Accelerator chips that use mild reasonably than electrons to hold out computations promise to supercharge AI mannequin coaching and inference. In concept, they may course of algorithms on the pace of sunshine — dramatically quicker than immediately’s speediest logic-gate circuits — however thus far, mild’s unpredictability has foiled most makes an attempt to emulate transistors optically.
Boston-based Lightelligence, although, claims it’s achieved a measure of success with its optical AI chip, which immediately debuts in prototype kind. It says that latency is improved as much as 10,000 instances in contrast with conventional , and it estimates energy consumption at “orders of magnitude” decrease.
The expertise underpinning it has its origins in a 2017 paper coauthored by CEO Yichen Shen. Shen — then a Ph.D. scholar learning photonic supplies at MIT below Marin Soljacic, a professor at MIT’s Division of Physics who runs the college’s photonics and fashionable electro-magnetics group — revealed analysis within the journal Nature Photonics describing a novel solution to carry out neural-network workloads utilizing optical interference.
Lightelligence was based months later, and Soljacic was one of many first to hitch its board of administrators.
“A scholar like Yichen comes by a couple of times in a professor’s profession, even at MIT. Yichen employed his background in nanophotonics and machine studying to invent this subject of utilizing optics for AI. He’s an actual pioneer,” mentioned Soljacic.
The chip in query — which is concerning the dimension of a printed circuit board — packs photonic circuits just like optical fibers that transmit indicators. It requires solely restricted power, as a result of mild produces much less warmth than the electrical energy, and is much less vulnerable to adjustments in ambient temperature, electromagnetic fields, and different noise. It’s designed to fit into present machines on the community edge, like on-premises servers, and can finally ship with a software program stack suitable with algorithms in generally used frameworks like Google’s Tensorflow, Fb’s Caffe2 and Pytorch, and others.
Lightelligence has thus far demonstrated MNIST, a benchmark machine studying mannequin that makes use of pc imaginative and prescient to acknowledge handwritten digits, on its accelerator. And it’s recorded matrix-vector multiplications and different linear operations — key elements of AI fashions — operating roughly 100 instances quicker than the state-of-the-art digital chips.
“We’re very happy to disclose our working optical chip AI computing system,” Shen mentioned. “Our prototype … is 100,000 instances quicker than the system demonstrated in our Nature Photonics paper and a fraction of the dimensions. The system is a real testomony to our workforce.”
Thus far, Lightelligence has raised $10.7 million in enterprise financing and has over 20 staff, together with a lot of trade veterans hailing from Columbia, Georgia Tech, Peking College, and UC Berkeley. Headlining the roster is Dr. Gilbert Hendry, who’s held varied roles at Google and Microsoft, and Maurice Steinman, a former AMD senior fellow.
Lightelligence stands largely alone within the optical AI accelerator area, but it surely competes with Lightmatter, which has raised double the quantity of funding ($22 million) for its personal chip. (Lightmatter’s CEO Nicholas Harris, curiously, was a coauthor on that Nature Photonics paper.)