Researchers have developed a new nanoelectronic device that could greatly reduce the energy used by artificial intelligence (AI). This device is inspired by how the human brain works.
A team from the University of Cambridge created a special form of hafnium oxide. It acts as a stable, low-energy "memristor." Memristors are designed to copy how neurons in the brain connect and communicate efficiently. These findings were published in Science Advances.
A Brain-Inspired Solution for AI Energy Use
Today's AI systems use standard computer chips. These chips constantly move data between memory and processing parts. This constant data movement uses a lot of electricity. As AI becomes more common, this demand is growing fast.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxNeuromorphic computing, which mimics the brain, offers a different way. It combines data storage and processing in one place. This could cut energy use by up to 70% and work at very low power. Systems built this way could also learn and adapt more easily, much like the brain.
Dr. Babak Bakhit, a lead author from Cambridge, explained that energy use is a big problem for current AI hardware. He said that to fix this, devices need very low currents, great stability, consistent performance, and the ability to switch between many different states.
Most current memristors create tiny conductive paths inside metal oxides. These paths can be unpredictable and often need high voltages. This limits their use in large-scale computing and data storage.
The Cambridge researchers tried something new. They made a hafnium-based thin film that changes states without relying on these paths. By adding strontium and titanium and using a two-step growth process, they created small electronic gates, called "p-n junctions," within the oxide where layers meet.
This setup lets the device smoothly adjust its resistance. It does this by changing the energy barrier at the interface, rather than forming or breaking paths. Bakhit noted that this design solves a big problem with existing memristors. He said that their devices show consistent performance because they switch at the interface.
Performance and Future Potential
The new devices use about one million times less current than some older oxide-based memristors. They also showed hundreds of stable conductance levels. These are important for analog "in-memory" computing.
In tests, the devices handled tens of thousands of switching cycles. They also held stored information for about a day. They even copied key biological learning behaviors, like spike-timing dependent plasticity. This is where connection strength changes based on signal timing between neurons.
Bakhit said these properties are needed for hardware that can learn and adapt, not just store data.
There are still challenges. The current manufacturing process needs temperatures around 700°C (1,292°F). This is too high for standard semiconductor production. Bakhit said this is the main challenge, but they are working to lower the temperature to fit industry standards.
Despite this, Bakhit believes the technology could eventually be used in chip systems. He noted that reducing the temperature and putting these devices on a chip would be a big step forward.
He added that this breakthrough came after years of trying different things. Progress happened when he changed the two-stage process to add oxygen only after the first layer formed. Bakhit spent almost three years on this project. He saw the first good results in late November. He believes that if the temperature issue is solved, this technology could be a game-changer due to its much lower energy use and promising performance.
