Nervana Wants to Make Special Hardware Just for Deep Learning
Deep learning, the current belle of the artificial-intelligence ball, has helped products like speech recognition and image search, from companies like Google, Microsoft and Facebook, make great leaps forward in recent years. It’s an older technique that was enlivened by massive quantities of data.
A new San Diego-based startup called Nervana Systems wants to make deep learning more accessible by developing custom hardware built to process all that data.
To design the hardware, it has collected a seed round of $600,000 in funding from a group of Silicon Valley investors who are primarily associated with Internet companies, not chips. They are Ali and Hadi Partovi, Sam Altman, Geoff Ralston, Scott Banister, Owen Van Natta, Eric Baker, Farzad Khosrowshahi, Dara Khosrowshahi, Allen & Company, Aditya Agarwal and Ruchi Sanghvi, and SV Angel.
Nervana CEO Naveen Rao said that should be enough money for the company’s small team — which previously worked at Qualcomm and only started Nervana in March — to create a demo version.
It’s too far out to nail things down, but the product should be priced in the tens of thousands of dollars, Rao said.
“We’re running out of juice already when it comes to silicon,” Rao said. “Silicon has hit a wall. The only way to do it better is novel architectures.”
Here’s how Rao described Nervana’s solution: Deep learning is generally conducted on graphic processing units (GPUs) rather than general-purpose central processing units (CPUs) because GPUs are designed to handle parallel processes. But since data structures live in the memory, they tend to be moved back and forth as they are processed. That generates a lot of overhead. Nervana’s solution moves things back and forth less, because the state of the algorithm resides on the chip, rather than in the memory.
The real opportunity of Nervana is that by making deep learning more accessible, it will hopefully help researchers and companies make more interesting breakthroughs in artificial intelligence, said Ali Partovi, who helped assemble the investor group. Partovi and his brother Hadi knew about Nervana early on because CTO Amir Khosrowshahi is their cousin.
As Ali Partovi put it, today’s deep learning breakthroughs have been in things like computers training themselves to recognize cats in videos and make fewer speech recognition errors — things a human 4-year-old could easily do.
What would be more impressive to laypeople might be breakthroughs in areas like processing weather data, satellite photos and surveillance video — things people have more trouble doing with their own brains.