Beyond von–Neumann Computing—Engineering Artificial Synapses for In–Memory Computing

Abstract

The processing of an input dataset by a deep neural network, e.g. for classification tasks, is a series of vector-matrix multiplications (input data * synapse weight). The size and number of matrices that need to be processed are too large to be stored in a Static Random Access Memory (SRAM) unit. For this reason, data need to be shuffled constantly between memory and processing unit in a classical computing architecture, leading to inefficiency and performance mitigation. Dedicated neuromorphic computing hardware may help to overcome these issues and is a promising technology for the post– von Neumann computing era. Especially memristive devices are suitable for this dedicated hardware. A crossbar of memristive devices stores the synaptic weights in a physical matrix. The weight can be read non-destructively and updated e.g. during training. The matrix vector multiplication can be performed in-memory by coding the synaptic weight to the device conductance and the input data to the read voltage amplitude. Several memristive technologies are under study for the implementation of neuromorphic hardware, and they rely on diverse physical mechanisms and materials [1]. However, improvements in the device characteristics are required for optimal hardware acceleration in both performance precision and energy consumption. Examples of challenges are the minimization of inter and intra device asymmetricity, non-linearity and stochasticity, and the increase of the dynamic range. In this talk I will present Phase Change Memories (PCM) and filamentary-based Oxide Resistive RAM (OxReRAM), the two most promising candidates to represent artificial synapses in neuromorphic hardware [2,3]. In PCM-based memristors, the conductance change relies on a phase transition in the material, while for filamentary-based OxReRAM, the synaptic weight depends on the rupture/formation of oxygen vacancies based conductive paths. Strength and challenges of the two classes of memristors will be discussed. For each technology, an innovative design and material stack concepts will be presented, demonstrating enhanced operational characteristics. The material-stack characterization by means of X-ray reflectivity and diffraction, Hall-measurements as well as the electrical characterization of the devices will be covered. A discussion on the ideal class of applications for a more efficient use of PCM- or OxReRAM-based synapses will follow. [1] Rajendran, B. & Alibart, IEEE J. Emerg. Sel. Top. Circuits Syst. 6, 198–211 (2016) [2] Abu Sebastian et al., Jour. Of Phys. D, 52, 44 (2019) [3] B. Govoreanu, et al., Proc. IEEE Int. Electron Devices Meeting, 31.6.1–31.6.4 (2011)