Thought Leadership
The growing energy appetite of AI applications is forcing researchers to rethink computing from the ground up. Spintronic devices could form the foundation of a new generation of energy-efficient computer architectures.
ChatGPT, image generators, optimization algorithms: the AI boom comes with a downside that rarely makes the headlines. Data centers around the world are consuming ever more electricity. Conventional semiconductor technology is increasingly bumping up against its physical limits. An international team of researchers, including University of Duisburg-Essen physicist Prof. Karin Everschor-Sitte, has now outlined in a review article published in Nature Reviews Physics how spintronic devices could fundamentally change the way we compute.
The problem: transistors at their limit
Conventional computers translate all information into binary sequences of zeros and ones. This principle has worked reliably for decades, but the exponentially growing data volumes of modern AI applications are turning it into a bottleneck. More computing power requires more transistors, and more transistors mean more waste heat and higher energy consumption. It is a vicious cycle that the industry has so far tried to break mainly through larger cooling systems and more efficient chip designs.
The way out: electrons have more to offer than their charge
Spintronics takes a fundamentally different approach. Rather than relying solely on the electrical charge of electrons, as conventional semiconductors do, it exploits a second physical property: the spin. In simplified terms, spin can be thought of as a kind of magnetic angular momentum of the electron. This additional degree of freedom opens up possibilities that go far beyond what classical transistors can achieve.
Magnetic materials bring several properties to the table that are highly attractive for novel computing approaches. They store information in a nonvolatile manner, respond extremely fast and exhibit complex dynamic processes including nonlinearity, controlled randomness and temporal feedback. These are all characteristics that classical transistors can only replicate through elaborate circuitry.
Magnetic neurons and probabilistic bits
The review article catalogues the key building blocks that can be realized from spintronic materials. These include artificial neurons and synapses based on spin, so called p-bits (probabilistic bits that deliberately work with uncertainties) as well as more complex architectures such as magnetic reservoir computing and Ising machines. The latter are considered particularly promising for combinatorial optimization problems that push classical computers to their limits.
“We are specifically researching how reservoir computing can be implemented using magnetic structures known as skyrmions,” explains Everschor-Sitte. Skyrmions are tiny, vortex-like magnetic structures that form in thin layers of certain materials and can serve as information carriers. “An important part of our work is also to develop new metrics that allow the performance of such systems to be reliably evaluated.”
Compatible with existing manufacturing
A decisive advantage over many other future technologies: spintronic devices do not have to start from scratch. Magnetic tunnel junctions, one of the core elements of the technology, are already used in commercial memory products such as MRAM chips. They can be integrated into existing CMOS manufacturing processes, which could significantly shorten the path from the laboratory to mass production.
Not a replacement, but a complement
Despite all the progress, the authors caution against inflated expectations. The optimal tuning of materials, devices and algorithms remains an open challenge. There is also a lack of standardized benchmarks to meaningfully compare the performance of spintronic hardware with established systems.
In the long run, the researchers see spin-based technologies not as a replacement for classical computer architectures but as a complement. The future is likely to belong to hybrid systems that combine different physical computing principles depending on the task at hand. When it comes to handling the growing computational demands of data-driven workloads in an energy-efficient way, the electron’s spin could turn out to be just the right twist.
