Krishna Palem thinks a little uncertainty in chips could extend battery life in mobile devices--and maybe the duration of Moore's Law, too.
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.
Chips that let errors happen, then correct them, use less power overall.
The computer chip has evolved from a simple integrated circuit to a microprocessor with millions of transistors.
what strikes me in this is that visual as well as audio circuits could benefit. music is NOT all about complete precision, but rather the depth and richness has a lot to do with minute variations from the norm. this is why a really good radio (like my old tube-type 1934 Airline) acutally sounds better than the vaunted Bose radio next to it. in video, it could bring a more natural look to displays and virtual. nature is not strictly linear, but has a lot of seemingly random variations.
We applied this concept in a DARPA project details forever secret. CPAs use this principal to save tons of time in accounting without sacrificing overall accuracy at least for taxes. Good going Palem. It is about time the semiconductor world caught with accounting.
Besides a chart available online, a video sample and audio sample that we could play on our computers and/ or cell phones would be nice. This way we could see and hear for ourselves what the difference is. More details on how that relates to cryptography would be good too. Using probability seems to come closer to the uncertainty principles that emerge at the quantum scale too. Wonder if this will open the door to quantum computing from another angle. A third area of interest is whether probability based results might also result in designs that survive circuit imperfections better, resulting in higher net yield.
"Palem's premise is that for many applications--in particular those like audio or video processing, where the final result isn't a number"
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.
Your concern was addressed in the article: "Palem's idea is to lower the operating voltage of parts of a chip--specifically, the logic circuits that calculate the least significant bits".
Even if it weren't mentioned, it should be a trivial matter for any electronic engineer to "extrapolate" how Palem's concept could be successfully realised.
Also, a single-bit error in an 8-bit value will produce an error of anywhere between 1 and 128. A change of 0 is not an error.
presumably feedback filtering,etc is a no-go? because of the accumulation of errors?
Manufacturing in the United States is in trouble. That's bad news not just for the country's economy but for the future of innovation.
Our list of the 50 most innovative companies, including the following:
On Twitter
Become a Fan on Facebook
Subscribe to the Feed
GaryB
119 Comments
Energy savings
A simple chart showing what sort of energy savings could be achieved with what sort of loss of accuracy would have helped here and is crucial for understanding what the significance of his work is.
Reply
lenhamilton
1 Comment
Re: Energy savings
Informative article, but I tend to agree with GaryB that even a simple chart would have added to its impact. Still, it is easy to see that implementation of these concepts where appropriate would be of tremendous advantage.
Reply