Sign In

Communications of the ACM

ACM News

Tripping the Light Fantastic

neuromorphic chip, illustration

Lightmatter's neuromorphic chip processes vast quantities of data simultaneously, similar to the way a human brain does.

Credit: Lightmatter

A team of researchers at the Massachusetts Institute of Technology (MIT) has developed a super-fast, prototype chip that uses light rather than electrons to crunch data.

The optical chip, which could one day process information as much as a million times faster than the chip in a typical Intel Core i7-based personal computer, won the grand prize in the MIT $100K Entrepreneurship Competition earlier this year.

Essentially, the researchers have found a way to manipulate light inside a chip so it processes vast quantities of data simultaneously, similarly to the way a human brain does.

"Today's computers do operations one by one," says Yichen Shen, a post-doctoral associate at MIT and co-founder of Lightmatter, a start-up the MIT research team has forged to bring its prototype chip to market. "We do operations all together, when light is passing through. It's all simultaneous. We change the architecture completely."

Such vastly improved computational power (when perfected) will enable manufacturers to bring extremely potent artificial intelligence (AI) to every desktop computer and many other computational devices, according to Shen.

"Deep learning algorithms that are traditionally computationally-hungry, or expensive, can now be efficiently deployed with our chips," says Darius Bunandar, another Lightmatter co-founder.

Personal assistants like Amazon Alexa, for example, would be able to engage in conversation much more quickly and efficiently when powered by an optics-driven chip, because they would no longer be forced to reach out to the cloud for the AI they need to process those responses. Instead, all of the AI processing work could be done locally within the Alexa unit, Shen says.

Moreover, an everyday smartphone, which also needs to reach out to the cloud for help with AI processing, would also be able to do AI computations on-board.

Other devices that could be turbocharged with intrinsically faster, on-board AI include many of the billions of Internet of Things devices connected to the Internet, devices that are expected to exceed 20 billion in number by 2020, according to a 2016 Gartner report.

"Lightmatter's approach can be seen as an important breakthrough in advancing AI," says Christian Weedbrook, founder and CEO of Xanadu Quantum Technologies, a quantum computing firm based in Toronto.  "Lightmatter's ability to harness photonics allows it to compute at the speed of light — which is much faster than what electrons can compute at."

Philip Walther, vice-dean of the Faculty of Physics, University of Vienna, agrees Lightmatter has "done an amazing job using phase shifters to get control of light. Because you're using light to transmit data, you're looking at very low power consumption and very little cooling needed, if at all."

Unlike conventional chips that rely on arithmetic, bit-by-bit processing, Lightmatter's chip processes data by using an architecture very similar to a biologic neural network, commonly referred to as neuromorphic computing.

"Traditional computer architectures are not very efficient when it comes to the kinds of calculations needed for neural network tasks," says Shen. "Such tasks typically involve repeated multiplications of matrices, which can be very computationally intensive in conventional CPU or GPU chips."

Nicholas Harris, another Lightmatter co-founder, adds, "After years of research, we came up with a way of performing these operations optically instead."

Moreover, while some researchers working in neuromorphic computing create an artificial neural network to mimic human brain processing, Lightmatter's network is physically based. "Rather than simulate the neural network, we physically implement it by controlling the strength of the connections between nanoscale optical waveguides," Harris says.

At that nanoscale, the optical chip "uses multiple light beams directed in such a way that their waves interact with each other, producing interference patterns that can be interpreted to convey the result of the intended operation," Shen says.

"Our current chip is just an experimental prototype; there is still a lot more work to do," Shen adds.   "When our chip is scaled-up in the future, it will be able to carry out, theoretically, about 10,000 TFLOPs per second, so we will be a million times faster than an i7 CPU when doing deep learning computing. In the long term, we hope it can become the next hardware platform for AI computing."

Similar work with light-driven chips is being done by a team led by Alexander Tait at Princeton University, which also created a neuromorphic, light-driven chip prototype that computes at ultrafast speeds.

Lightmatter is one of more than 160 start-ups that have emerged from the annual MIT $100K Entrepreneurship Competition, which has been going strong since 1990.

Collectively, those start-ups have gone on to raise more than $1.3 billion in venture capital and build $16 billion in market capitalization, and major companies have acquired more than 30 of those fledgling businesses.

Joe Dysart is an Internet speaker and business consultant based in New York City. 


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account