Lyric Semiconductor has developed a type of chip that won't give you a simple yes-or-no answer -- and that's the beauty of it.
And if the idea one day takes flight, the MIT spinout plans to accomplish things like drastically improving online recommendation engines, mapping an individual's genome cheaply in a matter of minutes, or more accurately detecting fraud and spam on the internet. In the process, power consumption would be slashed back at data centers.
"Yahoo uses a lot of servers for spam," said CEO Ben Vigoda. "We combine physics and computing in a new way. It is not digital. We use electrons and transistors in a different way."
The company's GP5 architecture is essentially based around the principles of probability, according to Vigoda. Digital chips, like the Intel or AMD processor in your laptop, operate by breaking down everything into series of 1s and 0s. Transistors either let electrons pass or they don't. A photo from your digital camera or a song from your MP3 player are really just collections of on/off signals that form a coherent whole because you don't detect the individual signals.
Analog devices, by contrast, mimic natural phenomena: photoresistant chemicals replicate reality on film and speakers vibrate.
Chips based around the GP5 architecture combine the probability of one event (a 50-percent chance of A visiting the Empire State Building) with another (a 40-percent chance of B appearing there) to determine the odds of them meeting (0.56 x 0.4 = 20 percent). The answer isn't definitive in a theoretical sense, but it's right.
What does that mean in practice? To solve a sudoku problem or fold a protein, a digital computer will look at a vast number of all possible combinations. A probability processor would look at the basic rules (e.g., you can't have the same number twice in the same line or these two amino acids can't bond) and build from there. In the end, that leads to less power consumption.
Others are mining the field of green IT through better chips, as well. Yesterday, Smooth-Stone came out of stealth with a processor that directly challenges Intel in the market for low-power server chips. Sanforce, meanwhile, has a controller that makes it more practical to adopt flash -- which can be problematic over time because it stores data in a "violent" manner -- for storage.
Although probability sounds like a somewhat loosey-goosey premise from which to derive answers, computing has already moved in that direction. Google's search engines are based around probability, also known as probabilistic computing. The field of artificial intelligence has largely shifted from rule-based computational systems for robots to probability.
The foundation of the field can be traced back to Thomas Bayes, an 18th-century clergyman who tried to prove the existence of God through equations. (Disclosure: Vigoda told me he became intrigued by probability after reading this 2003 article that, coincidentally, I wrote. What are the odds of that?)
Lyric's first chip will be an error correction controller for flash memory chips used to eliminate artifacts from data. Vigoda will publicly unveil Lyric's technology for the first time at the Flash Memory Summit taking place in Santa Clara this week.
The company's controller is 30 times smaller than conventional error correction components, consumes 12 times less power and provides more bandwidth. In some applications, the a GP5 chip could provide up to 1000 times better performance than a conventional processors.
Ultimately, Lyric would like to make programmable chips that could serve as co-processors inside servers or storage systems, according to Mira Wilczek. The company, in fact, is working on projects for DARPA, which is also one of the company's investors.
The history of the chip industry is filled with noble ventures for improving performance that flash and burn. The escalating performance and ever-declining prices of conventional digital chips make them tough to beat.
Lyric's strong points? Error correction is a looming hurdle for the data storage industry. Moreover, nearly every major flash vendor has a probability project underway. Thus, the need is clear and the foreign-ness of the concept is eroding. The company can make its chips using conventional silicon processes and works with industry stalwart TSMC.
It will also try to license its technology to large vendors (Samsung and Toshiba are two large flash vendors) in a way that will let them differentiate their products. ARM did something similar and now dominates the market for cell phones and smart grid.
That should increase the company's chances of success.