A few errors could be key to super-efficient computer chips
Summary: A group of university researchers has determined that "inexact" processors could yield huge gains in efficiency.
"Everyone makes mistakes" might be a nice "words-to-live-by" phrase, but it hardly seems like the path to greater performance from computer chips. Yet a group of university researchers has determined that processors that can tolerate a couple of errors could yield huge gains in efficiency.
The so-called inexact computer chip would be a whopping 15 times more efficient than current CPUs that include error-correcting mechanisms, managing to be faster and less power-hungry. The concept is based around the idea that certain applications can tolerate some mistakes in processing -- image-processing software, for example.
The researchers "pruned" certain circuits out of their chip design to help provide the efficiency boost. Of course, the more errors tolerated, the greater the performance gains, which could limit the real-world applications of the approach.
The inexact chip is initially being promoted for embedded applications, such as for hearing aids. It's also being used for a low-cost tablet in India where classrooms often lack electricity. It will be interesting to see if there's ever a way to harness this approach to mainstream computer chips, where significant performance gains are getting a bit harder to come by. Would there be a way to prune circuits and still deliver the computing experience expected by the typical user? After all, a few errors in the millions of pixels of a digital image is one thing, but you don't want them showing up in your Word document.
[Via The Verge]
Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.

Talkback
Stupid.
Images already have errors generated by every lossy compression algorithm.
Unfortunately, these same errors prevent it from being reliable in court.
Also doesn't work well for facial recognition, or anything needing a positive identification.
Doesn't hurt facial recognition
The court thing is right
But, regarding jessepollard's other statements:
"The place you don't want errors is in the instruction stream."
That's not where the errors are. They're in the calculation results from the non-errored instructions. And then, only in specific types of calculations.
"Images already have errors generated by every lossy compression algorithm."
True, but also irrelevant.
"Unfortunately, these same errors prevent it from being reliable in court."
True.
"Also doesn't work well for facial recognition"
False. Facial recognition is already a fuzzy logic affair. Actually, facial recognition is a perfect example of inexactness in computing and would be a great application for these chips.
"or anything needing a positive identification."
Definitely not true. When using biometric identification, "positive" is a graduating scale, whether it's voice identification, facial recognition, fingerprint recognition, or gate (walking style) recognition. The inexactness of images and motion from biological creatures makes 100% identification nearly impossible even with chips that produce no errors. Inexact calculations are the core of that type of technology and are perfect applications for these inexact chip technologies. Others would include 3D rendering, stock market prediction algorithms, AI, and many many more.
Of course, there are plenty of areas in computing where errors are completely unacceptable, but in the areas where errors are not critical, these chips could provide great benefits.
The Pentium FPU Flaw
Yeah, that's the ticket!
racingmustang
We are Pentium
RE: We are Pentium
You may assimilate as many units of the (alcoholic) beverage of choice, as you see fit.
Along with GPU's
Realistically, processors for routine work don't need to be significantly more efficient. So, this would be mainly for high-performance computing. If these "slightly error-prone" processors are that much more efficient, they could be combined with separate processors to handle the math that needs accuracy. The overall combination would probably provide substantial improvement yet with acceptable accuracy.
Uh, yeah...
ERROR ERROR ERROR ! ! !
.
NOPE, ain't gonna work.
.
This is very serious stuff...
It seems like saying a watch that is broken...
What are our kids learning these days, anyway??????? How to "cheat"?
But...
Head Shot!!
I wonder how aware those Indian children are of the inexact computers
ranban sunglasses
3+3=7
This is known already
Errors happen in computers anyway. What is necessary is an error-detection and remedy technology, as see in the biological computers.
Interesting!
Maybe it's just a helper chip alongside the main one(s) which gets sent various processing demands that are suited for it, like the graphics or audio manipulation or physics in gameplay mentioned here. For mobile phones maybe it could help with the A-D and D-A conversions? Losing a tiny bit of clarity for voice or video chat probably won't be noticed, and definitely wouldn't be on the scale of network dropouts. For many times the battery life, I'd take that tradeoff any day.
Clever idea!