Ongoing advances in electronics and computing have introduced opportunities to achieve things that once seemed inconceivable: build autonomous machines, solve complex deep learning problems, and communicate instantaneously across the planet. Yet, for all the advances, today’s systems—which rely on electronic processors—are grounded in a frustrating reality: the sheer physics of electrons limits their bandwidth and forces them to produce enormous heat, which means they draw vast amounts of energy.
As demand for fast and low-energy artificial intelligence (AI) grows, researchers are exploring ways to push beyond electrons and into the world of photons. They are replacing electronic processors with photonic designs that incorporate lasers and other light components. While there is skepticism among some observers that the technology can transform analog computing, researchers in the optical space are now building systems demonstrating significant benefits in AI and deep learning.
“Photonic processors can resolve the bottlenecks associated with today’s electrical computing systems. They are highly efficient from an energy perspective and they can overtake the standard clock rates of electronic systems by almost two orders of magnitude,” according to Maxim Karpov, a researcher at Switzerland’s École Polytechnique Fédérale de Lausanne (EPFL). However, considerable challenges remain in optimizing photonics in integrated circuits, including finding the right mix of materials to replace silicon, which does not perform well with optical, and improved packaging techniques.
Figure. Demonstrating the principle underlying the photonic processor by showing the spread of light in a matrix of phase-change materials.
Nevertheless, the technology is emerging from research labs and popping up in real-world systems, including from a handful of commercial startups. The possibilities are particularly enticing in areas such as deep learning, machine learning, and quantum computing. Technical advancements, including miniaturization and better packaging, are pushing the field forward at a rapid clip. Says Karpov, “Photonic computing and especially the area of integrated photonic computing, which uses silicon-based chips for optical signal processing, is actively evolving and beginning to make an impact.”
Seeing the Light
The idea of using light to speed processing is rooted in research from the 1980s. Yet, until recently, the idea had mostly stalled out. For one thing, the level of miniaturization required for components did not exist. For another, lasers and other components were not ready for primetime. As a result, the focus has mostly remained on eking out performance gains from conventional computing frameworks.
Although challenges still exist in the optical space—for example, it is not clear whether researchers package photonics in a way that actually delivers widespread benefits in general computing systems—the field is advancing. “With the rise of machine learning and artificial intelligence, photonic processors found a field in which it can shine. Increasing volumes of data bring current electronic technologies to their limits,” says Johannes Feldmann, a postdoctoral researcher at Oxford University in the U.K.
Optical technology, already widely used for cabling, communications, and increasingly, system interconnects, takes direct aim at the growing limitations surrounding Moore’s Law and von Neumann architectures. When compute-intensive tasks such as deep learning are tossed at electronic processors, they choke on advanced tasks and they devour energy. Moreover, scaling up systems to handle increasingly complex tasks is cost prohibitive. On the other hand, optics excels with low-precision linear functions. “This is where photonic processers challenge electronic processors: as hardware accelerators for artificial intelligence,” Feldmann says.
It is important to recognize the key differences between electrical and optical systems, says Nathan Youngblood, an assistant professor in the Electrical Engineering Department at the University of Pittsburgh. At the most basic level, electrical systems constantly change the number of electrons in a line, thereby charging and discharging metal interconnects that span two logic gates on the chip. “Fundamentally, optics is not limited by the charging and discharging of the interconnect line, so you can transfer data at much higher speeds. You don’t have a trade-off between energy consumption and modulation.”
Photons are attractive for performing computations because, unlike electrons, they can occupy the same physical state as other photons. This makes them more efficient and ideally suited to handle matrix-vector multiplications (MVMs) and convolutions used for deep learning. Light signals, which are modulated to encode input data vectors, are sent to the optical chip. “The light then propagates through the mesh of photonic waveguides. It is passively attenuated and mixed to transform it so that it conforms to the data matrix we want to use for multiplications,” Karpov explains. As the chip generates output, the light bearing the result of the multiplication operation is detected.
