Scientists at the University of Cambridge have developed a new type of nanoelectronic device that could reduce AI energy consumption by up to 70% by mimicking how the human brain processes information.
The device improves on existing memristor technology through controlled interface switching
The team engineered a modified hafnium oxide thin film by adding strontium and titanium and using a two-step growth process to create stable p-n junctions at material interfaces. Unlike conventional memristors that rely on forming and breaking conductive filaments—which cause unpredictable behavior and high voltage requirements—this design changes resistance by adjusting energy barriers at the interfaces. Lead author Dr. Babak Bakhit explained that this approach provides outstanding uniformity across switching cycles and devices, solving a key reliability issue in neuromorphic hardware.
Ultra-low power operation enables hundreds of stable conductance levels for analog computing
Tests showed the new devices operate at switching currents roughly a million times lower than some conventional oxide-based memristors. They can also achieve hundreds of stable conductance levels, which is essential for analog in-memory computing where data processing occurs directly within memory units. This capability supports more natural learning and adaptation in AI systems while drastically reducing the energy needed for data movement between separate memory and processing components—a major inefficiency in current AI hardware.
Energy efficiency gains could address growing power demands from widespread AI deployment
Modern AI systems consume large amounts of electricity due to constant data transfer between memory and processing units, a bottleneck that worsens as AI adoption expands across industries. By combining memory and processing in a single neuromorphic architecture—similar to brain function—the Cambridge innovation could cut energy use by as much as 70%. Dr. Bakhit emphasized that achieving such efficiency requires devices with extremely low currents, excellent stability, uniformity, and multi-state switching capability, all of which the new design demonstrates.
How does this chip differ from traditional AI hardware?
Unlike conventional chips that separate memory and processing, requiring energy-intensive data transfers, this neuromorphic device combines both functions in one place by mimicking neural connections, reducing energy use by up to 70%.
What makes the new memristor more reliable than previous versions?
It avoids filament-based switching, which causes unpredictable behavior, and instead uses controlled interface adjustments in engineered hafnium oxide films, ensuring uniform performance across cycles and devices.
Keep reading
