New brain-inspired AI chip promises major drop in data centre power consumption

A new type of AI chip developed by researchers at the University of Cambridge could reduce energy consumption in data centres by up to 70%, offering a potential breakthrough in tackling the rapidly rising power demands of artificial intelligence systems.

AI data centres are among the fastest-growing sources of global electricity use, according to the International Energy Agency, driven by the constant movement of data between memory and processing units. 

Each transfer consumes significant energy, making large-scale AI systems increasingly expensive to operate.

The Cambridge research team says their approach aims to eliminate this bottleneck entirely by redesigning how computation and memory interact.

Unlike conventional memristor-based devices that rely on unpredictable filament formation within metal oxides, the new chip uses a hafnium-based thin film enriched with strontium and titanium. 

Instead of forming filaments, it alters energy barriers at interface layers to process information more efficiently and consistently.

Dr Babak Bakhit of the University of Cambridge’s Department of Materials Science and Metallurgy, who led the research, said the interface-based switching method delivers improved stability and uniform performance across devices.

“Since our devices operate based on interface switching, we demonstrate excellent cycle-to-cycle and device-to-device uniformity,” the team noted in Science Advances.

In testing, the devices remained stable across tens of thousands of switching cycles and demonstrated spike-timing-dependent plasticity, a brain-inspired learning mechanism where connections strengthen or weaken based on timing.

Researchers say this capability is essential for enabling true in-memory computing, allowing chips to both store and process information in a way that mimics neural networks. 

The new design also supports more than 100 conductance states, a key requirement for advanced analogue computing that current memristor technologies struggle to achieve.

However, significant engineering challenges remain. The current fabrication process requires temperatures of around 700°C, which exceeds standard semiconductor manufacturing limits.

Dr Bakhit acknowledged the difficulty of the development process, noting the team faced years of setbacks.

“We’ve had thousands of failures before this, and we’ve been struggling for almost three years,” he said. “But we finally managed to get this temperature, and now we’re working on bringing it down further.”