Replicating the brain’s capabilities, an impossible task, may theoretically require thousands of H100, one of NVIDIA’s most powerful GPUs. At 700 watts per chip, we are looking at power consumption in the megawatt range. The brain runs on 20 watts. Scientists have taken inspiration from this remarkable organ to create chips that could cut conventional energy use by 70%.
Researchers at the University of Cambridge have developed a new brain-inspired nanoscale device that they say could dramatically reduce the enormous energy demands of artificial intelligence hardware. The team created an ultra-low-power “memristor”: a device that can both store and process information in the same location, much like synapses in the human brain.
In conventional computing architectures, memory and processing units are physically separated, requiring data to shuttle back and forth between these units for every task. This seemingly simple process consumes enormous amounts of electricity and is a significant contributor to AI’s exploding power demands.
Researchers have increasingly looked toward neuromorphic computing as a possible solution. Instead of mimicking the architecture of traditional computers, neuromorphic systems aim to emulate how biological brains operate. In the human brain, neurons and synapses simultaneously store and process information through dense networks of electrical and chemical signaling. This architecture is extraordinarily energy efficient.
At the center of many neuromorphic computing concepts is a component known as a memristor. Unlike conventional transistors, memristors can retain memory states even when power is removed. They also behave somewhat like artificial synapses whose connection strengths can change over time.
However, existing memristors come with major limitations. Most oxide-based memristors operate by forming and rupturing tiny conductive filaments within the material. These microscopic conductive pathways form somewhat randomly, making the devices unpredictable from one switching cycle to another. They also typically require relatively high voltages and consume more power than researchers would like for truly energy-efficient AI hardware.
In their study, published in the journal Science Advances, the Cambridge team took a completely different approach.
Instead of relying on conductive filaments, the researchers engineered a hafnium-oxide-based material that switches states through controlled changes at an internal electronic interface. By adding strontium and titanium into hafnium oxide and fabricating the material using a two-stage growth process, the team created what are effectively microscopic p-n junctions inside the device. These are the same kinds of electronic junctions used throughout conventional semiconductor electronics.
Rather than forming and destroying conductive pathways, the device changes its electrical resistance by modifying the height of an energy barrier at this internal junction. This allows for much smoother and more controllable switching behavior. According to the researchers, this solves one of the biggest problems in memristor technology: variability.
“Filamentary devices suffer from random behavior,” says lead author Dr. Babak Bakhit. “But because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device.”
The energy savings are substantial. The researchers report switching currents as low as 10⁻¹¹ amps. For context, that is roughly a million times lower than those of some conventional oxide-based memristors. According to the paper, the switching energy falls within the femtojoule-to-picojoule range, comparable to or lower than that of some of the most energy-efficient neuromorphic hardware demonstrated so far. Researchers say this brain-inspired approach could reduce energy consumption in computing by more than 70%.
Another breakthrough is the device’s analog behavior.
Traditional digital systems largely operate in binary states: on or off. Biological synapses do not work this way. Their connection strengths change gradually. To emulate this, neuromorphic hardware requires components capable of holding many stable conductance states rather than just two.
The new memristors demonstrated hundreds of distinct, stable conductance levels without saturating, which the researchers say is critical for brain-like analog computing. The devices maintained smooth conductance modulation over thousands of electrical pulses, with remarkable consistency across cycles.
The team also demonstrated several behaviors associated with biological learning.
One of these is spike-timing-dependent plasticity (STDP), a learning mechanism found in biological neural networks in which the strength of connections between neurons changes based on the relative timing of their signals. The artificial synapses successfully reproduced these timing-dependent learning behaviors within millisecond-scale learning windows.
In simple terms, the hardware itself begins to behave less like static memory and more like adaptive brain tissue capable of learning.
Despite the promising results, significant hurdles remain before the technology can become commercially viable. One of the biggest challenges is manufacturing compatibility. The current fabrication process requires temperatures of around 700 °C (1,292 °F), far higher than standard semiconductor manufacturing tolerances. The team is now working on lowering those temperatures so the devices can eventually be integrated into conventional chip fabrication processes.
For now, the technology remains firmly in the research stage. But if the manufacturing challenges can be solved, the work could represent a major step toward ultra-efficient AI hardware that consumes only a fraction of the energy used by today’s systems.
Source: University of Cambridge