Abstract:Systolic array has emerged as a prominent architecture for Deep Neural Network (DNN) hardware accelerators, providing high-throughput and low-latency performance essential for deploying DNNs across diverse applications. However, when used in safety-critical applications, reliability assessment is mandatory to guarantee the correct behavior of DNN accelerators. While fault injection stands out as a well-established practical and robust method for reliability assessment, it is still a very time-consuming process. This paper addresses the time efficiency issue by introducing a novel hierarchical software-based hardware-aware fault injection strategy tailored for systolic array-based DNN accelerators.

What problem does this paper attempt to address?

This paper attempts to address the problem of efficiently evaluating the reliability of deep neural network (DNN) hardware accelerators based on systolic arrays. Specifically, the paper focuses on how to quickly and accurately perform fault injection (FI) in safety-critical applications to assess the reliability and robustness of these accelerators in the face of hardware faults. ### Background and Challenges 1. **Importance of Reliability Evaluation**: In safety-critical applications, the reliability evaluation of DNN hardware accelerators is essential because hardware faults can severely impact the performance of DNNs. 2. **Limitations of Existing Methods**: - **Traditional Fault Injection Methods**: Although fault injection is a commonly used and effective reliability evaluation method, traditional fault injection methods are very time-consuming, especially on large-scale DNN accelerators. - **Hardware-agnostic Fault Injection Tools**: These tools do not consider the underlying hardware, thus failing to accurately simulate actual hardware behavior. - **Hardware-aware Fault Injection Tools**: While these tools consider the hardware, they usually require a significant amount of computational resources and time. ### Main Contributions of the Paper 1. **Proposed a New Hierarchical Software Fault Injection Strategy**: This strategy is specifically designed for DNN accelerators based on systolic arrays. By using a unified recursive equation (URE) system to model the systolic array core, the speed of fault injection is significantly improved. 2. **Developed an Open-source Tool SAFFIRA**: This tool implements the aforementioned fault injection strategy, reducing the fault injection time to 1/3 or even 1/2000 of the existing methods while ensuring accuracy. 3. **Introduced a New Reliability Metric - Faulty Distance**: This metric can better evaluate the classification performance of DNNs under fault conditions. 4. **Performance Evaluation on the Latest DNN Benchmarks**: The effectiveness and efficiency of the framework were validated. ### Experimental Results - **Permanent Fault Injection Experiments**: Permanent fault injection experiments were conducted on different quantized versions (8-bit and 16-bit integers) of the LeNet-5 network. The results showed that the 16-bit network performed better under permanent faults than the 8-bit network. - **Transient Fault Injection Experiments**: Transient fault injection experiments were conducted on the AlexNet, VGG-16, and ResNet-18 networks, evaluating the fault sensitivity and reliability of different networks. ### Conclusion The paper proposes an efficient hierarchical fault injection strategy that significantly improves the speed of reliability evaluation for systolic array-based DNN accelerators through software modeling and optimization. This method not only far exceeds existing methods in terms of time efficiency but also maintains a high level of accuracy. Additionally, the proposed faulty distance metric provides a new perspective for evaluating the fault robustness of DNNs.

SAFFIRA: a Framework for Assessing the Reliability of Systolic-Array-Based DNN Accelerators

Exploration of Activation Fault Reliability in Quantized Systolic Array-Based DNN Accelerators

Saca-FI: A Microarchitecture-Level Fault Injection Framework for Reliability Analysis of Systolic Array Based CNN Accelerator.

Single-Event Upset Analysis of a Systolic Array based Deep Neural Network Accelerator

Towards Reliable AI Applications Via Algorithm-Based Fault Tolerance on NVDLA

DeepVigor+: Scalable and Accurate Semi-Analytical Fault Resilience Analysis for Deep Neural Network

Special Session: Approximation and Fault Resiliency of DNN Accelerators

A Deep Investigation on Stealthy DVFS Fault Injection Attacks at DNN Hardware Accelerators

Algorithmic Strategies for Sustainable Reuse of Neural Network Accelerators with Permanent Faults

A Systematic Literature Review on Hardware Reliability Assessment Methods for Deep Neural Networks

DeepVigor: Vulnerability Value Ranges and Factors for DNNs' Reliability Assessment

A Survey of Design and Optimization for Systolic Array-based DNN Accelerators

A Systolic SNN Inference Accelerator and Its Co-optimized Software Framework

APPRAISER: DNN Fault Resilience Analysis Employing Approximation Errors

ONE-SA: Enabling Nonlinear Operations in Systolic Arrays for Efficient and Flexible Neural Network Inference

RescueSNN: Enabling Reliable Executions on Spiking Neural Network Accelerators under Permanent Faults

Systematic Reliability Evaluation of FPGA Implemented CNN Accelerators

DeepAxe: A Framework for Exploration of Approximation and Reliability Trade-offs in DNN Accelerators

Addressing the issue of processing element under-utilization in general-purpose systolic deep learning accelerators

Impact of High-Level-Synthesis on Reliability of Artificial Neural Network Hardware Accelerators

Evaluating Different Fault Injection Abstractions on the Assessment of DNN SW Hardening Strategies