Publication
MEMRISYS 2023
Talk

Analog Conductive Metal Oxide-HfO2 ReRAM artificial synapses for neuromorphic computing: physical modelling and stack optimization

Abstract

The energy required for the training and inference of modern neural networks on conventional von- Neumann computing architecture has been growing tremendously. In particular, the data transferring between the processing and memory units, physically separated by a common bus, results in energy and performance bottleneck. This has led to a tremendous effort to develop more energy-efficient AI hardware, such as specialized neuromorphic chips that mimic the behavior and/or the efficiency of biological neurons. In-memory brain-inspired neuromorphic architectures based on non-volatile memory elements, known as memristors, are a promising solution in this regard. When organized in crossbar arrays, memristive technologies such as Redox-based Resistive Switching Random Access Memories (ReRAMs) or Phase-Change Memories (PCMs) [1] can be used to perform vector-matrix multiplications (VMMs), the most energy-expensive and recurrent operation in AI’s tasks, in the analog domain, by exploiting Ohm’s and Kirchhoff’s law. Such technologies are scalable and can be densely integrated in the Back End of the Line (BEOL) of existing CMOS technology, to create a high performing synaptic analog signal processing accelerator. In this work, we focus on an emerging innovative concept of ReRAM based on an engineered Conductive Metal Oxide (CMO) layer [2]. The proposed Metal/CMO/Insulator/Metal filamentary ReRAM device shows superior characteristics such as gradual, linear, and symmetric conductance update, large number of states, good retention, and reproducibility of the switching characteristics with respect to conventional Metal/Insulator/Metal ReRAM type. The device optimization process and the material-stack characterization by means of Transmission Electron Microscopy and Electron Dispersive Spectroscopy, as well as the device electrical characterization will be discussed. In addition, through impedance spectroscopy experiment, an equivalent electrical circuit model of CMO-HfO2 ReRAM is extracted, enabling further system-level simulations of this technology. Finally, a physical understanding of the optimized CMO-HfO2 ReRAM in pristine state, after forming, in the high resistive state (HRS) and in low resistive state (LRS), explaining the role of the CMO layer in the resistive switching behavior, will be presented. The analog switching properties and CMOS and BEOL compatibility of the CMO-HfO2 ReRAM devices are promising for large-scale integration into crossbar arrays for neural network training and inference. References [1] Ielmini, D., Wong, HS.P., in. Nat Electron 1, 333–343 (2018). https://doi.org/10.1038/s41928-018-0092-2 [2] Stecconi T., et al. (2022), in Advanced Electronic Materials. 8. 10.1002/aelm.202200448.