HP Rewires Electronics
A new electronic device could lead to denser, faster kinds of memory, and processing chips that act more like the brain.
A novel building block for electronic circuits has been developed by researchers at HP Labs. Called a memristor, it could be used to make extremely compact, low-power memory chips, as well as processors that mimic natural neural networks to perform tasks such as face recognition and controlling robots.
The new memristors are inspired by a nearly 40-year-old theory. Electrical engineers build circuits out of three passive electronic devices: resistors, inductors, and capacitors. But in 1971, a researcher named Leon Chua predicted that a fourth device–the memristor–should be possible and that it would have useful electronic properties. Now the HP researchers have made a working memristor, showing that Chua was right.
Memristors can be used as simple switches that turn current off and on, acting something like the transistors used in today’s computer chips. But unlike transistors, which are engineered to flip quickly between just two distinct states, memristors can easily be set to a range of different resistance levels. What’s more, once they’re set, they remember the resistance setting. The ability to remember the setting makes them useful for faster, denser types of nonvolatile memory.
What’s more, memristors’ ability to be set to a range of values makes them similar in some ways to the synapses between neurons in the brain. That could make them useful for chips that can learn to recognize faces and speech potentially much better than conventional computer chips can. They could also be useful in processors that allow robots to, for example, learn to walk.
The device is actually very simple, says R. Stanley Williams, who led the work developing the memristor, which is described in the current issue of Nature. A memristor consists of two thin layers of titanium dioxide. These layers are sandwiched between two metal contacts. One of the layers of titanium dioxide has the normal ratio of titanium and oxygen atoms, but the other has fewer oxygen atoms than usual. The missing oxygen atoms create vacancies within the material, and these vacancies change its electronic properties. Ordinarily, titanium dioxide is an insulator, blocking the flow of electrons. But oxygen-deficient titanium dioxide readily conducts electrons.
The device begins in the off position, with the layer of insulating titanium dioxide blocking current. But when the researchers apply a positive voltage to one of the metal contacts, the oxygen atoms move from one titanium dioxide layer to the other; as a result, oxygen vacancies exist in both layers. “When that happens, the resistance of the material drops dramatically,” Williams says, allowing current to flow through the device. “We’ve actually built devices where the resistance of the device changes by six orders of magnitude.”
The device can be switched off by applying the opposite voltage, which drives the oxygen back into the first layer. The amount of resistance can be controlled by varying how long the voltage is applied.
The memristor’s electronic properties have led the HP researchers to pursue two potential applications for the device. One is employing it for nonvolatile memory similar to the flash memory used now in digital cameras and cell phones. The speed of memristors suggests that they could be far faster than flash memory and phase-change memory, another type of nonvolatile memory being developed now by Intel and others as a replacement for flash. And the simple devices can be packed densely, potentially allowing memristor chips to store more data than flash memory. Since HP doesn’t make computer chips, the company will likely license the technology to another company.
HP researchers are also making memristor-based chips that mimic the workings of neural networks in the brain. During learning, the connections between neurons in the brain change, some becoming stronger and others weaker over time. The strength of these connections can have a range of different values. It’s possible to simulate this range by making circuits of many transistors. But this can take a large number of transistors. Williams says that it’s possible to do the same thing with just one memristor, since it can be set to have a range of different resistance levels. That could reduce the size of such neural networks by 10,000 times, he says.
Williams envisions a system that uses transistors for the role of neurons, and memristors for the connections between them. Such a memristor-based system would be smaller and use much less energy than one built of transistors alone. Since it would have the ability to adapt in much the same way that brains adapt, it could be good at tasks, such as speech recognition, that are easy for animals but difficult for computers. The HP researchers hope to have prototypes of such chips by next year.
Be there when AI pioneers take center stage at EmTech Digital 2019.Register now