Scientists have developed a new type of nanoelectronic device that can significantly reduce the energy consumed by artificial intelligence systems. This innovation works by copying the way the human brain processes information, providing a more efficient alternative to today’s power-hungry AI hardware.
A research team led by the University of Cambridge has developed an improved version of hafnium oxide that acts as a highly stable, low-energy ‘memristor’. The memristor is a component designed to replicate how neurons connect and communicate in the brain. Their findings were published in the magazine scientific progress.
Why current AI systems use so much energy
Modern AI relies on traditional computer chips that constantly move data between memory and processing units. This round-trip transfer requires large amounts of power, and demand continues to increase as AI becomes more widely used across industries.
Neuromorphic computing offers a different approach. Similar to how the brain works, rather than separating memory and processing, it combines both in one place. This method can reduce energy usage by up to 70% while allowing the system to learn and adapt more naturally.
“Energy consumption is one of the key challenges in modern AI hardware,” said lead author Dr. Babak Bakhit from the University of Cambridge’s School of Materials Science and Metallurgy. “Addressing this requires devices with extremely low currents, excellent stability, excellent uniformity across the switching cycle and device, and the ability to switch between many different states.”
A new approach to memristor design
Most existing memristors work by forming small conductive filaments inside a metal oxide material. These filaments tend to behave unpredictably and often require high voltages, which limits their practicality in large-scale computing.
The Cambridge researchers took a different path. They designed a hafnium-based thin film that switches states through a more controlled mechanism. By adding strontium and titanium and using a two-step growth process, they created a small electronic gate known as a ‘pn junction’ at the interface between the layers.
Rather than relying on filament formation and breaking, this device changes resistance by adjusting energy barriers at these interfaces. This allows for smoother and more reliable switching.
Bakhit, who is also at the University of Cambridge’s School of Engineering, explained that the design solves a key problem in memristor development. “Filament-type devices suffer from random behavior,” he said. “But because our device switches at the interface, it exhibits excellent uniformity from cycle to cycle and device to device.”
Ultra-low power consumption and brain-like learning
Tests showed that the new device operates with approximately one million times lower switching current than traditional oxide-based memristors. It is also possible to achieve hundreds of stable conductance levels, which is essential for analog “in-memory” computing.
In laboratory experiments, the device remained stable through tens of thousands of switching cycles and maintained its programmed state for about a day. They also demonstrated important biological learning behaviors, including spike-timing-dependent plasticity, the process by which neurons strengthen or weaken connections based on timing.
“If you want hardware that can learn and adapt, not just store bits, you need these characteristics,” Bakhit said.
Remaining issues and future prospects
Despite promising results, there are still obstacles to overcome. Current manufacturing processes require temperatures of approximately 700°C, higher than typically tolerated in standard semiconductor manufacturing.
“This is currently the main challenge in our device manufacturing process,” Bakhit said. “However, we are currently working on ways to lower the temperature to make it more compatible with standard industry processes.”
If this problem can be solved, the technology could be integrated into practical chip-scale systems. “If we can lower the temperature and put these devices on a chip, that would be a big step forward,” he said.
Years of trial and error behind groundbreaking progress
This progress came after several years of experimentation and many setbacks. Bakhit said progress was only accelerated late last year when the company changed its manufacturing process to add oxygen only after the first layer was formed.
“I spent almost three years on this,” he said. “There were a huge number of failures. But at the end of November, for the first time, we got very good results. It’s still early days, of course, but if we can solve the temperature problem, this technology could be a game-changer, because the energy consumption is very low and at the same time the performance of the device is very promising.”
This research was partially supported by the Swedish Research Council (VR), the Royal Academy of Engineering, the Royal Society, and UK Research and Innovation (UKRI). The patent application was filed by Cambridge Enterprise, the university’s innovation arm.
