New Mathematical Framework Could Cut AI Memory's Energy Use by Thousands of Times

Researchers at the University of Edinburgh have developed a mathematical framework designed to sharply cut the energy needed to write information in future magnetic memory, a potential way to shrink the growing power footprint of artificial intelligence. The work, led by Dr. Elton Santos of the university's Institute for Condensed Matter Physics and Complex Systems, is published in the journal Advanced Materials.

Magnetic memory stores information by controlling the orientation of tiny magnetic regions, switching between two states to represent a digital "0" or "1." Conventional magnetic field pulses used to make that switch are often inefficient, consuming substantial energy and struggling to target very small regions as devices shrink. The Edinburgh team applied optimal control theory, a mathematical method for finding the most efficient route to a result, to calculate how a magnetic field should change moment to moment to flip a bit using minimal energy, rather than applying a simple fixed pulse.

In computer simulations of three ultrathin materials known as van der Waals magnets — Fe3GaTe2, Fe3GeTe2 and CrSBr — the optimized pulses reversed magnetization in about 1 to 10 picoseconds (a picosecond is one trillionth of a second), using fields more than 10 times weaker than standard methods. Switching required as little as 0.94 nanojoules in the simulations, compared with as much as 91.2 nanojoules for conventional field pulses. By adjusting properties that govern how the material's magnetic spins move, the researchers calculate switching energy could eventually fall into the femtojoule range, a quadrillionth of a joule.

The researchers say the broader modeling indicates optimized switching could reduce energy use by several orders of magnitude compared with established memory technologies, moving closer to the Landauer limit, the fundamental physical minimum energy required to erase a bit of information. Santos said the same mathematical approach could also be adapted to electrical currents and ultrafast laser pulses being explored for future memory and spintronics devices.