Â鶹ÒùÔº

New AI hardware on the horizon thanks to electrically programmable spintronic device

AI transformational impact is well under way. But as AI technologies develop, so too does their power consumption. Further advancements will require AI chips that can process AI calculations with high energy efficiency. This is where spintronic devices enter the equation. Their integrated memory and computing capabilities mimic the human brain, and they can serve as a building block for lower-power AI chips.

Now, researchers at Tohoku University, National Institute for Materials Science, and Japan Atomic Energy Agency have developed a new spintronic device that allows for the electrical mutual control of non-collinear antiferromagnets and ferromagnets. This means the device can switch magnetic states efficiently, storing and processing information with less energy—just like a brain-like AI chip.

The breakthrough can potentially revolutionize AI hardware via high efficiency and low energy costs. The team their results in Nature Communications on February 5, 2025.

"While spintronic research has made significant strides in controlling magnetic order electrically, most existing spintronic devices separate the role of the magnetic material to be controlled and the material providing the driving force," says Tohoku University's Shunsuke Fukami, who supervised the research.

These devices have a fixed operation scheme once fabricated, typically switching information from "0" to "1" in a binary fashion. However, the new research team's breakthrough offers a major innovation in electrically programmable switching of multiple magnetic states.

Fukami and his colleagues employed the non-collinear antiferromagnet Mn₃Sn as the core magnetic material. By applying an electrical current, Mn₃Sn generates a spin current that drives the switching of a neighboring ferromagnet, CoFeB, through a process known as the magnetic spin Hall effect.

Not only does the ferromagnet respond to the spin-polarized current, but it also influences the magnetic state of Mn₃Sn, enabling the electrical mutual switching between the two materials.

In their proof-of-concept experiment, the team demonstrated that information written to the ferromagnet can be electrically controlled via the magnetic state of Mn₃Sn. By adjusting the set current, they were able to switch the magnetization of CoFeB in different traces representing multiple states.

This analog switching mechanism, where the polarity of the current can change the sign of the information written, is a key operation in neural networks, mimicking the way synaptic weights (analog values) function in AI processing.

"This discovery represents an important step toward the development of more energy-efficient AI chips. By realizing the electrical mutual switching between a non-collinear antiferromagnet and a ferromagnet, we have opened new possibilities for current-programmable neural networks," said Fukami. "We are now focusing on further reducing operating currents and increasing readout signals, which will be crucial for practical applications in AI chips."

The team's research opens new pathways for improving the energy efficiency of AI chips and minimizing their environmental impacts.