Cientistas of Universidade of Cambridge have developed a new nanoelectronic device. The technology imitates the functioning of neurons and promises to reduce the energy consumption of artificial intelligence systems by up to 70%. The advance uses a modified version of hafnium oxide.
The component acts as a high stability and low consumption memristor. Ele combines data storage and processing in the same location, unlike traditional chips that constantly transfer information between memory and processing units. Essa transfer consumes large amount of electricity. The team published the results in the journal Science Advances.
Dispositivo mimics neural connections in the human brain
The new material modifies hafnium oxide with the addition of strontium and titanium. The two-step manufacturing process creates p-n junctions at layer interfaces. Esses electronic channels control resistance more predictably.
Diferente of conventional memristors, which depend on unstable conductive filaments, this alternates states through the energy barrier at the interfaces. The result is smooth, uniform switching. Testes showed switching currents about a million times lower than in similar devices. The component maintains hundreds of stable conductance levels.
Isso enables in-memory analog computing. The researchers highlight that the device replicates firing time-dependent plasticity, a key mechanism for learning in biological neurons.
- Opera with extremely low currents
- Apresenta high uniformity between cycles and devices
- Mantém programmed states for about a day
- Demonstra brain-like learning behaviors
Current AI Consumo drives search for alternatives
Modern artificial intelligence Sistemas uses a lot of energy because of the architecture von Neumann. Dados circulate between memory and processor all the time. Demand is growing rapidly with the expansion of AI in various sectors.
Neuromorphic computing emerges as a solution. Ela combines memory and processing, as occurs in the brain. Especialistas estimate a reduction of up to 70% in energy expenditure. The Cambridge device moves in this direction with superior stability.
Equipe led by Babak Bakhit overcomes technical challenges
Dr. Babak Bakhit, from Departamento from Ciência from Materiais and Metalurgia from Universidade from Cambridge, leads the work. Ele also has an affiliation with Departamento’s Engenharia. The researcher spent years experimenting until he adjusted the manufacturing process.
The controlled addition of oxygen after the first layer was decisive. Positive results emerged at the end of 2025. The device showed stability over tens of thousands of switching cycles.
Ainda therefore, the current process requires temperatures around 700°C. Isso makes integration with standard semiconductor manufacturing difficult. The team is now looking to reduce this temperature.
Potencial application on neuromorphic hardware
If they overcome the thermal hurdle, the researchers believe the device could be integrated into chips at scale. The advancement would pave the way for more efficient and adaptable AI systems.
The component offers exceptional uniformity and the ability to switch between many states. Essas features are essential for hardware that learns naturally rather than just storing bits.
Próximos search steps in Cambridge
The post in Science Advances details the mechanism of extended asymmetric pn junctions. The title of the article is “HfO2-based memristive synapses with asymmetrically extended p-n heterointerfaces for highly energy-efficient neuromorphic hardware”.
The team continues testing to improve industrial compatibility. The focus is on reducing temperature without losing performance. Sucesso at this stage can accelerate adoption in practical AI applications.