Researchers at Rice University have built yet another prototype of their probabilistic chip that is faster, more energy-efficient and wrong. By sacrificing accuracy, the chip is 15 times more energy-efficient that normal chips and is well suited for applications like video and audio processing where the human eyes and ears can make up for the computer’s inexactitude. Rice expects the chip to be in the real world next year in hearing aids and soon in a low-power tablet aimed at the Indian education market.
The research, which I covered in 2009 when Rice began discussing the concept of “probabilistic computing,” has advanced beyond the design stage to a prototype, which puts it closer to reality. When I touched base with the Rice folks last year, they had hoped to build their prototype during the summer, and could better explain how the chip worked through what researchers called “pruning.”
The crux of the breakthrough is that if a chip doesn’t have to get the right answer every time, it can perform faster and use less energy. Software helps keep the chips wrong answers within a functional realm and helps sooth out the bumps that wrong answers cause. Much like any underachiever can tell you, sometimes you don’t have to get 100 percent — 90 percent or even 85 percent can be enough. And like a student debating on how much to study, the more accurate you want to be, the more energy you burn. From the Rice release:
“In the latest tests, we showed that pruning could cut energy demands 3.5 times with chips that deviated from the correct value by an average of 0.25 percent,” said study co-author Avinash Lingamneni, a Rice graduate student. “When we factored in size and speed gains, these chips were 7.5 times more efficient than regular chips. Chips that got wrong answers with a larger deviation of about 8 percent were up to 15 times more efficient.”
The end result is a hearing aid that could run about four or five times longer on one battery. But it doesn’t have to stop there. Rice plans to use this type of chip in a tablet destined for India’s schools and has a contract with the government to produce 50,000 of them over the next three years. Above is an example of how the errors distort but don’t render the image impossible to see. Given our Retina displays on the latest iPads and love of HD I’m not sure I’d want this on my tablet, but it would improve battery life.
The broad technique of sacrificing accuracy for better performance and lower power consumption shows promise in other areas as well. Startup Lyric Semiconductor is trying to use it to improve Flash density and power consumption, while even programmers are exploring ways to write code for massively multicore chips that sacrifice accuracy for performance. So perhaps after decades of striving for straight As, computing might want to relax and accept a few Bs.