Cambridge researchers develop hafnium oxide memristor for low-energy AI chips

New neuromorphic chip material

Cambridge researchers have developed a nanoelectronic device using hafnium oxide that acts as a stable, low-energy memristor designed to mimic neural connections in the human brain. The work, published in Science Advances, addresses the energy consumption challenges in current AI hardware.

How it works

Unlike conventional AI systems that shuttle data between separate memory and processing units, this brain-inspired approach stores and processes information in the same place. The Cambridge team created a hafnium-based thin film that switches states differently from existing memristors.

Most memristors rely on conductive filaments inside metal oxide material, which behave unpredictably and require high voltages. The Cambridge device instead uses a two-step growth method with added strontium and titanium to form tiny electronic gates (p-n junctions) where layers meet.

This allows the device to change resistance smoothly by shifting the height of an energy barrier at the interface, rather than by growing or rupturing filaments. Lead researcher Dr. Babak Bakhit notes: "Because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device."

Performance metrics

• Switching currents approximately one million times lower than conventional oxide-based devices
• Produces hundreds of distinct, stable conductance levels
• Potential to reduce AI energy consumption by up to 70% compared to current hardware
• Excellent stability and uniformity across switching cycles

The researchers emphasize that effective AI hardware requires devices with extremely low currents, excellent stability, outstanding uniformity, and the ability to switch between many distinct states. This hafnium oxide memristor approach appears to meet these requirements while addressing the filament randomness problem that has limited previous memristor technologies.