SAN JOSE, Calif. — A veteran semiconductor executive launched a company to serve machine-learning startups that target embedded systems. He is already tracking more than 200 companies jumping into the fray that technology web giants such as Amazon, Google, and Facebook see as strategic.
“There’s a gold rush of ideas and market applications … it’s really a computing revolution — the emergence of neural networks as a broadly applicable and extremely powerful new method to solve a range of problems not well served by traditional algorithms,” said Chris Rowen, founder of Tensilica, who recently left a full-time position at Cadence Design Systems to start Cognite Ventures.
The AI field is so rich that Rowen has yet to determine whether he will mainly invest in startups, incubate a few, or launch one of his own in underserved or undiscovered areas he sees.
“My goal is to be smarter than the smart money,” said Rowen, who so far has just $10 million to spend but many years running companies focused on embedded systems, now generally called the Internet of Things.
Rowen sees some of the biggest opportunities in machine-learning tools. For example, he envisions tools to squeeze down the size of neural networks, generate code for them, and map them into fixed-point math routines or integrate them into a broad application framework. “Everyone has the same problems enabling rapid deployment,” he said.
The huge data sets and models used by web giants such as Google could be reduced by a factor of 500 to fit the needs of embedded systems, according to talks at a summit Rowen hosted at Cadence.
“There was lots of evidence of rapid progress adapting what was cloud-based technology by rethinking the algorithms and optimizing the networks,” he said, noting that the event drew nearly 200 people. “It was the biggest event ever held at the Cadence campus by a wide margin.”
Among AI market sectors, the autonomous vehicle segment is well served. Underserved areas include the human/machine interface, surveillance, and augmented personal devices, he believes. One area that Rowen does not see as a good target, ironically, is silicon.
“Embedded silicon for machine learning is probably more of a big-company game than a startup game because silicon generally is a big company game,” he said. “It’s highly likely that specialized inference engines will appear as accelerator subsystems like video codecs and audio DSPs in an SoC,” he added.
Cadence and Synopsys already offer such IP blocks, and other established chip companies have or are gearing up offerings, he noted. Silicon competition will play out in scaling performance and meeting specific market requirements
“Mainstream players are fully invested and producing some very good stuff; some companies will design their own inference engines, but they will have to compete with mainstream suppliers who have a pretty good handle on the issues,” Rowen said.
Next page: Learning to wield a big hammer
Rowen's taxonomy for embedded neural net apps. (Image: Cognite Ventures)