On the day after Intel announced its acquisition of San Diego machine learning startup Nervana Systems, investor Steve Jurvetson told me he was feeling a sense of satisfaction about a call he made three years ago, and how it has been playing out.
In a 2013 panel discussion at Silicon Valley’s Churchill Club, the DFJ partner said “machine learning” was his pick as the most important tech trend to watch for the next three to five years. “Just about anything you’ve heard [about] at Google that sounds interesting and new is based on machine learning,” Jurvetson said at the time. “Everywhere, technology is starting to percolate into an otherwise prosaic, non-tech industry—apply big data, apply machine learning—and revolutionize it.”
Just over a year later, Jurvetson led the Series A round of venture funding for Nervana Systems. Now, more than two years after Nervana was founded, Intel (NASDAQ: [[ticker:INTC]]) has endorsed Jurvetson’s top tech trend of 2013 by acquiring Nervana in a deal reportedly valued at more than $408 million (according to Recode).
“It’s not that often in a VC career you get to work with teams you’re sure will be featured in documentaries about how major industries changed,” wrote Matt Ocko of Data Collective Venture Capital, which joined DFJ in Nervana’s initial venture funding. “We can’t wait to see their next set of breakthroughs, delivered at even greater speed, with Intel behind them.”
(In the 2-1/2 years since it was founded, Nervana raised roughly $25 million. In addition to DFJ and Data Collective, other investors include Playground Global, CME Ventures, Lux Capital, Allen & Co, AME Cloud Ventures, Fuel Capital, Omidyar Technology Ventures, SV Angel, and unnamed seed investors.)
Nervana Systems co-founders (and Qualcomm expats) Naveen Rao, Amir Khosrowshahi, and Arjun Bansal set out to develop new artificial intelligence technology that emulates the human brain, where thousands of synapses transmit information between each neuron. The co-founders integrated their specialized expertise in neuroscience, distributed computing, and networking . As CEO Rao wrote in a blog post Monday, “Nervana started with the idea that we can engineer better solutions for computation by bringing together computer engineering, neuroscience, and machine learning.”
Rao added, “With this acquisition, Intel is formally committing to pushing the forefront of AI technologies.”
As Jurvetson drove between meetings yesterday afternoon, he said machine learning continues to be a top tech trend. In a self-described “sweeping generalization,” he declared that the new generation of iterative algorithms in computational mathematics that are being applied in machine learning, directed evolution, and generative design “is the most important advance in engineering since the scientific method.”
Nervana initially developed a software-based model of its technology, based on Nvidia graphics processors. Earlier this year, Nervana began offering its AI technology in the cloud to business customers, using a software-as-a-service model. For example, Nervana’s Web-based software has been used to help Paradigm software analyze prospective oil and gas fields, identify irregular trading patterns for the Chicago Mercantile Exchange, and enable Blue River Technology to refine its computer vision technology for use in agriculture.
At the same time, Nervana has been advancing its machine learning technology on a different front by developing a new type of semiconductor architecture based on the company’s neural-based approach to AI.
A Nervana “AI semiconductor” would appear to be ideally suited for Intel, as it enables the chipmaker to integrate machine learning into the silicon of Intel’s proprietary chip design—instead of running in the software atop graphics processors made by Nvidia and other chipmakers.
In a blog post that recaps the Nervana story, Jurvetson writes that one of the pre-eminent engineers hired by Nervana figured out how to rework the undocumented firmware of Nvidia’s processors to run deep learning algorithms faster than off-the-shelf GPUs “or anything else Facebook could find.”
Nervana says the souped-up capabilities of its Neon software framework enable it to run 10 times faster on Nvidia processors—faster even than Nvidia’s own software framework. While speed is always important in computer processing, it becomes especially important in machine learning, which rely on iterative algorithms that allow computers to learn patterns and data without explicit programming instructions.
Jurvetson writes that similar innovations in Nervana’s chip design will enable the company’s forthcoming chip to perform 55 trillion operations per second, with multiple high-speed interconnects typically seen in the networking industry making it possible to tie “a matrix of chips together into unprecedented compute fabrics.”
Nervana, which now has 48 employees, plans to remain in San Diego and operate under its own brand name. Combining Nervana’s expertise in artificial intelligence with Intel’s capabilities, technology resources, and huge market reach, Rao wrote, “will allow us to realize our vision and create something truly special.”
