A Chip That Digests Data and Calculates the Odds
By ASHLEE VANCE
Published: August 17, 2010
Complex as they may seem, traditional computers deal in a simple art. They rely on tiny switches that turn on and off, producing the streams of ones and zeros that software eventually translates into something meaningful to a human.
Times Topic: Computer Chips
Some computer scientists find solace in the degree of certainty that comes from trading in yes-or-no operations.
Lyric Semiconductor, a start-up that emerged from work at the Massachusetts Institute of Technology, looks to forgo this certainty in favor of probability. It unveiled plans this week to build a chip that can compute likelihoods. Such technology may help figure out which book someone will want to buy on Amazon.com or help create a better gene-sequencing machine.
“We decided there are lots of probability problems out there that are so important they deserved their own hardware,” said Ben Vigoda, the co-founder and chief executive at Lyric.
Most of Lyric’s nearly $20 million in financing has come from the Department of Defense and what the company will only refer to as a three-letter government agency. The military interest revolves around Lyric’s approach to determining the relationship between bits of information in a stream of communications and separating noise from useful data.
Over time, Lyric plans to step out from this military work and offer its wares to corporations working with large sets of data.
Today, retail Web sites throw lots of computing horsepower and algorithms at the prediction engines that try to determine which product someone might want to buy based on their past purchases and ratings. This is a grand probability puzzle, and it taxes traditional computers because they are made to deal with black-and-white questions.
Lyric, by contrast, can tell a computer that someone buys a mystery book 60 percent of the time and a science fiction book 30 percent of the time, and then hunt for the probability that the reader will like a new title that touches on these interests.
Similarly, the technology could be used to determine the best search results for an individual, or the likelihood that an e-mail message is spam. It could determine if a recent credit card purchase is fraudulent by comparing it with past purchases.
Crucially, Lyric claims it can perform these calculations using just a handful of the transistors inside a chip rather than the hundreds it takes today because it has created algorithms and chip designs geared toward probability. This means that companies could spend far less on computing gear for these complex tasks and save energy and space.
Lyric’s approach is similar to that of other chip companies that have filled niches in the past. Nvidia, for example, boomed after it created graphics chips that made things like video games and nuclear weapons simulations run faster.
Jon Stokes, the author of “Inside the Machine: An Illustrated Introduction to Microprocessors and Computer Architecture,” said Lyric faced “a really tough road ahead.”
“To take it to the next level requires an investment that reaches into the hundreds of millions of dollars, and it’s very risky and very hard to raise money for that sort of play right now,” Mr. Stokes said.
This week, the company unveiled one take on the technology that it hoped to license to makers of the memory that went into gadgets and computers. Lyric produced a small chip that could help catch errors that arose as those devices read and wrote data.