Columbia Engineering researchers presented an innovative solution: a 3D photonic-electronic platform that enables previously unheard-of bandwidth density and energy efficiency levels, opening the door for next-generation AI hardware. The study was published in Nature Photonics.
3D photonic chip module. Image Credit: Keren Bergman
Although AI systems hold the potential for revolutionary breakthroughs, energy inefficiencies and data transfer bottlenecks have hindered their development.
Keren Bergman, Charles Batchelor Professor of Electrical Engineering, led the study.
The study describes a novel approach to redefining energy-efficient, high-bandwidth data communication by combining photonics with cutting-edge complementary-metal-oxide-semiconductor (CMOS) electronics. This invention tackles important data flow issues, which continue to hinder the development of quicker and more effective AI technology.
In this work, we present a technology capable of transferring vast volumes of data with unprecedentedly low energy consumption. This innovation breaks through the long-standing energy barrier that has limited data movement in traditional computer and AI systems.
Keren Bergman, Charles Batchelor Professor, Electrical Engineering, Columbia University
Breakthrough in Data Communication
Alyosha Christopher Molnar, an Ilda and Charles Lee Professor of Engineering at Cornell University, and the Columbia Engineering team worked together to create a 3D-integrated photonic-electronic device that has a high density of 80 photonic transmitters and receivers in a small chip footprint. At only 120 femtojoules per bit, this device provides tremendous bandwidth (800 Gb/s) with remarkable energy economy. By greatly surpassing current standards, this innovation has a bandwidth density of 5.3 Tb/s/mm².
The chip's low-cost design, integration of photonic devices with CMOS electronic circuitry, and use of commercially made components pave the way for broad industry use.
Improving AI Hardware
The team's study addresses the long-standing scalability and energy efficiency issues by redefining how data is transferred between compute nodes. This technique eliminates conventional data localization limits and offers unparalleled energy savings and high bandwidth density by 3D combining photonic and electrical circuits.
AI systems can now effectively move enormous amounts of data using this cutting-edge technology, which supports distributed topologies previously unfeasible because of latency and energy constraints.
From massive AI models to real-time data processing in autonomous systems, the ensuing developments are set to unleash previously unheard-of performance levels, solidifying this technology as a foundation for future computing systems. Beyond AI, this strategy heralds a new era of high-speed, energy-efficient computing infrastructure and has revolutionary potential for disaggregated memory systems, telecommunications, and high-performance computing.
Collaborators
The Air Force Research Laboratory, Dartmouth College, and Cornell University's Molnar lab contributed to the cooperative study. Funding for the study came from the Advanced Research Projects Agency-Energy (ARPA-E) and the Defense Advanced Research Projects Agency (DARPA), highlighting its vital role in developing national technical capabilities.
Journal Reference:
Bergman, K., et al. (2025) 3D Photonics for Ultra-Low Energy, High Bandwidth-Density Chip Data Links. Nature Photonics. doi.org/10.1038/s41566-025-01633-0