Stanford study shows magnet-based memory is attractive for edge AI training
A group of researchers from Stanford University School of Engineering introduced a metallic compound as a potential solution for enhancing computer memory and facilitating faster processing. Additionally, their innovative hardware design can enable AI training on edge devices instead of relying on remote servers.
The overarching objective of this research was to minimize energy consumption and reduce the carbon footprint of computing and foster the development of a sustainable computing ecosystem.
“We want to enable AI on edge — training locally on your home computer, phone, or smartwatch — for things like heart attack detection or speech recognition. To do that, you need a very fast, non-volatile memory,” says Shan Wang, the Leland T. Edwards Professor in the School of Engineering at Stanford University.
The study focuses on spin-orbit torque magnetoresistive random access memory (SOT-MRAM), which relies on the phenomenon of electrons possessing a charge and exhibiting rotational motion, effectively transforming them into tiny magnets with polarization along their axes. This unique property of an electron spin reversal, known as magnetic dipole moment, is utilized by researchers to represent the ones and zeroes constituting bits and bytes of computer data.
A significant challenge associated with this approach was identifying a suitable SOT material that would allow for higher data storage density by aligning electron spin directions either upward or downward in the z-direction. In many materials, the flow of current in the x-direction tends to polarize electron spins in the y-direction. However, researchers says that manganese palladium three exhibited the desired properties for generating spin orientation in any direction due to its internal structure that lacks the symmetrical crystal arrangement typically found in other materials.
“Conventional materials only generate spin in the y-direction — that means we would need an external magnetic field to make switching happen in the z-direction, which takes more energy and space,” says Fen Xue, a postdoctoral researcher in Wang’s lab. “For the purpose of lowering the energy and having a higher density of memory, we want to be able to realize this switching without an external magnetic field.”
Scientists advocate for the utilization of this material in computer memories because of its seamless integration with existing manufacturing methods. “There’s no new tools or new techniques needed for this kind of material,” Xue says.
Moreover, the material retains its properties even after undergoing the post-annealing process. Currently, the researchers are focused on developing prototypes of SOT-MRAM using manganese palladium three to be integrated into real commercial devices.
Lanner Electronics, Landing AI develop computer vision solution for quality inspection
AI/ML | edge AI | energy efficiency | ESG | memory | research | Stanford University | training