Neuromorphic Spintronics

TL;DR

Neuromorphic Spintronics surveys how spintronic materials can enable brain-inspired, energy-efficient computing beyond traditional CMOS by integrating memory, computation, and adaptability. It outlines four pathways: computation based on fluctuations (stochastic, inverse, and token-based Brownian computing), spintronics-based neural networks (neural/synaptic building blocks and spintronic realizations), reservoir computing with magnetic textures, and spintronic memory technologies (MRAM and beyond). Key contributions include detailing hardware primitives (MTJs, skyrmions, domain walls, STT/SOT), evaluating 2D and emerging 3D textures, and discussing multi-physics and organic spintronics as routes to higher density and efficiency. The work emphasizes potential for in-memory, low-energy AI hardware and identifies challenges in fabrication, integration, and scalable performance, with a forward-looking view on 3D architectures, multiferroics, and edge implementations. In mathematical terms, spintronic neural computation can realize forward passes such as $a_j = f\left(\sum_i w_{ij} x_i\right)$, while stochastic and inverse schemes leverage probabilistic representations and energy landscapes to trade accuracy for efficiency.

Abstract

Neuromorphic spintronics combines two advanced fields in technology, neuromorphic computing and spintronics, to create brain-inspired, efficient computing systems that leverage the unique properties of the electron's spin. In this book chapter, we first introduce both fields - neuromorphic computing and spintronics and then make a case for neuromorphic spintronics. We discuss concrete examples of neuromorphic spintronics, including computing based on fluctuations, artificial neural networks, and reservoir computing, highlighting their potential to revolutionize computational efficiency and functionality.

Figures (5)

• Figure 1: Neuromorphic computing with spintronics-based devices offers many characteristics such as memory and adaptability, scalability, low-energy computation, and error tolerance.
• Figure 2: Illustration of the operating principle of stochastic computing using the example of multiplication of two numbers $p_1$ and $p_2$. The (in general approximate) result of the multiplication $p_3$ is obtained by connecting uncorrelated random bitstream representations of $p_1$ and $p_2$ using an AND gate.
• Figure 3: Illustration of inverse computing exploiting p-bits. (a) Schematic of the general idea behind inverse computing showing an example of an AND gate. The inverse for the output "0" is not unique. Three different inputs (00, 01, 10) correspond to the output "0". An ideal inverse computing scheme utilizing fluctuations yields all three input possibilities with equal probability. (b) Key characteristics of a p-bit: the bistable system has two energy ($E$) minima at states representing 0 and 1 separated by an energy barrier ($\Delta E$), which is of the order of $k_B T$, where $T$ is the ambient temperature. Under zero bias (blue), the p-bit fluctuates equally between the two states. Under positive/negative bias (brown/green), the energy landscape is tilted, and the 1/0 state is favored, thus changing the represented probability value. (c) Operation of a p-bit with nanomagnets: a majority gate setup can be used to implement an AND gate using p-bits. When the output bit C is fixed (indicated by the light red pin) to the 0 state, the A and B bits fluctuate between the three possible states. The jagged lines for the bits indicate that they can toggle between the two states, 0 and 1.
• Figure 4: Illustration of the basic principles of token-based Brownian computing. The tokens (cyan-colored circles) perform a random motion along the wires (black) and are influenced by the circuit elements (orange) according to their functions. a) One basic set of primitive elements consists of a Hub (circle) and a C-join (Square). The Hub is a triple junction that allows the token to move freely along any of the three connecting wires. The C-join is a four-way intersection that requires two tokens to enter the element from different directions, and it releases them along the other two directions. b) Implementation of a half-adder circuit using hubs and C-joins. In the specific example, the calculation of the addition of two bits (Input A and B) with the value 1 is shown. The wires marked in (light) blue indicate along which wires the tokens are allowed to move (before) after passing the C-join. Adding two bits with the value 1 thus leads to the desired result of carry 1 and sum 0, which is where the tokens exit the circuit and can be detected. The C-join that mainly contributes to this operation is marked with a dotted square.
• Figure 5: Illustration depicting various physical operational and readout modes of reservoir computing. The input can be supplied to the reservoir via different physical quantities like voltage, light, magnetic field, temperature, and strain. A plethora of different spintronics systems can play the role of the reservoir. The reservoir states can be read by measuring different physical quantities like temperature, resistance, or magnetization. Furthermore, the mode of reading out the reservoir state could be spatial or temporal. (Figure adapted from figures in everschor2024topologicalprychynenko2018magneticbourianoff2018potentialzheng2018experimental). Hopfion figure created by Ross Knapman.