A materials- and devices-centric approach to neuromorphic computing

The fundamental physics of spintronic devices makes them an excellent technology for realizing neuromorphic computing hardware. At the heart of all spintronic devices is a magnetic material that can be driven out of equilibrium with spin torques and host dynamics, such as phase transitions and criticality, self-oscillations, synchronization, stochastic resonance, and chaos, which form the basis of many brain algorithms. This paper highlights the potential of magnetic tunnel junctions for realizing the key building blocks, including on-chip non-volatile memory, probabilistic bits, stochastic oscillators, coherent oscillators, of echo-state networks, spiking neural networks, and population coding neural systems for edge AI applications. Antiferromagnetic spintronics offers a new paradigm to exploit a plethora of antiferromagnetic materials for energy-efficient computing applications. We present calculations of device performance, energy dissipation, and offer insights into technology-device co-optimization. System-level calculations of the performance of a spintronics-based spiking neural network for practical workloads that uses spiking antiferromagnetic neurons with auto-reset functionality, ferromagnetic synapses, and electrical interconnects are discussed. Our calculations reveal that the energy consumption of the network is limited by the ferromagnetic synapses and electrical interconnects, while the latency is dominated by neurons. Yet, spintronics spiking networks perform at a fraction of energy and latency cost compared to CMOS-only solutions, while spintronics systems are also much more area efficient. The paper concludes by summarizing the limits, challenges, and opportunities of spintronics for edge AI applications.