A new highly stable and energy-efficient memristor based on a hafnium oxide material can emulate the behaviour of synapses in the brain. The neuromorphic device could help dramatically cut the energy consumed by artificial intelligence (AI) hardware, say its developers at the University of Cambridge in the UK.
Today’s AI systems rely on conventional digital computers. These have separate processing and storage units and consume huge amounts of energy when performing data-intensive tasks. As global AI use is exploding, this energy consumption has already become unsustainable, says materials scientist Babak Bakhit, who led this new study.
An alternative way to process information
Neuromorphic computers could provide an alternative way to process information. As their name suggests, they are inspired by the architecture of the human brain. The circuits in these computers are made up of highly connected artificial neurons and artificial synapses that simulate the brain’s structure and functions. These machines have combined processing and memory units that allow them to process information at the same time as they store it, in the same way as a multi-tasking human brain. This means they could reduce energy consumption by as much as 70% compared with their digital counterparts.
Memory-resistors, or memristors, have become a fundamental building block of such neuromorphic architectures. This is because they can be engineered to behave very much like neurons in the human brain, which learn by reconfiguring the strengths of the connections (synapses) between neurons. Memristors excel in this respect as they can bring this learning functionality to the connections in electronic circuits.
First described theoretically in 1971, it was not until 2008 that researchers made the first practical version of a memristor. These devices are special in that their resistance can be programmed and subsequently stored. This is because, unlike standard resistors, the resistance of a memristor changes depending on the current previously applied to it – hence the “memory” in its name. What is more, the device “remembers” this resistive state even when the power is switched off.
Randomness in switching behaviour is a problem
All well and good, but most of today’s memristors unfortunately suffer from randomness in their switching behaviour because they rely on the formation of tiny conductive filaments in the materials making them up. These filamentary devices also typically require high forming and operating voltages and extra devices to avoid uncontrolled current changes that lead to permanent device failure. These challenges make such devices difficult to scale up for real-world applications, says Bakhit.
The researchers, who report their work in Science Advances, claim to have overcome the intrinsic stochasticity of memristive switching by exploiting a completely different switching mechanism – based on carefully engineered heterointerface physics rather than random filament switching. They achieved this by adding strontium and titanium to a hafnium-oxide thin film, which results in the formation of a p-n heterointerface. This junction allows the device to change its resistance smoothly by shifting the height of an energy barrier at the bottom interface through the migration of electro-ionic charges, explains Bakhit.
The new interfacial device has an ultralow switching current of less than or equal to 10-8 A, which is around 106 times lower than those of conventional oxide-based memristors. It also produces hundreds of distinct and stable conductance levels that can be easily modulated, a key prerequisite for analogue “in-memory” computing. And that’s not all: the device can also undergo tens of thousands of switching cycles without losing its programmed states for around a day.
Memristors could measure a single quantum of resistance
Looking ahead, the researchers say they will now be focusing on translating their material and device breakthrough into a functional computing system. “In particular, we are working on reducing the thin-film growth temperature (which currently stands at around 700 °C) so that it is compatible with standard semiconductor manufacturing (CMOS) tolerances,” says Bakhit. “We will then scale up device arrays to demonstrate large-scale integration.”
Ultimately, the goal is to move from individual devices to fully integrated neuromorphic chips that can compete with, or surpass, conventional AI hardware in both performance and energy efficiency, he tells Physics World.
