Nonlinear Processing with Linear Optics and Optical Generative AI
As artificial intelligence (AI) use grows, the demand for computing resources has increased dramatically. For example, a large language model like ChatGPT, requires a large amount of computing power to operate. The server cost alone is $100,000 per day as it performs calculations on tens of thousands graphical processors (GPU). These costs are significant and highlight the need for more energy-efficient and cost-effective computing platforms for AI applications and future growth. Optical transceiver technology products using silicon photonics has been announced in March 2025 by NVIDIA to communicate between GPUs with more power efficiency and higher speed. In addition to communication, can Optics also be used efficiently to compute?
Optical computing hardware has several advantages, such as high bandwidth parallelism and energy efficiency. However, one major limitation is the implementation of nonlinear calculations in the optical domain. Current solutions for nonlinearity require the use of digital computers in conjunction with linear optical hardware, which is costly and inefficient in terms of energy consumption due to the costly and power-consuming Optical-Electrical-Optical (OEO) conversions that are required many times for large models. This has been the major limitation of the scalability of optical computing hardware.
We will discuss an approach that achieves the equivalent of optical nonlinearity by relying on multiple linear scattering off data encoded onto a spatial light modulator (SLM) that uses low optical power to effectively synthesize a nonlinear operation. By exploiting this relationship, arbitrary nonlinear transformations can be programmed digitally, and light effectively performs an all-optical computation without requiring electronic switching or high peak power to achieve non-linearity. We will show recent results in Optical Generative AI using this method.