A new chip design from UC San Diego could fundamentally reshape how data centers power graphics processing units (GPUs), addressing one of the fastest-growing energy challenges in computing. Researchers have created a hybrid power converter that achieves 96.2% efficiency when converting the 48 volts supplied to data centers down to the 4.8 volts required by GPU processors, while delivering roughly four times more power output than previous piezoelectric-based designs .

Why Is GPU Power Conversion Such a Critical Problem?

Data centers face an escalating energy crisis. U.S. data centers consumed roughly 4% of national electricity in 2024, with projections that this figure could reach 12% by 2028 . GPU-dense AI infrastructure consumes an order of magnitude more energy than traditional servers, making efficient power conversion no longer a technical optimization but a business necessity.

The challenge lies in a fundamental mismatch: electricity arrives at data centers at high voltage (typically 48 volts), but GPU processors need much lower voltages, usually between 1 and 5 volts. This large voltage drop has become increasingly difficult to manage efficiently as computing systems grow more powerful and compact. Traditional converters using magnetic components like inductors have reached their practical limits after decades of refinement.

"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 in the Department of Electrical and Computer Engineering at UC San Diego.

Patrick Mercier, Professor of Electrical and Computer Engineering, UC San Diego Jacobs School of Engineering

How Does the New Piezoelectric Approach Work Differently?

Instead of relying on magnetic inductors, the UC San Diego team investigated piezoelectric resonators, which store and transfer energy through mechanical vibrations rather than magnetic fields. This fundamentally different approach offers several theoretical advantages: smaller size, higher energy density, improved efficiency, and easier manufacturing at scale .

However, earlier piezoelectric converter designs struggled with two critical problems: they couldn't maintain efficiency when handling large voltage differences, and they couldn't deliver sufficient power output. The UC San Diego team solved these limitations by creating a hybrid design that combines a piezoelectric resonator with small, commercially available capacitors arranged in a carefully engineered configuration.

The prototype chip successfully converted 48 volts down to 4.8 volts with peak efficiency of 96.2%, while delivering roughly four times more output current than previous piezoelectric designs. This hybrid approach creates multiple pathways for energy to flow through the system, reduces wasted power, and lessens strain on the resonator, enhancing both efficiency and power delivery with only a slight increase in chip size .

Steps to Understand the Real-World Impact of This Technology

• Energy Savings Potential: A 96.2% efficiency rating means only 3.8% of power is lost as heat during conversion, compared to significantly higher losses in traditional magnetic converters, translating to measurable reductions in cooling costs and overall data center power consumption.
• Scale Implications: Data centers operate thousands of power conversion units simultaneously, so even small efficiency gains multiply across entire facilities, potentially reducing annual electricity bills by millions of dollars for large-scale operations.
• Environmental Impact: Lower power consumption directly reduces carbon emissions from data centers, helping organizations meet sustainability commitments as AI workloads continue to grow exponentially.

What Challenges Remain Before Real-World Deployment?

Despite its promise, the technology remains in early development stages. The most significant challenge is that piezoelectric resonators physically vibrate, which means they cannot be attached to circuit boards using standard soldering techniques. New integration strategies will be needed to incorporate them into commercial electronic systems .

Researchers view future work as requiring improvements across three critical areas: materials science, circuit design optimization, and packaging methods. The team emphasizes that piezoelectric-based converters are not yet ready to replace existing power converter technologies in production data centers, but they represent a promising trajectory for overcoming the constraints of current systems.

"Piezoelectric-based converters aren't quite ready to replace existing power converter technologies yet, 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," explained Patrick Mercier.

Patrick Mercier, Professor of Electrical and Computer Engineering, UC San Diego Jacobs School of Engineering

Why Does This Matter Now?

The timing of this breakthrough aligns with a critical inflection point in AI infrastructure. Energy has become the primary constraint for scaling AI systems, surpassing GPU availability as the limiting factor. Financial services firms and other enterprises are increasingly evaluating data center locations based on power availability and cost, with some companies choosing geographically distant locations like Iceland specifically for access to renewable energy at five to ten times lower cost than metropolitan hubs .

The UC San Diego innovation addresses a fundamental bottleneck in this infrastructure: the conversion of power from the grid to the processors themselves. By improving efficiency at this critical junction, the technology could enable more powerful AI systems to operate within existing power budgets, or alternatively, allow the same computational capability with significantly reduced energy consumption and associated costs.

The research was published in Nature Communications and was supported in part by the Power Management Integration Center, an Industry-University Cooperative Research Center funded by the National Science Foundation .