Cuts energy 70% with brain-like chip
- University of Cambridge researchers reported a brain-inspired memristor made from modified hafnium oxide that could cut artificial-intelligence hardware energy use by 70%. - The device switches with currents as low as about 10^-8 amperes and about a million times lower than some oxide memristors. - The work targets AI’s memory bottleneck by merging storage and computing in one device. (cam.ac.uk)
Artificial intelligence chips waste power mostly by shuttling data between memory and logic, and Cambridge researchers say they built a device that avoids much of that. (cam.ac.uk) The team, led by the University of Cambridge, reported a brain-inspired memristor made from modified hafnium oxide in *Science Advances* in March 2026. They said the design could reduce AI hardware energy use by as much as 70%. (cam.ac.uk) (science.org) A memristor is a component that stores information and processes it in the same place, more like a synapse than a standard chip. Conventional processors separate those jobs, creating the “von Neumann bottleneck” that burns time and electricity on data movement. (science.org) (cam.ac.uk) Most memristors switch by growing tiny conductive filaments inside an oxide, and those filaments can behave unpredictably. The Cambridge device instead switches at an interface between layers, using p-n junctions formed by adding strontium and titanium to hafnium oxide in a two-step growth process. (cam.ac.uk) (science.org) That change let the device operate at switching currents of roughly 10^-8 amperes, with the university saying that is about a million times lower than some conventional oxide-based devices. The paper also reports retention longer than 10^5 seconds and hundreds of conductance levels. (cam.ac.uk) (science.org) Those details matter because always-on AI at the edge — in sensors, wearables, cameras, and local agents — is limited by battery life and heat. Neuromorphic hardware tries to trade some general-purpose flexibility for lower-power inference and learning close to where data is generated. (science.org) (mdpi.com) The Cambridge group said hafnium oxide is already used in semiconductor manufacturing, which could make the materials stack easier to adapt than more exotic alternatives. But the work is still a device-level result, not a commercial chip shipping in phones or servers. (cam.ac.uk) (science.org) So the news here is not that data centers suddenly got 70% more efficient this week. It is that one research team showed a lower-current, more uniform memristor architecture that could make future low-power AI hardware more practical. (cam.ac.uk) (science.org)
