A Universal, Analog, In-Memory Computing Primitive for Linear Algebra Using Memristors

The increasing demand for data-intensive computing applications, such as artificial intelligence (AI) and more specifically machine learning (ML), raises the need novel hardware architectures capable of massive parallelism in performing core algebraic operations. Among new paradigms, in-memory (IMC) with analogue devices is attracting significant interest its large-scale integration potential, together unrivaled speed energy performance. Here, we present a fully-analogue, universal primitive executing linear algebra operations regression, generalized least-square minimization system solution without preconditioning. We study impact main circuit parameters on accuracy bandwidth analytical closed-form expressions SPICE simulations. Scaling challenges due to parasitic resistance/capacitance their key are discussed. Finally, comparison existing solvers belonging same IMC framework made assess advantages disadvantages proposed circuit.