Probability calculations are at the heart of the software most of us rely on daily, whether it is or iTunes recommending new products based on previous purchases, a spam filter weeding out junk e-mail or a credit card processing program searching for potentially fraudulent transactions. Making these applications faster and more accurate has generally meant throwing more number-crunching capacity at them, but one Cambridge, Mass.–based start-up claims to have developed a cheaper and more energy-efficient approach that eschews digital processing.

Rather than performing serial calculations like a digital processor, Lyric Semiconductor's probability-processing technology includes a circuit design that the company says performs probability calculations in parallel, taking up less space and using less power.

Spam filters, search engines and even genome sequence analysis applications are simultaneously checking alternatives, computing their probabilities and choosing those options that have the greatest likelihood of being correct, says Lyric vice president of product development Dave Reynolds, who developed Lyric's technology with fellow co-founder Ben Vigoda at the Massachusetts Institute of Technology (M.I.T.). "Digital computers are good at processing 1s and 0s but not at processing probabilities," he adds.

Vigoda says, "Unlike digital gates that can only take 0s and 1s as input and output 0s and 1s, Lyric's gate circuits can take inputs that are between 0s and 1s such as 0.7 or 0.234. These in-between values represent probabilities."

Lyric's first offering, unveiled Tuesday at the Flash Memory Summit and Exhibition in Santa Clara, Calif., is an advanced error-correction technology called Lyric Error Correction (LEC) for flash memory that performs with a few probability gate circuit transistors what it takes hundreds of binary digital transistors to accomplish, Reynolds says.

Slippery Slope
The demand for flash memory—used for fast, easy storage and access to information stored in cell phones, computers, digital cameras and other devices—has grown rapidly. To meet this demand for ubiquitous memory, electronics makers have had to pack in more and more data, even as they shrink the sizes of their gadgets. The downside is that one in every thousand bits stored in a flash memory comes out wrong when the memory is read, Reynolds says. As the ability to increase the volume of information improves, the number of errors will approach one wrong bit in every 100 bits, he adds.

"We've been heading down a road of using digital logic more and more. If things become complicated, we just add transistors," says Jim Handy, an analyst with the semiconductor market research firm Objective Analysis in Los Gatos, Calif. "As flash storage has progressed, it's common to have more errors, so we have needed more algorithms with fancier math and a greater number of transistors to perform the corrections."

Flash memory drives today use a microcontroller for digital error correction, but these become more complex and more expensive as flash density grows. Reynolds claims that the LEC will be 30 times smaller and consume 12 times less energy than the current generation of flash memory error correctors. Adds Vigoda, "The LEC product proves that we can build products using these probability gate circuits and bring significant value versus digital technology."

Lyric's goal is to license LEC to flash memory–makers so that they could build this new approach to error correction into their products. "We are in discussions with multiple flash vendors, but we do not have anything we can announce at this time," Reynolds says.

Probability Platform
Now that Lyric has launched its LEC the company plans to expand its probability offering by developing the general-purpose programmable probability-processing platform (GP5) by 2013. Instead of being a particular specialized fixed-function chip like the LEC, the GP5 could be used presumably by any application that deals heavily with probabilities. When computing probabilities it will do the work of a 1,000 conventional processors, such as those from Intel or AMD, Vigoda says.

Lyric, which has received more than $18 million in funding from the Defense Advanced Research Projects Agency (DARPA) to perform basic research into probability-processing technology, is also working on a programming language called Probability Synthesis to Bayesian Logic. PSBL would essentially enable a computer to offload probability-related problems to a GP5 while assigning other types of processors—central and graphics processing units, for example—to do what they do best, whether it is serving Web pages or rendering graphics. "We're not trying to solve all problems, just probability-based problems," Reynolds says.

Lyric's technology has a lot of potential, although it is difficult to know at this point whether their approach will succeed, Handy says, adding, "Lyric's approach is to go back to the starting point and not use conventional digital logic."

If successful, Lyric's probability-processing technology could greatly improve data compression and encryption (in addition to error correction), while also driving down their cost. All three functions use almost the same math, Handy says. Whereas advances in error correction have big implications for data storage, a new approach to data compression could lead to, for example, more efficient use of the radio waves and spectrum bandwidth (a boon to wireless communications). Meanwhile, enhanced encryption could make data security ubiquitous because, Handy adds, "Every connection becomes a secure connection."