[Paper Review] RxNN: A Framework for Evaluating Deep Neural Networks on Resistive Crossbars

Shubham Jain, Abhronil Sengupta+2 more|arXiv (Cornell University)|Aug 31, 2018

Advanced Memory and Neural Computing4 citations

TL;DR

RxNN is a fast and accurate simulation framework that evaluates large-scale deep neural networks (DNNs) on resistive crossbar hardware by modeling device and circuit non-idealities through a Fast Crossbar Model (FCM). It enables application-level accuracy assessment and re-training, revealing up to 32% accuracy degradation due to non-idealities and demonstrating up to 26% accuracy improvement via re-training.

ABSTRACT

Resistive crossbars designed with non-volatile memory devices have emerged as promising building blocks for Deep Neural Network (DNN) hardware, due to their ability to compactly and efficiently realize vector-matrix multiplication (VMM), the dominant computational kernel in DNNs. However, a key challenge with resistive crossbars is that they suffer from a range of device and circuit level non-idealities such as interconnect parasitics, peripheral circuits, sneak paths, and process variations. These non-idealities can lead to errors in VMMs, eventually degrading the DNN's accuracy. It is therefore critical to study the impact of crossbar non-idealities on the accuracy of large-scale DNNs. However, this is challenging because existing device and circuit models are too slow to use in application-level evaluations. We present RxNN, a fast and accurate simulation framework to evaluate large-scale DNNs on resistive crossbar systems. RxNN splits and maps the computations involved in each DNN layer into crossbar operations, and evaluates them using a Fast Crossbar Model (FCM) that accurately captures the errors arising due to crossbar non-idealities while being four-to-five orders of magnitude faster than circuit simulation. FCM models a crossbar-based VMM operation using three stages - non-linear models for the input and output peripheral circuits (DACs and ADCs), and an equivalent non-ideal conductance matrix for the core crossbar array. We implement RxNN by extending the Caffe machine learning framework and use it to evaluate a suite of six large-scale DNNs developed for the ImageNet Challenge. Our experiments reveal that resistive crossbar non-idealities can lead to significant accuracy degradations (9.6%-32%) for these large-scale DNNs. To the best of our knowledge, this work is the first quantitative evaluation of the accuracy of large-scale DNNs on resistive crossbar based hardware.

Motivation & Objective

• To address the lack of scalable, accurate simulation frameworks for evaluating large-scale DNNs on resistive crossbar-based hardware.
• To quantify the impact of device and circuit non-idealities—such as sneak paths, non-linearities, and process variations—on DNN accuracy in real-world applications.
• To enable model-in-the-loop re-training of DNNs to mitigate accuracy degradation caused by crossbar non-idealities.
• To provide a fast yet accurate alternative to SPICE-level simulations for crossbar-based DNN evaluation at scale.
• To establish a foundation for designing energy-efficient, reliable resistive crossbar systems for large-scale DNN inference and training.

Proposed method

• RxNN extends the Caffe deep learning framework to integrate crossbar-aware inference and training pipelines.
• It decomposes DNN layers into crossbar operations and maps them to resistive crossbar hardware using a Fast Crossbar Model (FCM).
• FCM models vector-matrix multiplication in three stages: non-linear input (DAC), non-ideal conductance matrix (crossbar core), and non-linear output (ADC).
• The FCM captures input-dependent errors from driver resistance, sensing resistance, sneak paths, interconnect parasitics, and stochastic writes.
• FCM achieves 4–5 orders of magnitude speedup over SPICE simulations while maintaining high accuracy in predicting DNN inference errors.
• RxNN supports end-to-end evaluation of large-scale DNNs (e.g., ImageNet models) and enables iterative re-training with crossbar-aware loss functions.

Experimental results

Research questions

• RQ1 To what extent do device and circuit non-idealities in resistive crossbars degrade the inference accuracy of large-scale DNNs?
• RQ2 How do non-idealities such as non-linearities, sneak paths, and process variations affect DNN accuracy across different network depths and architectures?
• RQ3 Can a fast and accurate simulation framework like RxNN enable effective re-training to recover accuracy lost due to crossbar non-idealities?
• RQ4 How does the accuracy degradation vary across different DNN architectures (e.g., AlexNet, VGG-16, ResNet-50) when mapped to resistive crossbars?
• RQ5 What is the trade-off between system energy efficiency and application-level accuracy in resistive crossbar-based DNN accelerators?

Key findings

• Resistive crossbar non-idealities cause significant accuracy degradation in large-scale DNNs, ranging from 9.6% to 32% across six ImageNet models.
• The largest accuracy drop occurs between ideal 6-bit crossbar designs and non-ideal implementations, indicating that non-idealities—not just limited precision—significantly impact performance.
• Re-training DNNs using RxNN recovers up to 26% of the accuracy loss in GoogleNet, 8% in VGG-16, and 9% in AlexNet after only 150 iterations.
• Deeper layers (e.g., Conv3) exhibit more pronounced feature distortions due to accumulated non-idealities, confirming that degradation increases with network depth.
• The Fast Crossbar Model (FCM) achieves 4–5 orders of magnitude speedup over SPICE simulations while accurately modeling input- and weight-dependent errors.
• The study demonstrates, for the first time, a quantitative, application-level evaluation of large-scale DNNs on resistive crossbar systems, revealing critical challenges for real-world deployment.