Magnetic tunnel junction devices and circuits for in-memory, neuromorphic and radiation hard computing




Alamdar, Mahshid

Journal Title

Journal ISSN

Volume Title



The magnetic tunnel junction is a memory device at the core of emerging magnetic random access memory technology. As CMOS technology is approaching its physical limits, spintronics, with benefits like non-volatility and normally-off behavior, is a promising candidate for next-generation artificial intelligence computing applications. In this dissertation, I show prototypes of magnetic tunnel junction in-memory computing devices that use spin-orbit torque switching, with device tunnel magnetoresistance up to 203%, close to the expected highest 200% seen in this type of magnetic tunnel junctions, and average resistance-area product close to the unpatterned film. Device cycle-to-cycle variation in switching voltage curtailed to 10% by controlling the domain wall initial position, which corresponds to 90% accuracy in a domain wall-magnetic tunnel junction full adder simulation. Repeatability of writing and resetting the device and an inverter circuit of two devices are also shown in this work. Another promising feature of magnetic tunnel junctions is their intrinsic hardness to radiation, which makes them an attractive candidate to be used in high radiation environments. We investigate the radiation hardness of perpendicular magnetic anisotropy spin-orbit torque magnetic tunnel junctions to different types of irradiations, such as gamma and heavy-ion irradiation. We observe gamma ionizing dose up to 1 Mrad(Si) does not alter the magnetic switching behavior of the film. Then we examine the effects of heavy ion irradiation through Ta¹⁺ bombardment and realize high enough heavy ion fluence will cause displacement damage and intermixing, mostly at the bottom layers of the stack, which degrades the magnetic properties of the device. We propose a multi-weight synapse using a notched magnetic tunnel junction for neuromorphic computing. We observe the high stability of the resistance states and confirm it by implementing it in a CIFAR-100 inference task and checking the accuracy. Employing the multi-weight switching capability of the proposed synapse, we design an analog content addressable memory based on magnetic tunnel junctions. Some future directions are proposed to optimize these devices for more separated and bigger range bounds that is desired for this application. These results make strides in using magnetic tunnel junctions for in-memory, neuromorphic and radiation hard computing.


LCSH Subject Headings