A memristor design could change how AI chips work, cutting power use while enabling learning and adaptation. It points to a shift in computing systems.
Researchers from the University of Cambridge have developed a nanoelectronic device that could reduce energy use in AI hardware by mimicking the human brain. The team created a hafnium oxide–based memristor that is stable and operates at low energy, replicating how neurons connect.
Current AI systems rely on chips that move data between memory and processing units, leading to high energy use as demand grows across industries. Neuromorphic computing offers an alternative by storing and processing data in the same place, which can reduce energy use by up to 70%. This approach also enables systems to learn and adapt in a way similar to the human brain.
To support such systems, hardware must operate at low currents while maintaining stability, uniform switching, and the ability to handle multiple states. However, most memristors depend on conductive filaments in metal oxides, which behave unpredictably and require high voltages, limiting their scalability.
A different approach uses a hafnium-based thin film that switches states without relying on filaments. By introducing additional elements and using a two-step growth process, internal electronic junctions are formed at layer interfaces. This allows resistance to change by adjusting an energy barrier, instead of forming or breaking filaments.
This interface-based switching improves consistency across cycles and devices, addressing a limitation of traditional designs. The devices also operate at currents about a million times lower than some oxide-based memristors and can achieve hundreds of conductance levels, making them suitable for analog in-memory computing.
Tests show that these devices can endure thousands of switching cycles and retain their states for about a day. They also demonstrate learning behavior similar to biological systems, where connections change based on signal timing.
Despite these advantages, fabrication remains a challenge due to processing temperatures of around 700°C, which are not yet compatible with semiconductor manufacturing. Work is ongoing to reduce this temperature for easier integration into existing chip processes.
If this issue is resolved, the technology could be integrated into chip-scale systems, offering a path toward low-power, adaptive AI hardware. The development follows several years of experimentation, with progress achieved after refining the material growth process to improve device performance.
