Cambridge Develops Low-Energy Hafnium Memristor

Researchers at the University of Cambridge have developed a brain-inspired hafnium-oxide memristor that may reduce the energy use of artificial intelligence systems. The device combines memory and computation in the same component, unlike the conventional von Neumann model where data must move between memory and processors. The findings were published in Science Advances on 20 March.

Why the Device Matters

Modern AI systems consume large amounts of energy because data transfer between memory and processing units is costly. In many cases, moving information across chips or servers uses more power than the calculations themselves. The new memristor is designed to reduce this bottleneck by performing computation where data is stored.

How a Memristor Works

A memristor is a memory-resistor device with variable resistance. It can remember its resistance state after the current is removed. In neuromorphic computing, such devices act like artificial synapses, where resistance levels represent connection strength between neurons. This makes them useful for on-chip learning and brain-like processing.

Key Features of the Cambridge Device

• The device uses a hafnium-based oxide film with a titanium oxynitride layer.
• Its resistance changes smoothly through a p-n junction mechanism, rather than unstable filament formation.
• It required about a million times less current to switch than conventional oxide memristors in lab tests.
• The researchers estimated that wider use could cut AI energy consumption by up to 70%.
• The device showed synapse-like behaviour, including linear response and spike-timing-dependent plasticity.

Potential and Limitations

Hafnium oxide is already used in advanced CMOS transistors, which supports possible industrial scaling. The device also withstood tens of thousands of switching cycles and retained programmed states for about a day. However, fabrication currently needs temperatures of around 700 °C, which is higher than standard semiconductor manufacturing conditions. Further work is needed to align the process with commercial chip production.