Analytical Model for Parasitic Resistances of Crossbar Arrays Suitable for Open-loop Programming Schemes Reliability Analysis
Authors:
- 6 October 2024
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Abstract
The stringent demands for area efficiency and low power consumption in edge devices have driven the investigation of open-loop programming schemes for In-Memory Computing (IMC) systems, which show promise for further improving energy efficiency in Artificial Intelligence (AI) applications. However, the adoption of these programming schemes requires extensive reliability analysis to ensure that signal degradation along metallization lines does not prevent the devices from reaching their target states. Unfortunately, accurate circuit simulations are impractical for AI applications due to the large number of crossbars involved, leading to infeasible simulation times. Additionally, existing analytical models of crossbar parasitic effects primarily focus on output signal degradation, overlooking the voltage drop at each cell. In this work, we modify a wellknown analytical model for parasitic resistances in resistive crossbars to provide a more accurate estimation of voltage levels at internal array nodes. The proposed model reduces the average error in estimating the voltage drop across cells during programming operations by up to $30 \%$ compared to the previous approach, while offering a significant speed-up in simulation time w.r.t. traditional SPICE analysis. Finally, we apply the model to assess two open-loop programming schemes for crossbars up to a $128 \times 128$ array, revealing reliability issues that could cause malfunctions in the IMC system if not addressed during the design phase.
Keywords
Funding Information
- Deutsche Forschungsgemeinschaft
This publication has 19 references indexed in Scilit:
- Open-loop analog programmable electrochemical memory arrayNature Communications, 2023
- MemTorch: An Open-source Simulation Framework for Memristive Deep Learning SystemsNeurocomputing, 2022
- A Flexible and Fast PyTorch Toolkit for Simulating Training and Inference on Analog Crossbar Arrays2021
- Analog architectures for neural network acceleration based on non-volatile memoryApplied Physics Reviews, 2020
- Circuit Reliability Analysis of RRAM-based Logic-in-Memory Crossbar Architectures Including Line Parasitic Effects, Variability, and Random Telegraph Noise2020
- Reinforcement learning with analogue memristor arraysNature Electronics, 2019
- Efficient evaluation model including interconnect resistance effect for large scale RRAM crossbar array matrix computingScience China Information Sciences, 2018
- The Next Generation of Deep Learning Hardware: Analog ComputingProceedings of the IEEE, 2018
- Technological Exploration of RRAM Crossbar Array for Matrix-Vector MultiplicationJournal of Computer Science and Technology, 2016
- Scaling-up resistive synaptic arrays for neuro-inspired architecture: Challenges and prospect2015