|
March/April 2008 TR10: Probabilistic ChipsKrishna Palem thinks a little uncertainty in chips could extend battery life in mobile devices--and maybe the duration of Moore's Law, too. By Erika Jonietz
Krishna Palem is a heretic. In the world of microchips, precision and perfection have always been imperative. Every step of the fabrication process involves testing and retesting and is aimed at ensuring that every chip calculates the exact answer every time. But Palem, a professor of computing at Rice University, believes that a little error can be a good thing. Palem has developed a way for chips to use significantly less power in exchange for a small loss of precision. His concept carries the daunting moniker "probabilistic complementary metal-oxide semiconductor technology"--PCMOS for short. Palem's premise is that for many applications--in particular those like audio or video processing, where the final result isn't a number--maximum precision is unnecessary. Instead, chips could be designed to produce the correct answer sometimes, but only come close the rest of the time. Because the errors would be small, so would their effects: in essence, Palem believes that in computing, close enough is often good enough. Every calculation done by a microchip depends on its transistors' registering either a 1 or a 0 as electrons flow through them in response to an applied voltage. But electrons move constantly, producing electrical "noise." In order to overcome noise and ensure that their transistors register the correct values, most chips run at a relatively high voltage. Palem's idea is to lower the operating voltage of parts of a chip--specifically, the logic circuits that calculate the least significant bits, such as the 3 in the number 21,693. The resulting decrease in signal-to-noise ratio means those circuits would occasionally arrive at the wrong answer, but engineers can calculate the probability of getting the right answer for any specific voltage. "Relaxing the probability of correctness even a little bit can produce significant savings in energy," Palem says. Within a few years, chips using such designs could boost battery life in mobile devices such as music players and cell phones. But in a decade or so, Palem's ideas could have a much larger impact. By then, silicon transistors will be so small that engineers won't be able to precisely control their behavior: the transistors will be inherently probabilistic. Palem's techniques could then become important to the continuation of Moore's Law, the exponential increase in transistor density--and thus in computing power--that has persisted for four decades. When Palem began working on the idea around 2002, skepticism about the principles behind PCMOS was "pretty universal," he says. That changed in 2006. He and his students simulated a PCMOS circuit that would be part of a chip for processing video, such as streaming video in a cell phone, and compared it with the performance of existing chips. They presented the work at a technical conference, and in a show of hands, much of the audience couldn't discern any difference in picture quality. |
Audio »
Share »
Digg this
Add to del.icio.us
Add to Reddit
Add to Facebook
Slashdot It!
Stumble It!
Add to Newsvine
Add to Connotea
Add to CiteUlike
Add to Furl
Googlize this
Add to Rojo
Add to MyWeb
Favorite
Print
E-mail
Moore's Law
12/22/2008
Massachusetts Institute of Technology
Comments
GaryB on 02/19/2008 at 4:28 PM
42
lenhamilton on 02/25/2008 at 12:33 PM
1
kitk on 02/27/2008 at 2:01 PM
55
Silacon on 02/27/2008 at 4:19 PM
41
carlii on 02/27/2008 at 10:43 PM
25
Rich43 on 12/05/2008 at 5:40 PM
2
I'm afraid the result of video processing most definitely is a number! It is in fact a stream of numbers.
I don't know if Palem actually said what was written down by the author of this article or whether the author translated what Palem actually said into something else in an attempt to simplify for the readership or whether the author didn't actually understand themselves; but what what Palem is trying to convey is this:
In some applications 100% accuracy of the output of the digital electronics is not critical. In areas of video and audio small errors in the data streams - so long as they are not perceptible by the human senses - is acceptable.
On the other hand, applications such as flight control systems in fighter aircraft, if the digital systems outputted small errors then that would not be tolerated and could be fatal.
So depending on the application, the presence of a small amount of error on the output can be tolerated.
Now, like other engineers/experts in electronics and computing, I am initially sceptical. Why?
What Palam is talking about is reducing the voltage of operation of the electronics to the point where it fails to function correctly.
How does this translate into 'small' errors in the output? Why is it that the errors are small, why should those errors not be large?
He has demonstrated video using the concept where the errors in the data stream were not discernable.
As an electronics engineer, I'm extrapolating what I believe Palam is talking about.
Logic gates are made up of transistors.
A chip for processing video will be made up of tens of thousands, hundreds of thousands of logic gates. Reducing the voltage will mean that the transistors don't switch correctly, thus leading to logic gates ouputting errant values.
Let's suppose we're working with monochrome video, with 8 bits per pixel, giving 256 shades of grey to simply the discussion.
Those logic gates work together to generate a stream of values, each value being in the range of 0-255. A single errant bit change in an output value can cause the output to change from its correct value by 0 - 128.
For the errant value to be imperceptible, what you would like is the difference from correct to errant value to be small, perhaps 1 or 2.
It's small and you might not notice it when watching the video stream.
How can that be realised? That's the issue I have.
Anyone of those logic gates in the digital circuit could produce an errant output value, leading to a completely different value being output in the 8 bit data stream, not just errors of 1 or 2, but 128 too. And here we're only talking about a single bit being in error.
If more than one bit error occurs, then the error value is greater still.
So my question is, how has Palam managed to constrain the magnitude of the error to ensure that it is not perceptible?
That's the key to this whole thing. (which hasn't been even mentioned in the article)
If reducing the operating voltage results in errors being perceptible in the output data stream then it's useless.