PCMOS Microchip Puts Errors To Good Use

Didn't you hate when your college calculus processor insisted that an answer had to be "exact," and that "close enough" just wouldn't cut it? Clearly, you should have attended Rice University or Nanyang Technological University in Singapore. A team of researchers from both institutions have come together in order to create a revolutionary type of computing that thrives on random errors. You read correctly: this stuff actually benefits from impreciseness.

Dubbed PCMOS (for probabilistic CMOS), this technology has been used in a microchip that requires 30 times less electricity while running seven times faster than today's best technology. The crew unveiled their findings at the International Solid-State Circuits Conference in San Francisco, and while it may be tough to fathom, it's the "probabilistic logic" that enables it to operate with drastically less power. Better still, PCMOS piggybacks on the "complementary metal-oxide semiconductor" technology, or CMOS, that chipmakers already use, which means that said chipmakers won't have to buy new equipment to support PCMOS. As you can imagine, that's about where the similarities end, as PCMOS ditches the traditional usage of Boolean logic and replaces it with a radical probabilistic approach that was conceived by Rice University Professor Krishna Palem (pictured) and his doctoral student, Lakshmi Chakrapani.

According to Shekhar Borkar, an Intel Fellow and director of Intel's Microprocessor Technology Lab: "A significant achievement here is the validation of Rice's probabilistic analogue to Boolean logic using PCMOS. Coupled with the significant energy and speed advantages that PCMOS offers, this logic will prove extremely important because basic physics dictates that future transistor-based logic will need probabilistic methods."

You see, existing silicon transistors become increasingly noisy as they decrease in size, forcing engineers to deal by boosting the operating voltage to overpower the noise and ensure accurate calculations. Clearly, the more the voltage is boosted, the more power will be wasted in operation. Palem has stated that "PCMOS is fundamentally different," in that it "lowers the voltage dramatically and deals with the resulting computational errors by embracing the errors and uncertainties through probabilistic logic."

As for the future? The Rice-NTU team is hoping to follow-up its proof-of-concept work on encryption with proof-of-concept tests on microchips for cell phones, graphics cards and medical implants. For example, in a streaming video application on a mobile phone, the small display -- combined with the human brain's ability to process less-than-perfect pictures -- "results in a case where the picture looks just as good with a calculation that's only approximately correct."

Palem, who directs Rice's Value of Information-based Sustainable Embedded Nanocomputing Center, sums up the purpose nicely with this: "Our goal is green computing. We're looking for applications where PCMOS can deliver as well as or better than existing technology but with a fraction of the energy." So, who said imperfections weren't beneficial in computing?


Comments
3vi1 5 years ago

The chip doesn't put errors to good use, nor thrive on them - it just makes more of them than a normal chip. It's a bucket full of holes that they're trying to sell you as "50% lighter!".

It's like that coworker of yours that spends 7x less time and 30x less brainpower doing work because he gives educated guesses as answers instead of analyzing the data in detail.

It might seem like this could speed up video and other non income-tax-related stuff, but what happened back when nVidia was using less precise algorithms in products? People bought ATI because the picture "looked better".

Devices are still going to need boolean logic for most functions, so this is single-purpose overhead. I cannot believe this would give you anywhere near 30x times the battery life in an actual device - as a lot of that power's going towards screen illumination, sound output, etc.

Using JPEGs can speed up reading graphics from your hard drive by 100x by using a level of acceptable loss (error) to make the files that much smaller. Is your computer 100x faster for general purposes if you only use JPEGs for your graphics?

bob_on_the_cob 5 years ago

[quote user="3vi1"]It's like that coworker of yours that spends 7x less time and 30x less brainpower doing work because he gives educated guesses as answers instead of analyzing the data in detail.[/quote]

Sums the article up in one sentence. You amaze me 3vi1.

Post a Comment
or Register to comment