Spintronics for Neuromorphic Engineering
299
with modern day computational algorithms and hardware [7]. As a result, braininspired computing using artificial neural networks such as depicted in Fig. 1c have
become the de facto approach for classification tasks such as computer vision and
speech recognition.
For instance, DeepMind’s AlphaZero is an artificial intelligence (AI) designed to
master the games of chess, shogi, and go [8]. Rather than performing computation
using consumer desktop processors, AlphaZero relies on optimized and purpose-built
application-specific integrated circuits (ASIC). Over 5000 units of Google’s proprietary tensor processing units (TPU) are used in AlphaZero, with each TPU estimated
to consume around 200 W [9]. Each TPU alone far exceeds the power utilization of
the biological brain of 20 W. Neuromorphic chips have been developed to function as
artificial synapses and neurons using conventional CMOS transistors. For instance,
IBM’s TrueNorth utilizes 5.4 billion transistors across 1 million programmable
neurons and 256 million configurable synapses, and operates at comparatively lower
power than general purpose microprocessors [10, 11]. These brain-inspired neuromorphic engineering and computing models are still implemented using conventional transistor circuits. Having a deeper understanding of the biological brain
and emulating brain processes and attributes allow for the development of highly
efficient brain-inspired computing technologies. The prospects of beyond CMOS
technologies are also promising.
Emerging memory technologies such as spintronics offer a promising prospect for
fulfilling such needs and applications. In utilizing spintronic device physics, one can
engineer neuromorphic primitives that mimic the functionality of biological neurons,
synapses, and processes in a top-down approach. Spintronic devices are inherently
non-volatile, high-speed, radiation-hard, have zero static power requirements, practically infinite endurance, and can thus overcome the drawbacks inherent in transistorbased devices. Spintronic devices have also rapidly gained technological maturity in
the form of magnetic random access memory (MRAM) that use magnetic tunnel junctions (MTJ) for data storage applications. In these MRAM devices, MTJs are deterministically switched to store binary information. In research, spintronic devices that
rely on the motion of domain walls (DW) to achieve multi-state capabilities as well as
in devices that operate in the stochastic regime have been demonstrated. These characteristics have opened opportunities and possibilities in developing brain-inspired
functional primitives such as artificial neurons and synapses.
2 Fundamentals of Spintronic Device and Physical
Phenomena
In addition to the fundamental charge property of electrons, spintronic devices also
make use of the intrinsic electron spin and its associated magnetic moment. In the
absence of external influence such as applied magnetic fields and spin currents, a
299
with modern day computational algorithms and hardware [7]. As a result, braininspired computing using artificial neural networks such as depicted in Fig. 1c have
become the de facto approach for classification tasks such as computer vision and
speech recognition.
For instance, DeepMind’s AlphaZero is an artificial intelligence (AI) designed to
master the games of chess, shogi, and go [8]. Rather than performing computation
using consumer desktop processors, AlphaZero relies on optimized and purpose-built
application-specific integrated circuits (ASIC). Over 5000 units of Google’s proprietary tensor processing units (TPU) are used in AlphaZero, with each TPU estimated
to consume around 200 W [9]. Each TPU alone far exceeds the power utilization of
the biological brain of 20 W. Neuromorphic chips have been developed to function as
artificial synapses and neurons using conventional CMOS transistors. For instance,
IBM’s TrueNorth utilizes 5.4 billion transistors across 1 million programmable
neurons and 256 million configurable synapses, and operates at comparatively lower
power than general purpose microprocessors [10, 11]. These brain-inspired neuromorphic engineering and computing models are still implemented using conventional transistor circuits. Having a deeper understanding of the biological brain
and emulating brain processes and attributes allow for the development of highly
efficient brain-inspired computing technologies. The prospects of beyond CMOS
technologies are also promising.
Emerging memory technologies such as spintronics offer a promising prospect for
fulfilling such needs and applications. In utilizing spintronic device physics, one can
engineer neuromorphic primitives that mimic the functionality of biological neurons,
synapses, and processes in a top-down approach. Spintronic devices are inherently
non-volatile, high-speed, radiation-hard, have zero static power requirements, practically infinite endurance, and can thus overcome the drawbacks inherent in transistorbased devices. Spintronic devices have also rapidly gained technological maturity in
the form of magnetic random access memory (MRAM) that use magnetic tunnel junctions (MTJ) for data storage applications. In these MRAM devices, MTJs are deterministically switched to store binary information. In research, spintronic devices that
rely on the motion of domain walls (DW) to achieve multi-state capabilities as well as
in devices that operate in the stochastic regime have been demonstrated. These characteristics have opened opportunities and possibilities in developing brain-inspired
functional primitives such as artificial neurons and synapses.
2 Fundamentals of Spintronic Device and Physical
Phenomena
In addition to the fundamental charge property of electrons, spintronic devices also
make use of the intrinsic electron spin and its associated magnetic moment. In the
absence of external influence such as applied magnetic fields and spin currents, a