Artificial intelligence (AI) is poised to drive groundbreaking technological advancements, but its progress has long been hindered by issues like energy inefficiencies and bottlenecks in data transfer. Now, researchers at Columbia Engineering have unveiled a game-changing solution: a 3D photonic-electronic platform that dramatically improves both energy efficiency and bandwidth density. This innovation represents a critical step toward creating faster, more capable AI hardware.

Published in Nature Photonics, the study, led by Keren Bergman, Charles Batchelor Professor of Electrical Engineering, introduces a novel approach that integrates photonics with advanced complementary metal-oxide-semiconductor (CMOS) electronics. By combining these two technologies, the researchers have developed a high-speed, energy-efficient data communication system that directly addresses one of the biggest hardware challenges in AI—moving large amounts of data quickly without consuming excessive power.

In the paper, Bergman explains, “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.”

To achieve this, the Columbia Engineering team collaborated with Alyosha Christopher Molnar, Ilda and Charles Lee Professor of Engineering at Cornell University, to create a 3D-integrated photonic-electronic chip. This chip features an impressive 80 photonic transmitters and receivers within a compact footprint. The result is a platform that delivers a high bandwidth of 800 Gb/s while consuming just 120 femtojoules per bit—far surpassing existing data transfer benchmarks. With a bandwidth density of 5.3 Tb/s/mm², this breakthrough sets a new standard for energy-efficient, high-speed communication.

Designed with cost-effectiveness in mind, the chip integrates photonic devices with CMOS electronic circuits and uses components manufactured in commercial foundries. This means the technology is primed for wide-scale adoption across industries.

This development redefines how data is transmitted between compute nodes, addressing the significant bottlenecks that have plagued energy efficiency and scalability in AI systems. Through the 3D integration of photonic and electronic chips, this platform achieves energy savings and bandwidth density that were previously unattainable. It breaks free from traditional data locality constraints, allowing AI systems to transfer large volumes of data efficiently, and supports distributed architectures that were once impractical due to energy and latency limitations.

While this breakthrough has transformative potential for AI, its implications extend far beyond that. The advancements achieved with this platform open up new possibilities for high-performance computing across a variety of applications. From large-scale AI models to real-time data processing in autonomous systems, this technology promises to significantly enhance computing power.

Moreover, the approach holds transformative potential for fields such as telecommunications, disaggregated memory systems, and other high-performance computing applications. With its focus on energy-efficient, high-speed computing infrastructure, this innovation is poised to usher in a new era of computing that can handle the growing demands of modern technologies.

By Impact Lab