Google has quietly acquired Provino Technologies, a start-up developing network-on-chip (NoC) systems for machine learning, an IEEE Spectrum investigation has discovered.
The latest processors for AI applications are home to thousands—or even hundreds of thousands—of cores, each of which needs to move vast swathes of data.
NoC technologies could accelerate communications on such “many-core” chips by replacing traditional buses and direct wires with an architecture familiar from large computer networks and the Internet, based on routers directing packets of data.
“Technology for communication simply hasn’t improved in the same way as that for computation,” says Md Farhadur Reza, assistant professor of Computer Science at the University of Central Missouri. “NoC’s decentralized architecture can have applications running multiple tasks in parallel and communicating with each other at the same time. And that means performance will improve, throughput will improve, and your wires will be shorter.”
Ex-Apple engineer Shailendra Desai founded Provino in 2015 to provide a platform called iFabric for developing NoC chips. The start-up was based in Silicon Valley with a second office in Ahmedabad, India.
In 2018, Provino raised $8 million in a Series A funding round led by Dell Technologies Capital, the investment arm of the computing multinational. At the time, the company identified chip development for “machine learning/artificial intelligence, consumer and automotive applications” as its focus.
“Provino’s technology has a fresh approach to system-on-chip design, addressing the challenging requirements of next generation chip design in the burgeoning artificial intelligence and machine learning markets,” wrote Daniel Docter, Managing Director of Dell Technologies Capital at the time.
Not only are NoC architectures faster and less prone to data bottlenecks than traditional chips, they are also inherently scalable, reconfigurable and fault tolerant. “There are multiple paths between the two nodes, so even if one link is down, you can still route a packet another way,” says Reza. “This makes it the most efficient architecture for neural networks.”
Neural networks are famously computationally intensive, particularly during training runs that rely on frequent communications between the “neurons” and memory.
Google started developing its own application-specific integrated circuits (ASICs) for neural networks in 2015. The hardware, called tensor processing units (or TPUs), are deployed within Google’s data centers to power AI products like Translate, Photos, Search, Assistant, and Gmail.
The precise specifications of Google’s TPUs are unknown, although some Google researchers have been studying NoC technologies for years.
In early February, Google bought 20 patents and patent applications for NoC communications and power control from Provino, for an undisclosed sum. This appears to be Provino’s entire portfolio of intellectual property.
While neither Desai nor Provino responded to requests for comment, Google later confirmed to Spectrum that it had bought Provino. It provided no further details on the acquisition, but the purchase could signal a move within Google to adopt NoC technologies.
A wholesale transition to NoC is unlikely to happen overnight, says Reza: “There are still a lot of challenges from the architectural and the algorithmic point of view. Routing is at the heart of NoC and there are many questions about how to design the routers, the algorithms they use, their buffers and the capacity of the links.”
Nevertheless, anything that promises to improve the efficiency—and reduce the staggering power usage—of machine learning systems, especially at the scale at which Google operates, can only be a good thing for the future of sustainable AI.