A synapse on a chip switches a million times lighter

The computer chip has spent half a century getting faster by getting smaller, and that era is visibly running into a wall that physics, not engineering, is putting up. Chips still consume large amounts of energy, and the energy appetite of running ever-larger artificial-intelligence models has become an industry-level worry: the International Energy Agency projects that data centres on AI workloads could consume more than 1,500 terawatt-hours of electricity a year by 2030. A paper published in Nature Electronics on 22 April, from Babak Bakhit and Manish Chhowalla at the University of Cambridge, proposes a different way forward, one borrowed not from the digital computer but from the brain.

The device is a memristor, a circuit element first theorized back in 1971 and only recently becoming practical, whose defining trick is that its resistance remembers the charge that has passed through it. Where an ordinary logic transistor is strictly on or off, a memristor can hold a continuous range of conductance values. That is exactly the property a brain synapse has: the connection between two neurons strengthens or weakens with use rather than flipping between two states. Build a circuit from memristors, and a single device can store a weight and compute with it in the same place, which is how the brain manages to do so much with so little power.

The Cambridge team built theirs from a thin film of hafnium oxide, a material already familiar from commercial chip manufacturing, which is a small but important detail. Novel materials are easy to dream up and hard to fab; a memristor made from a material the industry already knows how to deposit at scale is one that could plausibly ride on existing production lines. By controlling the structure of the film at the atomic scale, the group produced a switch whose conductance could be tuned smoothly and, crucially, stably.

The headline numbers are what make it notable. The device operates at switching currents roughly a million times lower than conventional silicon, and the team estimates such hardware could cut the energy demands of large AI systems by as much as seventy percent. In benchmark tests, an array of the memristors displayed spike-timing-dependent plasticity, the same biological rule by which real neurons strengthen the synapses between cells that fire in sequence. A small simulated neural network built on the devices classified handwritten digits and simple images with accuracy comparable to a conventional graphics processor, while drawing a small fraction of the power.

The honest caveats are the usual ones for a device this young. It remains a laboratory demonstration, a set of individual cells rather than a wafer-scale array, and the leap from a lab device to something that can sit on a fab line is the part that has defeated more than one promising architecture. The authors themselves estimate that any commercial deployment is years away. Several start-ups, among them Mythic AI and Rain AI, are pursuing similar in-memory computing designs, and a partnership with the British firm Pragmatic Semiconductor is only now beginning to explore early commercial routes.

None of that blunts the direction of the result. For years the assumed path to cheaper, more efficient AI has been to shrink the transistor and run it harder. A memristor that behaves like a synapse, and switches at currents a million times lower, suggests a different bargain: give up the universal digital transistor and build hardware shaped like the task it runs. The energy problem of AI is not going to be solved by a clever algorithm alone. It may, eventually, be solved by a chip that thinks a little more like the thing it is modelling.