Energy-efficient Cambridge memristor could power next-generation brain-like AI chips

Most modern AI systems run on traditional digital computers, where processing and memory are handled separately. This setup demands enormous amounts of power, especially for data-heavy tasks. With AI usage rapidly increasing, scientists like materials researcher Babak Bakhit warn that this level of energy consumption is becoming increasingly difficult to sustain.

Neuromorphic computing offers a different path. Inspired by the brain, these systems combine processing and memory in one place, using networks of artificial neurons and synapses. This design allows them to handle tasks more like a human brain - processing and storing information simultaneously - which could cut energy use by up to 70% compared to conventional systems.

At the heart of these architectures are memristors, or “memory resistors.” Unlike standard components, memristors can adjust and retain their resistance based on previous electrical activity, much like how synapses strengthen or weaken through learning. First theorized in 1971 and physically realized decades later, they’ve become key to building brain-like computing systems.

However, most existing memristors suffer from unpredictable switching behavior. This is largely due to their reliance on tiny conductive filaments that form randomly within the material. These devices also tend to require high voltages and additional safeguards, making them difficult to scale for practical use.

The Cambridge team tackled this problem by abandoning filament-based switching altogether. Instead, they engineered a new mechanism using a carefully designed interface within the material. By adding strontium and titanium to a hafnium oxide film, they created a p - n junction that allows the device to change resistance smoothly and reliably through controlled charge movement.

The results are impressive. The new memristor operates at an ultra-low current - millions of times lower than conventional versions - while offering hundreds of stable, adjustable states. It can also switch repeatedly over tens of thousands of cycles without losing its stored information for extended periods.

Looking ahead, the researchers aim to refine the manufacturing process, so it aligns with standard semiconductor production methods. Their ultimate goal is to scale this technology into full neuromorphic chips that could rival - or even outperform - today’s AI hardware in both efficiency and performance.