Resilience of Digital and Analog RRAM-Based ML Models to Device Variability: A Comparative Study

This paper presents a comprehensive study on the impact of variability in resistive random-access memory (RRAM) devices on the inference accuracy of machine learning (ML) models. Using VGG with CIFAR-10 and ResNet-50 with CIFAR-100 datasets, we performed quantization-aware training to establish baseline accuracies against IEEE FP32 precision. Our results reveal that inference accuracy remains stable within specific variability limits: up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0 \%}$</tex> for ResNet and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5 0 \%}$</tex> for VGG when variability is measured relative to the absolute mean of weights. Beyond these thresholds, accuracy declines rapidly. VGG models, with fewer layers and classes, exhibit greater resilience, tolerating up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$200 \%$</tex> variability, while ResNet models fail between 30-60%. Digital RRAM-based computation shows higher accuracy at low variability levels, whereas analog RRAM-based computation demonstrates superior resilience at higher variability, tolerating up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 0 \%}$</tex> variability compared to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3 0 \%}$</tex> for digital.