Cambridge researchers build memristor that cuts AI hardware energy use
Researchers at the University of Cambridge have built a nanoelectronic device that stores and processes information in the same location, rather than shuttling data back and forth like standard computer chips. The device, published in Science Advances, could help reduce the electricity consumed by AI systems.
The component is a memristor made from hafnium oxide modified with strontium and titanium. It mimics how brain synapses adjust their connections. Today's AI hardware consumes vast amounts of electricity, and demand continues to rise.
Why conventional memristors fall short
Most memristors rely on tiny conductive filaments forming inside an oxide layer. These filaments behave erratically-varying from one device to the next and even from one cycle to the next. They also require high voltages to start and high currents to operate, which works against energy efficiency.
The Cambridge team took a different approach. Instead of depending on filament growth, they built a thin film that switches through what they call an interfacial mechanism. The key action happens at the boundary between two oxide layers, where the energy barrier can be adjusted smoothly.
"Filamentary devices suffer from random behaviour," said Dr. Babak Bakhit, the lead researcher. "But because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device."
What the lab results showed
The memristors operated at switching currents at or below 10-8 amperes. Energy per update ranged from roughly 2.5 picojoules to 45 femtojoules. The devices produced hundreds of stable, replicable conductance levels without saturating over 6,000 applied spikes.
That gradual tuning matters. Neuromorphic hardware works best when it can adjust weights in many small steps rather than jumping between a few rigid states. The memristors endured more than 10,000 pulse-switching cycles, with binary states remaining stable for up to 500,000 seconds.
The devices also reproduced several learning behaviors associated with biological synapses, including paired-pulse facilitation, paired-pulse depression, short-term synaptic plasticity, and spike-timing dependent plasticity. These traits allow hardware to learn from timing and sequence, not just store digital bits.
The manufacturing problem
The current fabrication process requires about 700 degrees Celsius. Bakhit called that "currently the main challenge," since it exceeds standard semiconductor manufacturing tolerances. The team is working to lower the temperature so the technology can integrate with industry processes.
The study also demonstrates device-level behavior but stops short of a full network-level hardware demonstration. The authors say that remains for future work.
Bakhit said the breakthrough came after three years of failed attempts. In late November 2024, the team modified the deposition method by adding oxygen only after the first layer had been grown. A patent application has been filed by Cambridge Enterprise.
What this means for AI systems
If the temperature issue can be solved, this type of memristor could help build AI hardware that uses far less electricity while handling learning tasks in a more brain-like way. That could matter for data centers, edge devices, and future AI systems that need to process large amounts of information without driving energy use higher.
Compute-in-memory designs like this could save more than 70% of current computing power consumption, according to the research.
Your membership also unlocks:
