The rapid growth of generative AI and large language models is pushing up both computing demand and power consumption. A new UNIST project will explore whether more of that computation can be performed with light, reducing the energy required for data movement and signal processing in conventional electronic AI hardware.

Led by Professor Il-Sug Chung of the Department of Electrical Engineering, with Professors Jongeun Lee and Heein Yoon as co-investigators, the project has received KRW 9 billion in government funding for research from 2026 through 2030. UNIST will lead the project in collaboration with Lambda Innovation.

The team will integrate lasers, photonic integrated circuits (PICs), and electronic integrated circuits (EICs) into a single photonic-electronic computing system. Directly modulated vertical-cavity surface-emitting lasers (VCSELs) will generate high-speed optical signals, while optical interference will perform core neural network computations with less reliance on energy-intensive electronic data movement.

To scale the architecture, the researchers will combine time-division multiplexing with optical signal distribution, allowing a limited number of physical optical channels to process larger numbers of inputs and weights.

The project will also address one of the central challenges in photonic neural networks: nonlinear activation. Optical systems can perform linear operations such as matrix multiplication efficiently, but nonlinear operations often require electronic post-processing. The team will instead exploit nonlinear effects arising from coherent optical interference and detection, reducing the need to move between optical and electronic processing.

The researchers aim to demonstrate an integrated system with approximately 100 optical channels and an energy efficiency of 100 trillion operations per second per watt (100 TOPS/W) at the optical computing core. Light sources, PICs, EICs, packaging, system boards, and AI algorithms will be co-designed to support larger AI workloads.

“Meeting the computational demands of generative AI will require new approaches that move beyond the energy-efficiency limits of conventional electronic processors,” said Professor Chung. “By integrating lasers, photonic circuits, and electronic circuits, we aim to establish core technologies for the next generation of AI computing hardware.”

The project extends Chung's work in heterogeneous photonic-electronic integration. Since 2025, his team has led a separate KRW 11 billion national project to develop an all-in-one frequency-modulated continuous-wave (FMCW) LiDAR chip and system, integrating lasers, silicon photonics, electronic circuits, and optical transmit-and-receive components.

While the LiDAR project applies photonic-electronic integration to 3D sensing, the new project extends the technology to AI computing. The two efforts will provide a technology base spanning sensing and computation, with potential applications in autonomous systems, robotics, physical AI, aerospace, and defense.