Cambridge researchers develop low-energy chip material inspired by brain structure
Researchers at the University of Cambridge have created a new type of hafnium oxide device that could reduce the energy consumed by AI hardware by as much as 70 percent. The material mimics how neurons connect in the human brain, storing and processing information in the same location rather than shuttling data between separate memory and processing units.
The device is a memristor - a component designed to switch between different electrical states with minimal power consumption. Current AI systems waste enormous amounts of electricity moving data back and forth between memory and processors. As AI adoption expands across industries, global demand for electricity to power these systems is exploding.
A different switching mechanism
Most existing memristors rely on tiny conductive filaments that form inside metal oxide material. These filaments behave unpredictably and require high operating voltages, limiting their usefulness in large-scale systems.
The Cambridge team took a different approach. By adding strontium and titanium to hafnium oxide and using a two-step growth method, researchers created tiny electronic gates called p-n junctions at layer interfaces. The device changes resistance by shifting an energy barrier at these interfaces rather than growing or destroying filaments.
This mechanism solves a fundamental problem in memristor design: unpredictable behavior. "Filamentary devices suffer from random behaviour," said lead researcher Babak Bakhit from Cambridge's Department of Materials Science and Metallurgy. "But because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device."
Performance in testing
Laboratory tests showed the hafnium-based devices achieved switching currents about a million times lower than conventional oxide-based devices. The memristors also produced hundreds of distinct, stable conductance levels - a requirement for analog in-memory computing.
The devices reliably endured tens of thousands of switching cycles and maintained their programmed states for around a day. They also reproduced spike-timing dependent plasticity, a biological learning mechanism where neuron connections strengthen or weaken depending on signal timing.
Manufacturing challenges remain
The current fabrication process requires temperatures around 700 degrees Celsius, higher than standard semiconductor manufacturing tolerances. Bakhit said this is the main obstacle to scaling the technology.
"If we can reduce the temperature and put these devices onto a chip, it would be a major step forward," he said. The team is working on methods to lower the manufacturing temperature to align with standard industry processes.
The breakthrough came after three years of unsuccessful experiments. In late November, Bakhit modified the two-stage deposition method by adding oxygen only after the first layer had grown, producing the first promising results.
A patent application has been filed. The work appears in Science Advances and was supported by the Swedish Research Council, the Royal Academy of Engineering, the Royal Society, and UK Research and Innovation.
Learn more about Generative AI and LLM systems and their hardware requirements, or explore resources for AI for Science & Research professionals.
Your membership also unlocks:
