Although processors have gotten smaller and faster over time, few computers can compete with the speed and computing power of the human brain. And none comes close to the organ’s energy efficiency. So some engineers want to develop electronics that mimic how the brain computes to build more powerful and efficient devices.
A team at IBM Research, Zurich, now reports that nanosized devices made from phase-change materials can mimic how neurons fire to perform certain calculations (Nat. Nanotechnol. 2016, DOI:10.1038/nnano.2016.70).
This report “shows quite concretely that we can make simple but effective hardware mimics of neurons, which could be made really small and therefore have low operating powers,” says C. David Wright, an electrical engineer at the University of Exeter who wrote a commentary accompanying the new article.
The IBM team’s device imitates how an individual neuron integrates incoming signals from other neurons to determine when it should fire. These input signals change the electrical potential across the neuron’s membrane—some increase it, others decrease it. Once that potential passes a certain threshold, the neuron fires.
Previously, engineers have mimicked this process using combinations of capacitors and silicon transistors, which can be complex and difficult to scale down, Wright explains in his commentary.
In the new work, IBM’s Evangelos Eleftheriou and colleagues demonstrate a potentially simpler system that uses a phase-change material to play the part of a neuron’s membrane potential. The doped chalcogenide Ge2Sb2Te5, which has been tested in conventional memory devices, can exist in two phases: a glassy amorphous state and a crystalline one. Electrical pulses slowly convert the material from amorphous to crystalline, which, in turn, changes its conductance. At a certain level of phase change, the material’s conductance suddenly jumps, and the device fires like a neuron.
The IBM team tested a mushroom-shaped device consisting of a 100-nm-thick layer of the chalcogenide sandwiched between two electrodes. In one demonstration, they used the neuronlike device to detect correlations in 1,000 streams of binary data. Such a calculation could spot trends in social media chatter or even in stock market transactions, Wright says.
He also points out that the devices fire faster than actual neurons, on a nanosecond timescale compared with a millisecond one. The neuron mimics, Wright says, are another step toward hardware that can process information as the brain does but at speeds orders of magnitudes faster than the organ. “That could do some remarkable things.”