IBM Corp. is on a mission to redefine computing. The scientists at its T.J. Watson Research Center in New York are building quantum processors, while their colleagues at the San Jose lab is pioneering neuromorphic chip design. And now the company’s Zurich-based semiconductor unit is marking a breakthrough of its own that could take the effort in an entirely new direction.

As detailed in today’s issue of Nature Nanotechnology, the group has managed to show for the first time that artificial neurons can be created using so-called phase-change materials. These compounds are characterized by their ability to quickly switch between a disorganized amorphous state and a crystalline state in response to electric simulation. Each chemical configuration in turn corresponds to a different value, which is similar to how media content is recorded on Blueray discs. The difference is that Big Blue’s implementation doesn’t keep information in a digital format but rather an analog arrangement more akin to volatile memory. And like biological brains, the technology can not only ingest information but also carry out calculations at the same time.

This property gives IBM’s artificial neuron implementation several major advantages over traditional chips with separate processing and memory components. The biggest benefit is that the technology removes the need to shuffle data back and forth between different parts of the semiconductor, which saves a tremendous amount of power. And second, it avoids the delay involved in the transfer operation. The resulting time savings can provide a substantial speed improvement for certain workloads, particularly those that use artificial intelligence under the hood.

The neuromorphic chip architecture that is undergoing development at IBM Corp’s San Jose lab is based on similar principles, and it’s already showing impressive performance. An early silicon prototype that the company demonstrated to the press in 2014 was reportedly able to process video footage “100 times faster” than a laptop using just 0.00001 percent the power. Since the company’s Zurich unit is breaking new ground with its phase-shift technology, a similar proof-of-concept is not yet available, but the early experiment that landed the team in Nature Nanotechnology looks very promising.

The scientists organized a few hundred of their artificial neurons into a makeshift processor (pictured above) that IBM says racked up billions of switching cycles during trials. Its output amounted to “multiple years of operation at update frequency of 100 Hz” according to the company. That’s nowhere near what commercial chips are capable of, but given the project’s small size and early stage, it’s a major feat nonetheless. The results are made especially notable by the fact that updating each neuron took a mere 120 microwatts on average while a standard LED lightbulb consumes some 60 million.

Such research is costly, but experimenting with both silicon and phase-shift neuromorphic chips allows IBM to hedge its bets. The company should in time be able to identify which approach has a higher likelihood of yielding a commercially viable product and refocus its efforts accordingly while still retaining sizable a head-start over the competition. Big Blue envisions the technology being implemented in server chips to speed up artificial intelligence algorithms, real-time analytics software and other latency-sensitive workloads. Such an addition could give its POWER processor series a much-needed advantage over the Intel Xeon CPUs that currently dominate the data center market.

Image via IBM