• Post category:国际新闻

New Paradigm in AI computing is here.

Startup Lightelligence, a builder of optical AI accelerators, is pleased to announce the unveiling of its optical AI accelerator prototype, the first of its kind in the world. Lightelligence’s optical AI chip, developed by a team of MIT scientists and industry veterans, aims to leapfrog the existing digital AI accelerator/high-performance computing industry and be a game-changer for AI processing.

Lightelligence’s prototype device uses light instead of electrons to carry out core mathematical computations. Processing information with light could one day enable such devices to run AI algorithms ten to a hundred times faster than today’s best AI chips.

“Our initial project at MIT was a very simple proof of concept” shared Yichen Shen, CEO of Lightelligence. “The MIT computing system had the form factor and the computational power comparable to the 1943 ENIAC which was the first general purpose electronic computer. The Lightelligence demo we’re revealing has the computational power and form factor comparable to a personal computer circa 2000. We’re currently developing our next system that aims to match or even exceed the performance of the best digital electronic AI accelerator. From there, we will leverage our learnings to build AI accelerators that leave digital electronics behind.”

The Lightelligence optical chip demo was used to run MNIST, a benchmark machine learning model that uses computer vision to recognize handwritten digits. The demo showed that Matrix-Vector multiplication, which is a key component of machine learning, can be completed by the Lightelligence optical chip 100 times faster than the state-of-the-art electronic chips.

Michael Hochberg, CTO of Elenion Technologies, shared this about the work: “Doing computation with light has, until now, been wildly impractical, because researchers used conventional optics, usually in an attempt to create an optical analog to a transistor. This was fighting against the physics of light, because light doesn’t want to act like the electrons in a transistor, and this work never ended up with commercial products. What Lightelligence is doing is different: They’re working with the physics of light, instead of against it, and building integrated systems-on-chip to scale up the speed and capability of their optical processors. Silicon photonics is now emerging as the dominant technology for coherent optical links between datacenters and cities, and for connecting computers together within the datacenter. Lightelligence is leveraging this industry trend to get their devices to market quickly and Elenion, as a leader in silicon photonics, is pleased to support their ongoing efforts. It’s really exciting to see this optical computation work becoming real so rapidly.”

Yichen Shen co-founded Lightelligence after studying nanophotonics and machine learning as a Ph.D. student in the lab of Marin Soljacic at MIT. Together with Soljacic and several others, Yichen published a paper in the journal Nature Photonics in 2017 describing a newly discovered way to perform neural-network computations using optical interference.

“A student like Yichen only comes through rarely in a Professor’s career, even at MIT. Yichen is a real visionary and a pioneer in this field of using integrated optics for AI,” said Lightelligence co-founder and MIT Professor, Marin Soljacic.

Lightelligence plans to deploy its technology for AI computing, a rapidly growing market. McKinsey & Company projects the AI chipset market to grow to $65 billion by 2025. The Lightelligence software team is building the full software stack to allow for AI algorithms in commonly used frameworks to be easily run on the optical accelerator. A few use-cases for Lightelligence’s future optical AI chips include datacenters, security cameras, and autonomous vehicles.

About Lightelligence: Lightelligence is a startup company that has spun out of MIT and is based in Boston. The company has raised $10.7M in venture financing and has over 20 employees. Lightelligence is the first company to demonstrate a stand-alone, fully-integrated optical accelerator system to run an AI algorithm.