“Many of the obstacles that have prevented the technology from advancing are now being solved.”
Since photons can propagate within the chip in an ultra-efficient manner, the system puts the power it draws to maximum use. At the same time, the light modulation speed can easily hit tens of gigahertz, which radically boosts the throughput of the system in comparison to electronic components. Finally, multiple elements can be placed on a single chip, including modulators, detectors, and even light sources. This makes the technology ideal for a wide variety of other uses, including optical data transmission, spectroscopy, LiDAR, MRI scans, and even optical circuit switching in datacenters.
Fueling these advances are new photonic platforms that use materials such as silicon nitride and lithium niobate, and fabrication processes for extremely low-loss photonic waveguides based on these materials. Meanwhile, a growing number of commercial foundries are equipped to build photonic integrated circuits (PICs), and startup companies such as Lightelligence, Lightmatter, and Optalysys are introducing solutions that address various advanced computing and communications tasks. Says Youngblood, “Many of the obstacles that have prevented the technology from advancing are now being solved.”
Designs on Speed
Significant breakthroughs in photonic computing have appeared over the last few years. Some of the biggest advances revolve around the fundamental way these systems are designed and constructed. Packaging and interconnects are evolving rapidly and enabling even more sophisticated capabilities. Says Karpov, “We’re seeing a gradual transition from isolated chip-based photonic components to more complex photonic systems, where various technologies and material platforms are integrated on the same chip in a hybrid way.”
In fact, packaging is crucial. “Without the right packaging, you cannot fully take advantage of the functionality and performance of the integrated circuits,” says Paul Fortier, senior engineer for Photonic Packaging Development at IBM’s assembly and test facility is Bromont, Canada. “Legacy photonic assembly is too often manual, time consuming, and difficult to bring the technology into high-volume production.” IBM is focused on developing low-loss optical interconnects, thermal management packaging, integrating photonics with microelectronics, and further miniaturizing components. The goal is to “allow light to get on and off the chips with the highest bandwidth in the smallest space, all the while being low-cost, reliable, and scalable for automation,” he says.
In February 2021, the field took a significant leap forward when a group of researchers—including Karpov, Young-blood, and Feldmann—introduced a new architecture for photonics that combines processing and data storage on a single chip. With this design, the group developed a hardware accelerator for MVMs that serves as the basis for neural networks. Relying on different light wavelengths that do not interfere with each other, they were able to build a processor that handles complex parallel calculations and produces highly accurate results on convolution operations.
The technology framework, which is built using microresonator-based optical frequency combs (microcombs), provides a simple and straightforward way to parallelize computing operations on photonic processors, Karpov explains. Microcombs create ultra-compact light sources providing multiple equidistantly spaced optical frequencies. They allow the photonic tensor core to accommodate simultaneous data transfer and computing at speeds comparable to those of fiber networks, while generating near zero heat.
Nevertheless, further advances are required to push the technology into the mainstream—particularly in areas such as machine vision, which require ultra-fast calculations, Youngblood says. One obstacle, for now, is that photonic devices are physically larger than electronic transistors—even if the computing density, speed, and output is considerably higher for photonics. This makes optical chips unsuitable for certain tasks and situations. Another factor is that certain types of optical processors, such as free-space designs that rely on diffractive optics, introduce barriers related to the stability of the setup and the slow modulation speeds of spatial light modulators, Feldmann explains.
Scaling can also be a problem because photonic architectures still rely on electronic control circuits, which create a bottleneck. “The photonic processor itself could easily handle much higher data rates,” Feldmann says. “The photonic chip operates at a very low power level. However, the electronic control circuit driving it introduces much higher power requirements.” This means that further improvements in electronics are necessary to drive better photonic performance.
Finally, photonic foundries remain relatively immature, and the fabrication of photonic circuits must be more reproducible in the specifications of individual components such as multiplexers and light sources, he adds.
