Riccardo Bertacco1
Politecnico di Milano1
So far, edge computing has been implemented by exploiting conventional Boolean computing in microsystems. This solution is largely inefficient due to the following reasons: (i) Joule dissipation and latency due to continuous data transfer, (ii) the usage of general-purpose digital platforms for patterning recognition requires high-performing computing platforms, such as GPUs. This implies that edge computing should adopt a different technology platform instead of CMOS. CMOS is a powerful technology platform for binary logic, but it is not designed for emerging unconventional computing paradigms, such as neuromorphic computing.<br/>Spintronic architectures can make a difference in this field essentially thanks to these crucial features: (i) non-volatility and plasticity of the ferroic order parameters (magnetization in ferromagnets, ferroelectric polarization in ferroelectric systems), (ii) logic-in-memory functionalities offered by the ferroic control of low-dissipation spin currents, (iii) foreseen CMOS compatibility, (iv) much lower operating voltages (1V or below) than PCRAM or RRAM technologies.<br/>In this context, we are developing ultra-low power spin-computing devices for Edge Artificial Intelligence (AI), exploiting ferroic order (ferroelectricity) as a state variable, and enhanced spin- charge interconversion in Ferroelectric Rashba Semiconductors (FERSC) to manipulate such state variable.<br/>FERSC have been disclosed in 2013 with reference to germanium telluride (GeTe) (1), a CMOS compatible phase-change semiconductor featuring an intriguing property: the ferroelectric control of the spin texture and spin-charge interconversion. We are exploiting these unique features in FESO memtransistors in which the nonvolatile ferroelectric configuration (state variable) controls the transconductance of the device (computing functionality) through ferroelectric-dependent spin-to-charge conversion.<br/>In this paper we focus on the application of FERSC materials to memcomputing by benchmarking a neural network based on memtransistors (2) to current implementations on a microcontroller. As specific application we analyze the case of binary classification for human activity recognition based on data from inertial motion units. An energy saving by a factor 100 is estimated for a perceptron with 6 neurons and 18 synapses.<br/> <br/>(1) D. Di Sante, P. Barone, R. Bertacco, S. Picozzi, Advanced Materials. 25, 509–513 (2013).<br/>(2) Varotto, S. <i>et al.</i> Room-temperature ferroelectric switching of spin-to-charge conversion in germanium telluride. <i>Nat Electron</i> <b>4, </b>740–747 (2021)