Listen to article | 5 min
[February 19, 2022] – Earlier this month IARPA announced $4.8 million in funding for a project that is led by Boston-based electro-photonic computing startup Lightmatter and which is joined by BU and Harvard researchers. IARPA has conceptual affinity with DARPA, the Defense Dept.’s long-time running technology seeding lab, but in IARPA’s case the funding comes out of the Office of the Director of National Intelligence.
Lightmatter’s photonic computing efforts to date have centered on the idea of a supercomputer based on photonic silicon hybrid chips that specialize in performing large matrix computations. Proponents see the photonic domain as a logical next step beyond electronics, as the window on Moore’s Law closes. While barriers are significant, the photonic approach arguably requires shorter leaps forward than quantum computing techniques that have been extensively covered in the press. It seems photonic computing as represented in efforts by Lightmatter and others could be reasonably added to the basket of strange new brews being applied in the search for better AI processing today.
While optical data communications are well established – think “The Internet” — it’s early going for computing based on light waves. Lightmatter is one of few that has created photonic arrays that are capable of computation and which are matched with layers of digital and analog electronics that take care of much of the activity auxiliary to such computation.
A prototype board based on its chip was shown last year, as well as a box with six such chips and boards. So, any supercomputer is in the future.
Note that IARPA’s interest here is at the other end of the computing spectrum — farther from the data center, and closer to the edge. Among the obstacles Edge AI faces is power consumption, and today’s approaches to AI tend to use a lot of it.
Analog processing that calculates in memory has been put forward as a way to avoid costly (in power terms) movement of data from processor to memory in this way. The Lightmatter crew and others take a different tact. They see the optical domain as another way to save trips to memory for multiply-accumulate work that marks today’s neural nets for AI. Lightmatter is not alone in the quest.
Lightmatter is led by Nicholas Harris, a former MIT researcher whose interest in how computers work pans right down to the atomic sale, and which has led him to explore ways of coupling deep learning and Electro-Photonic Computing. His experience is by no means solely in the academic lab, as he served for a time as an R&D lead at Micron Technology in the high country of Idaho from where he originated. Lightmatter has raised more than $113 million from investors including Google Ventures and Spark Capital.
Edge AI-based silicon photonics is a bridge afar. Early going means basic chip architecture and processes need to be worked out before reliability, scalability and cost can be determined. There are still multiple approaches being taken to building in quite basic functionality as on-chips component. For example, how do assure that you are effectively shuttling light on chip? For its part, Lightmatter uses a MEMS-like structure which they call NOEMS, for “nano-opto-electro-mechanical”, based on phase shifters that operate at CMOS-compatible voltages. In effect this method guides a wave with the force of electrostatic charge, I think.
Cost is not an issue for IARPA, which is likely looking to improve the reach and ability of surveillance drones. But there are a lot of ramifications to this work If the present project bears fruit, it could further photonic computation in general terms, not just for AI apps. Moreover, what furthers photonic computation could also help spur quantum computing advancement.
A 2021 funding round saw Lightmatter gain backing from no less than he round was led by Viking Global Investors with participation from GV (formerly Google Ventures), Hewlett Packard Enterprise (HPE), Lockheed Martin, Matrix Partners, SIP Global Partners, Spark Capital, and others. – Jack Vaughan
IARPA project press release – Businesswire
Company site – Lightmatter
Citations/Nick Harris – Google Scholar
Useful silicon photonic back rounder – Ghent U.
Sally Ward-Foxton Lightmatter coverage from HotChips 2021 – EETimes