ORNL Develops High-Efficiency GaN Power Converters for AI Data Centers
New gallium nitride semiconductors enable switching speeds up to 20 times faster than silicon, slashing the footprint and cost of power infrastructure.
Researchers at Oak Ridge National Laboratory (ORNL) have developed and validated a high-efficiency power converter utilizing gallium nitride (GaN) semiconductors. This advancement aims to drastically reduce the physical size and operational costs of power systems within AI data centers.
Developed at ORNL's Grid Research Innovation and Development Center (GRID-C), the project utilized GaN components provided by ROHM Semiconductor. These semiconductors allow for switching speeds 10 to 20 times faster than traditional silicon-based semiconductors while simultaneously reducing energy loss. The resulting converters are smaller and lighter, which ORNL notes will lower costs associated with delivery, installation, and ongoing maintenance. The project was managed by UT-Battelle for the U.S. Department of Energy's Office of Science.
The Silicon Bottleneck
Conventional power converters have long relied on silicon semiconductors to manage the transformation of current and voltage. However, as artificial intelligence scales, the demand for power systems has outpaced the capabilities of traditional silicon. Silicon-based systems are inherently bulkier and suffer from higher energy loss during the switching process, creating a bottleneck for facilities that require extreme power density and efficiency.
Scaling AI Infrastructure
The transition to GaN is critical for the scalability of modern AI infrastructure. Enterprise data centers typically house hundreds or thousands of servers, with each individual server requiring between four and eight converters. Prasad Kandula, an ORNL researcher, noted that size and weight add up quickly at this scale. By reducing the dimensions and weight of these individual components, operators can significantly lower the total cost of ownership and create a more flexible, affordable facility footprint.
Future Implications
While the current validation focuses on the efficiency and speed of the GaN components, the broader goal is the integration of these systems into the next generation of AI data centers. The ability to pack more power into a smaller space without increasing energy waste is expected to be a primary driver in reducing the environmental and financial overhead of large-scale compute clusters. Future developments will likely focus on the widespread deployment of these converters to meet the escalating energy demands of generative AI workloads.