Korean researchers develop next-gen optical modulator to tackle AI data center power, bottleneck woes

Korean researchers have developed a technology that could simultaneously resolve the power consumption and data transmission bottleneck problems plaguing AI data centers.

KAIST announced Tuesday that a research team led by professor Kim Sang-hyeon of the School of Electrical Engineering had developed a next-generation optical modulator combining the strengths of different semiconductor materials. The team worked in collaboration with Han Jae-hoon of the Korea Institute of Science and Technology, Kim Jong-min of the Korea Advanced Nano Fab Center and Samsung Electronics' packaging division.

AI data centers rely on vast numbers of servers and semiconductor chips exchanging large volumes of data in real time. In that process, optical modulator performance is a key factor determining data transmission speed and power efficiency. Conventional silicon-based optical modulators, however, have been vulnerable to high heat generation and temperature fluctuations, and faced an inherent trade-off: boosting efficiency came at the cost of speed, and increasing speed reduced efficiency.

To overcome these limitations, the research team retained a Mach-Zehnder structure — known for its resilience to temperature changes — while bonding a thin film of indium phosphide-based semiconductor, which offers superior electro-optical responsiveness, onto a silicon waveguide, the channel through which light travels. By layering a specialized semiconductor material capable of more sensitive light control onto conventional silicon, the team reduced the device's physical size while achieving more efficient control of optical signals.

The team succeeded in simultaneously achieving world-leading modulation efficiency and high-speed operating bandwidth in an ultra-compact device measuring just 500 micrometers in length — tens of times smaller than the width of a human hair. The advance means more data can be transmitted faster using less power, improving both energy efficiency and data processing performance in AI data centers.

The research is considered particularly significant because it offers a practical solution to the twin challenges of rising power consumption and data transmission bottlenecks facing AI data centers. The device's compact size and low-power characteristics are expected to make it a key component in next-generation optical communication chips for data centers and in co-packaged optics applications.

"This technology contributes to reducing power consumption, and the thermal stability of the Mach-Zehnder structure enables stable operation in the high-heat environment of data centers without the need for separate thermal control," Kim said. "We are aiming to achieve the level of performance and yield needed for practical application in AI data center optical interconnects through further optimization of the device structure and fabrication process."