This book describes an extensive and consistent soft error assessment of convolutional neural network (CNN) models from different domains through more than 14.8 million fault injections, considering different precision bit-width configurations, optimization parameters, and processor models. The authors also evaluate the relative performance, memory utilization, and soft error reliability trade-offs analysis of different CNN models considering a compiler-based technique w.r.t. traditional redundancy approaches.
Author(s): Geancarlo Abich, Luciano Ost, Ricardo Reis
Series: Synthesis Lectures on Engineering, Science, and Technology
Publisher: Springer
Year: 2023
Language: English
Pages: 142
City: Cham
Preface
Acknowledgements
Contents
Acronyms
1 Introduction
[DELETE]
1.1 Hypothesis to Be Demonstrated in This Book
1.2 Book Goal
1.3 Original Contributions of This Book
1.3.1 Evaluation of SOFIA Consistency w.r.t. RTL
1.3.2 Early Soft Error Assessment of ML Models Executing on Resource-Constrained IoT Edge Devices
1.3.3 Legacy of Tools Integrated into the SOFIA Framework
1.4 Book Outline
2 Background in ML Models and Radiation Effects
[DELETE]
2.1 Basic Concepts of ML Models
2.1.1 Unsupervised Learning
2.1.2 Supervised Learning
2.1.3 ML for the IoT Edge Devices
2.2 Radiation Environment an its Effects on Semiconductors Devices
2.2.1 Classification of SEEs
3 Related Works
[DELETE]
3.1 Soft Error Assesment Considering Virtual Platforms
3.1.1 Contribution in Soft Error Assessment Considering Virtual Platforms
3.2 Soft Error Reliability Assessment of Machine Learning Techniques
3.2.1 Review of System-Level Soft Error Mitigation Techniques
3.2.2 Contribution in Machine Learning Soft Error Assessment and Mitigation
4 Soft Error Assessment Methodology
[DELETE]
4.1 Fault Injection Frameworks
4.1.1 RTL Fault Injection Module
4.1.2 SOFIA Framework
4.2 Fault Classification
4.3 System Reliability Assessment Metrics
4.3.1 Supported Soft Error Mitigation Techniques
5 Early Soft Error Consistency Assessment
[DELETE]
5.1 Soft Error Consistency Assessment for Single-Core Processors
5.1.1 Experimental Setup
5.1.2 FI Simulation Performance of SOFIA w.r.t. RTL
5.1.3 Soft Error Reliability Mismatch
5.1.4 Closing Remarks
6 Soft Error Reliability Assessment of ML Inference Models Executing on Resource-Constrained IoT Edge Devices
[DELETE]
6.1 Soft Error Reliability Assessment of the CIFAR-10 CNN
6.1.1 CIFAR-10 CNN Developed with CMSIS-NN
6.1.2 Soft Error Reliability Assessment of CIFAR-10 CNN Execution on Resource-Constrained IoT Devices
6.1.3 The Impact of Soft Errors in Memory Units of Edge Devices Executing CIFAR-10 CNN
6.1.4 Impact of Thread Parallelism on the Soft Error Reliability of CIFAR-10 CNN
6.2 Soft Error Reliability Assessment of the MobileNet CNN
6.2.1 MobileNet CNN Developed with CMix-NN
6.2.2 The Impact of Precision Bitwidth on the Soft Error Reliability of the MobileNet CNN Execution on Resource-Constrained IoT Devices
6.2.3 The Impact of Soft Errors in Memory Units of Edge Devices Executing the MobileNet CNN
6.2.4 Applying Lightweight Soft Error Mitigation Techniques to Embedded Mixed Precision Deep Neural Networks
7 Conclusions and Future Work
[DELETE]
