Not all advances in AI come from developing bigger models and using more data. An innovative University of Cambridge project suggests that refining tiny hardware components may present a more effective way of boosting performance. Martin Cooper MBCS reports.
Researchers have unveiled a new chip material inspired by the human brain that could dramatically cut the energy needed to run artificial intelligence. The team at the University of Cambridge says the approach mimics how neurons both store and process information, removing the constant back-and-forth between memory and the processor that makes modern AI so power hungry.
‘Energy consumption is one of the key challenges in current AI hardware’, said lead author Dr Babak Bakhit from Cambridge’s Department of Materials Science and Metallurgy.
Mimicking the brain’s way of doing things
Current AI hardware follows a decades-old design in which memory and computation are separate. That means data must be shuttled across a chip every time a model performs a calculation, a process that consumes a surprising amount of electricity. The Cambridge group argues that the brain’s way of doing things — where synapses both hold information and help compute — offers a more efficient blueprint.
At the heart of the work is a tiny device called a memristor, a component that can change its electrical resistance and thereby store information. Unlike conventional memory, memristors can also participate in computation, so a single circuit can both remember and calculate. The Cambridge researchers developed a memristor made from a modified form of hafnium oxide, a material already familiar to chipmakers, and report that it switches reliably at very low currents and can represent many distinct states — properties that are useful for hardware that learns.
For you
Be part of something bigger, join BCS, The Chartered Institute for IT.
The team’s press release notes that the new material ‘could slash AI energy use’ and describes the device as ‘inspired by the human brain’. Those phrases capture the promise: if chips could perform more work where the data lives, data centres and edge devices alike would need far less power to run the same models.
One of the practical advantages of the Cambridge design is stability. Many experimental memristors rely on tiny conductive filaments that form and break inside the device; those filaments can be unpredictable and wear out. The Cambridge device instead uses a different switching mechanism at an internal interface, which the researchers say gives much more uniform behaviour and a longer useful life. In tests, the component operated with currents a million times lower than some alternatives and could hold hundreds of stable levels of resistance — a useful trait for representing the subtle weight changes that underpin learning.
Running AI outside a data centre
For technology enthusiasts, the implication is straightforward: more efficient hardware could make powerful AI cheaper to run and easier to deploy outside the data centre. Imagine smartphones that run advanced language models without draining the battery, or sensors that learn locally without sending every bit of data to the cloud. The Cambridge work points towards that future.
There are, however, real hurdles before this becomes a product. The material needs to be processed at high temperatures during fabrication, a step that is not yet compatible with standard chip manufacturing. Engineers will need to find ways to integrate the new film into existing production lines or adapt the process to run at lower temperatures. Until that is solved, the technology will remain a laboratory project rather than a mass market solution.
The researchers have demonstrated that their memristor can mimic a basic form of learning found in biological synapses: the strength of its response changes depending on the timing of electrical pulses. That behaviour, known as synaptic plasticity, is a cornerstone of how brains adapt and learn. Embedding such dynamics in hardware could allow devices to learn continuously and efficiently, rather than relying solely on software updates.
For now, the work is a reminder that progress in AI won’t come only from bigger models and more data. Sometimes the biggest gains come from rethinking the hardware itself and borrowing a few lessons from the most efficient computer we know: the human brain.
Take it further
Interested in this and similar topics? Explore BCS' books and courses:
