Cornell Researchers Unlock Strategies for Compact Optical Computers

Researchers at Cornell University have made significant strides in the field of optical computing, revealing new strategies aimed at designing energy-efficient systems that utilize light instead of electricity. Their findings, published in Nature Communications, address a critical challenge: how to make optical computers small enough for practical use.

The study investigates the theoretical limits of light in computational processes. Just as digital computers require time and memory, optical systems need sufficient physical space for light waves to propagate and perform essential calculations. Lead authors Francesco Monticone, an associate professor of electrical and computer engineering, and Yandong Li, a postdoctoral researcher, explored scaling laws related to free-space optics and photonic circuits. They analyzed how the size of optical devices must increase as they tackle more complex tasks.

Monticone emphasized the challenges of optical computing, stating, “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.” He noted that photons, the particles of light, are more challenging to confine than electrons, making compact designs crucial for the technology’s viability.

To optimize space usage, the researchers drew inspiration from a deep-learning technique known as “neural pruning.” This method safely eliminates redundant parameters with minimal impact on performance. Li explained, “We specifically analyzed the connectivity pattern of these optical devices—how light waves overlap and interact in the entire device.” By developing optics-specific pruning methods grounded in wave physics, they reduced network complexity while preserving accuracy.

Their innovative approach demonstrated that an optical computing system could be as small as 1% to 10% the size of traditional counterparts while still performing the same tasks. To illustrate the potential, the team estimated the physical dimensions required for an optical computer to execute linear operations in large language models like ChatGPT, which can contain 100 billion to 2 trillion parameters. They suggested that a free-space optical setup could, in theory, function at this scale within a device only about 1 centimeter thick.

The researchers also identified a trend indicating diminishing returns in inference accuracy as the optical device size increases. This finding suggests that for certain applications, it is better to balance device size with performance. While fully optical computers remain a long-term aspiration, Monticone and Li see more immediate opportunities in hybrid systems. In these setups, light can efficiently handle rapid, energy-intensive linear operations, while electronic components manage nonlinear functions and decision-making processes.

Monticone expressed some skepticism regarding the potential for optical computers to replace or significantly enhance current technologies such as GPUs. He stated, “There are limitations other than size that make me personally skeptical about whether optical computers will really replace, or drastically accelerate, things like GPUs.” Nonetheless, he highlighted the promise of optical computing for applications like imaging and computing in resource-limited environments, noting that “space is not necessarily the bottleneck some people feared.”

As researchers continue to refine optical computing technologies, the implications for energy efficiency and performance in various applications could be profound, potentially paving the way for a new era of computational capabilities.