New Cambridge human brain-inspired chip could slash AI energy use — new type of memristor has roughly a million times lower switching current than conventional devices
The devices use a self-assembled p-n junction inside the oxide film instead of conductive filaments.
Researchers at the University of Cambridge published a paper in Science Advances earlier this month describing a new type of hafnium oxide memristor. The highlight of the new technology is that it operates at switching currents roughly a million times lower than conventional oxide-based devices.
The team, led by Dr. Babak Bakhit from Cambridge's Department of Materials Science and Metallurgy, engineered a multicomponent thin film that forms an internal p-n junction, enabling the device to switch states smoothly at currents below 10 nanoamps while producing hundreds of distinct conductance levels.
Memristors are two-terminal devices that can store and process data in the same physical location, eliminating the energy-intensive data shuttling between separate memory and processing units in conventional computer architectures. Neuromorphic systems built from memristors could reduce computing power consumption by more than 70%, according to the paper.
Most existing HfO2-based memristors rely on filamentary resistive switching, where conductive paths grow and rupture inside the oxide. These filaments exhibit stochastic behavior, resulting in poor device-to-device and cycle-to-cycle uniformity that limits computational accuracy.
A different approach - adding strontium and titanium
The Cambridge team took a different approach by adding strontium and titanium to hafnium oxide and depositing the film in a two-step process, thereby creating a p-type Hf(Sr,Ti)O2 layer that self-assembles a p-n heterointerface with an underlying n-type titanium oxynitride layer. Resistance changes occur by shifting the energy barrier height at this interface rather than by growing or breaking filaments.
"Filamentary devices suffer from random behavior," Bakhit said in a Cambridge press release announcing the work. "But because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device."
The devices demonstrated switching currents at or below 10-8 amps, retention exceeding 105 seconds, and endurance beyond 50,000 pulse-switching cycles. Using identical 1.0 V spikes comparable to biological neural signaling, the researchers achieved a conductance-modulation range exceeding 50 times across hundreds of distinct levels without saturation.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Synaptic update energy ranged from approximately 2.5 picojoules down to around 45 femtojoules. The devices also reproduced spike timing-dependent plasticity and maintained stable synaptic operation across roughly 40,000 electronic spikes.
One significant hurdle remains
The current deposition process requires temperatures of around 700°C, which exceeds standard CMOS manufacturing tolerances. "This is currently the main challenge in our device fabrication process," Bakhit said. "But we're now working on ways to bring the temperature down to make it more compatible with standard industry processes."
All materials used in the device stack are fully CMOS-compatible, and a patent application has been filed through Cambridge Enterprise.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
Luke James is a freelance writer and journalist. Although his background is in legal, he has a personal interest in all things tech, especially hardware and microelectronics, and anything regulatory.
-
usertests I've been hearing about memristors for decades. It's not coming soon, like this: https://www.tomshardware.com/tech-industry/artificial-intelligence/thermodynamic-computing-could-slash-energy-use-of-ai-image-generation-by-a-factor-of-ten-billion-study-claims-prototypes-show-promise-but-huge-task-required-to-create-hardware-that-can-rival-current-modelsReply
However, this does seem like the path we want to take. Low power, processing-in-memory, which could be similar to how neurons work.
Neuromorphic systems built from memristors could reduce computing power consumption by more than 70%, according to the paper.
Not 99.9%? And not to be confused with classical computing. -
Diogene7 The endurance seems far too low (10⁴–10⁵ cycles), and the retention time is also very limited (10⁴–10⁵ seconds!!!).Reply
For comparison, the FerroElectric Spin-Orbit (FESO) concept from the French lab Spintec could achieve an endurance of at least 10⁵–10⁶ cycles for inference, with retention times on the order of years or even decades.
It therefore appears to be a much more promising concept. -
Dementoss Greatly improved efficiency, is exactly what all computing needs, from AI slop farms, down to home PCs and mobile devices.Reply -
usertests Reply
It probably has no relevance to typical CPU/GPU designs, only AI/NPUs. They only talk about neuromorphic computing in the press release.Dementoss said:Greatly improved efficiency, is exactly what all computing needs, from AI slop farms, down to home PCs and mobile devices.