By studying the theoretical limits of how light can be used to perform computation, Cornell researchers have uncovered new insights and strategies for designing energy-efficient optical computing systems.
The research, published Sept. 29 in Nature Communications, addresses one of the key challenges to engineering computers that run on light instead of electricity: making those devices small enough to be practical. Just as algorithms on digital computers require time and memory to run, light-based systems also require resources to operate, including sufficient physical space for light waves to propagate, interact and perform analog computation.
Lead authors Francesco Monticone, associate professor of electrical and computer engineering, and Yandong Li, Ph.D. '23, postdoctoral researcher, revealed scaling laws for free-space optics and photonic circuits by analyzing how their size must grow as the tasks they perform become more complex.
"Optical computing can be powerful, especially in terms of energy efficiency, but if you need an optical setup that's as large as an entire room to accomplish a meaningful AI inference task, such as image classification, then your optical computer isn't very practical. Keep in mind that photons are much harder to confine in small spaces than electrons," Monticone said. "We wanted to understand the fundamental tradeoffs between task complexity, performance and minimum physical size, and then find ways to design optical inference systems that use the available space optimally."
To address this, the researchers found inspiration in a deep-learning technique called "neural pruning," which safely removes, or prunes, redundant parameters with little impact on performance.
"We specifically analyzed the connectivity pattern of these optical devices - how light waves overlap and interact in the entire device," Li said. "Then we developed optics-specific pruning methods grounded in wave physics to penalize the overlapping of light waves. This allowed us to simplify the network considerably with minimal loss in accuracy."
Using this technique, the researchers found that an optical computing system performing the same task could be 1% to 10% the size of its conventional counterpart. To put their findings in perspective, the researchers estimated how large an optical computer would need to be to perform the linear operations in large language models such as ChatGPT - more specifically, at a scale of 100 billion to 2 trillion parameters. They found that a free-space optical setup could, in principle, perform computations at this scale in a device roughly 1 centimeter thick. To approach this theoretical limit, emerging optical devices such as ultra-thin metasurfaces and spaceplates could be promising candidates, as the researchers noted in an earlier paper.
The study also revealed a trend of diminishing returns in inference accuracy as the optical device becomes larger, meaning that for some applications it is better to strike a balance between device size and task performance, according to the researchers.
While fully optical computers remain a challenging long-term goal, Li and Monticone see more immediate applications in hybrid systems, where light handles fast, energy-intensive linear operations, and electronics provides nonlinear functions, branching logic, decision-making and general-purpose programmability.
"There are limitations other than size that make me personally skeptical about whether optical computers will really replace, or drastically accelerate, things like GPUs," Monticone said. "But for many applications such as imaging and computing in resource-limited edge scenarios, optics could work extremely well, and we show that space is not necessarily the bottleneck some people feared."
The research was supported by the Air Force Office of Scientific Research and the Office of Naval Research.
Syl Kacapyr is associate director of marketing and communications for Cornell Engineering.
