UC San Diego Prototype Chip Converts GPU Power at 96.2% Efficiency Using Piezoelectric Resonators

Engineers at UC San Diego built a prototype chip that converts 48-volt data center power to the 4.8 volts GPUs need at 96.2% peak efficiency, using piezoelectric resonators instead of the magnetic inductors found in conventional power converters. The results, published April 9 in Nature Communications, show the design delivers roughly four times more output current than previous piezoelectric-based converters, according to ScienceDaily and EurekAlert.

The Bottleneck

Data centers distribute electricity at 48 volts. GPU processors need between 1 and 5 volts. Bridging that gap efficiently, inside a small physical footprint, is a DC-DC step-down conversion problem that every data center faces. Conventional converters use magnetic inductors, and after decades of refinement, they are approaching their physical performance limits.

“We’ve gotten so good at designing inductive converters that there’s not really much room left to improve them to meet future needs,” said Patrick Mercier, professor of electrical and computer engineering at UC San Diego and the study’s senior author, as quoted by ScienceDaily.

The Hybrid Approach

The UC San Diego team, led by first author Jae-Young Ko, a Ph.D. student, combined a piezoelectric resonator with small commercially available capacitors in a configuration that creates multiple pathways for energy to move through the system. Piezoelectric components store and transfer energy through mechanical vibrations rather than magnetic fields, offering potential advantages in size, energy density, and manufacturing scalability, according to EurekAlert.

The hybrid design reduces wasted power and lessens strain on the resonator, enabling the system to handle large voltage conversions more effectively than either approach alone, ScienceDaily reported.

Production Reality

The chip is still early-stage. Piezoelectric resonators physically vibrate, which means they cannot be soldered to circuit boards using standard techniques. New integration strategies are needed before the technology can be deployed in production hardware.

“Piezoelectric-based converters aren’t quite ready to replace existing power converter technologies yet,” Mercier told ScienceDaily. “But they offer a trajectory for improvement. We need to continue to improve on multiple areas, materials, circuits and packaging, to make this technology ready for data center applications.”

The Agent Infrastructure Angle

The timing matters because GPU compute is the binding constraint on AI agent deployment. CoreWeave signed a multi-year GPU capacity deal with Anthropic this week. AWS brought its DevOps and Security agents to general availability. xAI is building its second massive data center cluster in Memphis. Every percentage point of efficiency at the power conversion layer compounds across thousands of GPUs. A technology that meaningfully reduces power waste per GPU rack changes the economics of the infrastructure buildout that agent scaling depends on.

The project was supported in part by the Power Management Integration Center, an Industry-University Cooperative Research Center funded by the National Science Foundation.