Researchers have produced graphene memristive synapses that enable high accuracy neuromorphic computing. With record-breaking 16 conductance states in a single device, paired with a non-volatile nature of the synapse, the graphene-based solution hints at energy efficient hardware for neuromorphic computing and for integration of artificial neural networks (ANNs) with emerging technologies such as the Internet of Things. The research, published in Nature Communications, uses Graphenea graphene.

Neuromorphic computing attempts to emulate the neural structure of the human brain. While powerful supercomputers can rival or even exceed the brain in number of operations performed per second, the brain is indisputably superior in terms of energy and area efficiency. Moreover, the brain has the ability to learn by continuously adapting to external stimuli that vary with time. ANNs try to emulate the learning process by modulating synaptic weights assigned to connections between neurons, the brain’s building blocks. An increasingly common way to construct ANNs is from a network of memristors.

Memristive devices are man-made technology that comes the closest to emulating biological brain behaviour. Such devices are electrical resistance or conductance switches that retain a state of internal resistance based on the history of applied voltage and current. A major advantage of resistive memory devices is their ability to support multiple memory states, allowing for a single device to encompass multiple bits of memory and therefore possess a higher data storage density. This, in turn, can lead to the development of smaller, more efficient devices, which are highly advantageous for applications such as the Internet of Things (IoT) and mobile devices capable of utilizing ANNs.

The research, conducted at Penn State University, shows programmable conductance in graphene field effect transistors (GFETs), similar to effects observed before in oxide-based memristors. Compared to state of the art materials, the graphene solution enables more than a single bit, due to multiple conductance states, which brings a real neuromorphic computer a step closer to reality.