Blog Post

Startup Makes Probability-Based Chips for Big Data Apps

Stay on Top of Enterprise Technology Trends

Get updates impacting your industry from our GigaOm Research Community
Join the Community!

Lyric Semiconductor, a Cambridge, Mass.-based startup that was spun out of MIT in 2006, today announced a revolutionary new approach to processing data called probability processing. It wants to make chips that are fit for a world drowning in information and dealing with data deluge.

The company was started by Dr. Ben Vigoda, a chip industry veteran, and MIT Scientist and chip industry veteran David Reynolds. It’s backed by Analog Devices CEO and lead partner of Stata Ventures, Ray Stata, with more than $20 million in funding from DARPA. It’s coming out of stealth today.

The basic premise behind Lyric is quite simple: In the “Big Data” future, companies will need a new approach to processing data. They call this technology probability processing, and it’s been under development for nearly a decade.

For over 60 years, computers have been based on digital computing principles. Data is represented as bits (1s and 0s). Boolean logic gates perform operations on these bits. Lyric has invented a new kind of logic gate circuit that uses transistors as dimmer switches instead of as on/off switches.  These circuits can accept inputs and calculate outputs that are between 0 and 1, directly representing probabilities – levels of certainty. A digital processor steps through these operations serially in order to perform a function. In order to improve efficiency even further, Lyric’s processors are designed to perform many probability computations in parallel. (Lyric Press Materials)

What that means is that these chips would be ideal for big data applications that solve modern day computing problems like search, fraud detection, spam filtering, genome sequence analysis and financial modeling. It is not that general-purpose chips can’t do these tasks well; it’s that they’re not as efficient, argues Lyric’s Vigoda. Big data applications aren’t the only ones where these chips can come in handy. As the networks of tomorrow get beefier, more complex and carry larger amount of information, Lyric’s chips could make routing and switching more efficient.

The company is commercially launching the technology as an application targeting Flash memory chips. It’s smartly timed, especially since the size of flash memory chips keep increasing. According to some estimates, one in 1000 bits stored in Flash memory come out wrong. As the density of Flash memory chips increase, error rates increase, and correcting those errors has become an expensive proposition. To solve this problem, the company has made an error correction processor called the Lyric Error Correction (LEC) for Flash Memory. It’s 30 times smaller and consumes 12 times less power.

What’s next? The company is working on a new programmable processor called GP5.

The GP5 could enable performance gains of 1,000X over today’s digital x86-based systems such as the processors from Intel and AMD.  The GP5 will run code written in Lyric’s own probability programming language called PSBL (Probability Synthesis to Bayesian Logic), an expressive computer programming language for working with probability based computations.  Lyric will leverage its probability processor and programming technologies to deliver disruptive total systems to its customers.

The big risk for the company is in getting adoption for its technology. It needs to find buyers — which could range from networking gear maker to specialist server makers — to buy into its approach to making chips. Lyric will also face the additional challenge of getting folks to learn a new way to program these chips.  That said, no one can doubt that Lyric’s approach is ambitious and audacious.

Related Research from GigaOM Pro (subscription required): Pushing Processors Past Moore’s Law

7 Responses to “Startup Makes Probability-Based Chips for Big Data Apps”

  1. I thought that the transistors went out like the dodo bird when Computer chips were invented.
    It is so complicated that any one using this new technology will probably require a PHD. in Computer Tech. Programming and Analysis.
    Its not for the average Joe.

  2. While I agree that the age of big data requires probabilistic approaches, I fear Lyric may have a tough time convincing a business audience of this.

    Few questions that businesses ask of their massive data — such as, “were sales significantly higher yesterday?” — rarely require operations over billions of events to produce exact counts.

    Yet few executives like answers that begin with “probably…” — they want precision, even false precision.

    That said, I hope these guys may succeed in proving me wrong, and save the world a lot of CPU cycles in the process.

    • Interesting development: fuzzy chips! The tough part is not only getting new equipment purchases of the hardware, but convincing human capital (brains) to invest time into PSBL. Besides, its pretty clear that proprietary languages have problems (especially ones like PSBL that have trademarks next to their names … Oracle Java patents wars anyone?). They’ll have to get a few marquee customers which could be expensive (maybe Bernanke and the Federal Reserve will be good candidates to see if probability programming can avert the next financial crisis lol). @Michael would probably agree with me that this announcement is unlikely to stop Hadoop which is just getting warmed up!

      • Eddie and Michael,

        Great points. I didn’t go into details here, but getting folks to sign on for using their programming protocol isn’t going to be easy.


        I would disagree with you on one topic: in the future all businesses have to be asking complex questions from their IT systems. Without that they are just not a relevant enterprise. So perhaps that is why more folks are going to need these (or these kind of) chips ;-)