in a year and a half.