The first two sessions in the “IOWN and Internet of Light” track at Upgrade 2024 focused on IOWN and its capabilities for optical transport. The next three discussed how photonics can improve computation, especially with machine learning (ML) and artificial intelligence (AI). In his talk, Physics & Informatics (PHI) Lab Scientist Tatsuhiro Onodera explained the mathematics behind deep neural networks and how optics can solve related challenges.
At the heart of a neural network is matrix-vector multiplication, which applies multiplication and addition – or multiply-accumulate (MAC) operations – on successive layers of neurons. What’s interesting is how these networks scale. “If you have n neurons,” Onodera said, “you have n squared operations.” The problem is that computing these operations digitally requires energy that also grows by n squared. Thus, the predicted hockey-stick curve in AI-driven global energy consumption.
Energy consumption for photonics, however, scales as n instead of n squared. Onodera pointed to eyeglasses, which in effect compute (input, scenery; output, corrected vision) without requiring any energy. The same kind of passive computation – and energy savings – occurs on a photonic chip. The PHI Lab has made additional progress by showing that any appropriately controlled optical system can be used for ML acceleration; and (in collaboration with NTT’s Device Technology Lab) by programing the refractive index distribution for enhanced computation on a lithium niobate nanophotonic chip.
In the next session, PHI Lab Scientist Ryan Hamerly said one way around the high hurdle of digital computation is to look back to the pre-digital era. Specifically, to the electronic mixer, a critical part of the super heterodyne receiver that became the basis for all modern radios. Applying that electronic mixing concept to the photoelectric effect allows you to perform many MACs, with potentially very low power and distortion.
Hamerly said there are other ways to achieve that result, such as thin film lithium niobate (TFLN). In any case, photoelectric mixing allows for a delocalized MAC that involves both cloud-based encoding of weights and clients that use those algorithms to co-process data locally. The advantages of this “magic state” include lower energy and latency and more privacy. In an experiment at MIT, where Hamerly is also a visiting scientist, he and his colleagues generated a magic state in one lab and then sent it over fiber on an 86 km roundtrip back for processing in another lab. The outcome exhibited standard eight-bit precision and orders-of-magnitude better energy efficiency than state-of-the-art CMOS electronics. They used off-the-shelf telecom equipment, but Hamerly said that you could also implement it on a chip. PHI Lab Postdoctoral Fellow Saumil Bandyopadhyay expanded on why this is so.
Building optical systems traditionally required a tabletop and a collection of optical components. Over time, however, miniaturized form factors made it possible to build complex optical systems on a silicon chip. The first stop for silicon photonics was communications, but it is now taking off as a platform for computing, light-detection-and-ranging (LIDAR) and other applications, including neural network processing. Another collaboration with MIT led to fabricating a chip that could encode data in the optical domain and do the computations optically. “We’re able to classify signals in this chip in optics in less than half a nanosecond,” Bandyopadhyay said. “And perform a classification… with the same accuracy we get on a digital system.”
In the final session, NTT Research President and CEO Kazuhiro Gomi interviewed PHI Lab Scientist Tim McKenna on TFLN. An alternative to silicon, lithium niobate has been used for data modulation in fiber optics communications because it exhibits the optoelectric effect. Another of its properties is high non-linearity, which enables the manipulation of light as it travels from one point to another.
Lithium niobate and TFLN (as it became known during its relaunch in the 2010s) are chemically the same material, but they diverge in terms of form, fabrication and application. McKenna said TFLN is one of the most exciting, emerging photonic technologies to handle looming communication challenges, such as the need for more, faster and less expensive interconnections in purpose-built AI data centers. Sensing and LIDAR are two other applications. And the special properties of TFLN, especially non-linearity, position it to be used for computing, a primary focus of the PHI Lab.
Lithium niobate has been around for many decades. PHI Lab Distinguished Scientist and Stanford University Professor Bob Byer was involved with some of the first crystal growth of this material in the 1960s. The recent breakthroughs in nanofabrication and manufacturability of TFLN have generated renewed interest, and not only in academic circles. “We’re also at the point where it can be commercialized,” McKenna said. “It’s not just a peculiarity that happens in an academic lab, it actually is going to be entering the marketplace and play a really important role here on out.”