Cut AI energy 70% with brainlike chips
- University of Cambridge researchers said in April their hafnium-oxide memristor could cut AI hardware energy use by up to 70% by merging memory and compute. (cam.ac.uk) - The eye-catching detail is the device’s switching current — about a million times lower than some conventional oxide memristors — while avoiding unstable filament behavior. (cam.ac.uk) - It matters because AI’s power problem is shifting hardware attention beyond GPUs, but neuromorphic gains are still workload-specific and not yet mainstream. (nature.com)
AI’s energy problem is really a chip architecture problem. Most of today’s systems still burn power shuttling data back and forth between memory and compute, and that g(cam.ac.uk)odified hafnium oxide — that could cut AI hardware energy use by up to 70%. The result was published in *Science Advances* in March 2026 and highlighted by Cambridge on April 23, 2026. (cam.ac.uk) ### What is the new thing here? The device is a memristor — basically a component that can store information and process it in the same physical plac(nature.com). Cambridge’s version uses a modified hafnium oxide film with added strontium and titanium, built so the switching happens at an interface rather than through the usual random conductive filaments. (cam.ac.uk) ### Why does that save so much energy? A lot of AI power use comes from moving data, not just doing math. Neuromorphic hardware tries to dodge that by putting memor(cam.ac.uk)y as much as 70% in this kind of setup because the device runs at ultra-low power and avoids constant back-and-forth traffic. (cam.ac.uk) ### Why isn’t this just another memristor claim? Because the usual problem with memristors is messiness. Many rely on tiny conductive filaments forming and (cam.ac.uk)y modulating an energy barrier at an internal junction instead. The team says that gave it much better uniformity from cycle to cycle and device to device. (cam.ac.uk) ### What’s the standout number? The switching current. Cambridge says the device operated at currents about a million times lower than(cam.ac.uk)nt is the whole game if you want dense, efficient AI hardware instead of a lab curiosity. (cam.ac.uk) ### Does this mean GPUs are getting replaced? No — not anytime soon. GPUs are still the workhorses for training and a lot of inference, and the broader AI hardware push is still centered on parallel accelerators. But the f(cam.ac.uk)euromorphic chips, compute-in-memory designs, photonics, and other accelerators — because one architecture will not be optimal for every workload. (nature.com) ### Where could neuromorphic chips win first? Probably at the edge — sensors, wearables, always-on devices, and other places where battery lif(cam.ac.uk)getting closer to commercial usefulness, especially for ultra-low-power, real-time applications. That is a different lane from giant data-center training clusters, but it is still a huge market. (nature.com) ### What’s the catch? The catch is that “up to 70%” is not a universal AI discount. It depends on the workload, the surrounding system, the software stack, and whether the device can be manufactured reliably at scale. Even advocates of neu(nature.com)y and deployment at scale before it becomes a standard part of mainstream AI infrastructure. (nature.com) ### So what should you take from this? This is best read as a real hardware signal, not a solved industry transition. Cambridge seems to have shown a promising way to make memristors more stable and dramatically lower-power. If that survives scaling and productization, AI efficiency stops being only a “buy m(nature.com)e right chip for the right job” story. (cam.ac.uk)