A Bright Future?
The jury is still out on whether photonics will deliver niche benefits to computing or revolutionize the space. “It is possible that photonics may play a role in some analog aspects of computing, such as in neural network systems. And there is great potential for photonics to help relieve the communications bottleneck at the edge of electronic chips,” says Rod Tucker, Melbourne Laureate Emeritus Professor at the University of Melbourne and former director of the Institute for a Broadband-Enabled Society (IBES).
However, Tucker believes formidable challenges remain for swapping out digital electronic processing with digital photonic processing. As a result, a general-purpose photonic computer is not likely to appear anytime soon. “There is no photonic device that come anywhere near a digital electronic gate in terms of miniaturization, low energy consumption per operation, logic level restoration, and noise suppression. And no photonic device can store a bit of digital data as efficiently or as long as an electronic memory cell,” he explains.
Moreover, Tucker says, “There have been recent examples of clever experiments that show how photonic devices can emulate digital electronics, but the challenges emerge when one tries to scale up to the processing capacity of a state-of-the-art electronic chip containing millions of ultra-low-energy devices.” He believes a focus on direct and fair comparisons with state-of-the-art digital electronics is paramount.
Feldmann says critics often miss the mark on the role of photonics. “An optical general-purpose processor is not very close to reality, but photonic processors shine currently in accelerating AI workloads.” For example, Lightmatter—a photonics AI startup rooted in MIT—generates 1.2 million inferences per second on a ResNet50 deep learning architecture versus 300,000 on an Nvidia DGX GPU. “This is for a full hybrid electro-optic system,” he notes. “Other startups that focus on Theoretical Operations Per Second (TOPs/W) also beat electronics by a substantial margin.”
The benefits of this alone could be significant. For instance, AI-related computing already consumes a considerable chunk of global energy, and the trendline indicates that increased demand for computing resources is nearly certain in the future, particularly as autonomous vehicles, robotics, and other machines demand more data-intensive input and output. Global sustainability hangs in the balance. “Photonic processors could reduce power consumption substantially,” Feldmann points out.
The biggest gains, however, would likely center on radically higher clock rates and parallelization that take machine learning and deep learning to an entirely different level—and unlock previously unachievable results. Optical signals can be modulated at up to 100 GHz, which opens the door to new and different uses. “For now, photonic processing makes sense where both high throughput and a high level of parallelization is needed on linear operations, such as matrix-vector multiplications,” Feldmann says.
Although electronic microprocessors will continue to serve as the backbone of computing for the foreseeable future, photonic systems could begin to change computing—and many aspects of life. As researchers learn how to fully integrate electronic and photonic components into single systems and package them effectively, Markov sees a bright future for the field. Ultimately, “The technology is likely to lead to a variety of application-specific photonic processors that will support ongoing advances in digital technology and the rise of quantum computing.”
Further Reading
Patterson, D., De Sousa, I., and Archard, L.
The future of packaging with silicon photonics. Chip Scale Review, January 2017, https://www.ibm.com/downloads/cas/M0NL8N85
Feldmann, J., Youngblood, N. Karpov, M., Gehring, H., Li, X., Stappers, M., Le Gallo, X., Fu, A., Lukashchuk, A., Raja, A.S., Liu, J., Wright, C.D., Sebastian, A., Kippenberg, T.J., Pernice, W.H.P., and Bhaskaran, H.
Parallel Convolutional Processing Using an Integrated Photonic Tensor Core. Nature, 589, pp.52–58 (2021), February 23, 2021.
Shastri, B.J., Tait, A.N. Ferreira de Lima, T., Pernice, H.P., Bhaskaran, H., Wright, C.D., and Prucnal, P.R.
Photonics for Artificial Intelligence and Neuromorphic Computing, Nature Photonics, 15, pp. 102–114 (2021), January 29, 2021.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment