“Inexact” Chips Are 1500% More Efficient than Today’s Processors

Researches from Rice University in Houston, Singapore’s Nanyang Technological University (NTU), Switzerland’s Center for Electronics and Microtechnology (CSEM) and the University of California, Berkeley have recently made public their achievements in “inexact” computing.

What the researchers did was practically give up on perfect accuracy and see what happens when you design a chip that will give you “good-enough” results. Many of you might wonder what good could that do and how could “inexact” results be acceptable. Well, a simple answer might be a comparison between a pin-point accuracy of a “flying” bullet and the “acceptable” target miss by a missile. When shooting a “flying” bullet or any other small and accurate projectile, the objective is to accurately hit the desired target. 

When firing an intercontinental nuclear missile, a target miss of 10 meters might actually be acceptable while being completely unacceptable for a pin-point small projectile. The same can be said about facial recognition. All state security and intelligence agencies are running a lot of facial recognition software every second of every day. Facial recognition is an approximate matter and the likely results are finally displayed and agents and experts decide if the pictures are indeed representing the same person. Around 84% of Facebook’s initial funding came from In-Q-Tel, an investment fund for the CIA. So there is clear interest from state agencies into such technologies.

The way these chips achieve such high efficiency is “pruning.” This is basically a way of giving up some functional units of a CPU that are used very rarely and mostly for verification purposes. “Pruning” also gives up most of the performance gains for energy efficiency. If the researchers see that one special way of achieving a result yields an increased performance, they will confine the voltage of the chip to a lower level while keeping the previous performance level. So if there is one optimization that allows the chip to computer more or faster, the concept chooses a smaller voltage target that would allow the same result rather than enjoying the higher performance at the current voltage.

“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. 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,” said study co-author Avinash Lingamneni, a Rice graduate student. As you can see form the picture below, the image on the left is perfectly rendered while the next two are produced with an allowed relative error of 0.54 and 7.58 percent respectively. Obviously, the second and third imaged are clearly acceptable and discernable. This is quite good news for those into image recognition. It’s a good thing they bring higher efficiency, but the likely applications don’t sound too friendly.

Posted in: Hardware, News, Preview