Fault Classification and Prediction of AI Accelerators Based on Activation Maximization

This research proposes a method based on neural network interpretability techniques to effectively classify and predict AI accelerator failures. When an AI accelerator's Processing Element (PE) fails, different factors such as the model, input, and location of the failure can affect the output of the model in various ways. This method utilizes neural network interpretability, particularly activation maximization techniques, to classify AI accelerator failures into critical and non-critical based on the extent of their impact on the accelerator's performance. This approach maximizes the interpretability of the model in real-time failure prediction and classification, and compared to traditional method(FTM), the proposed method(AMM) has only a slight 0.6% decrease in fault coverage rate while reducing the classification time by 61.5%. It also avoids precise identification of the fault location and facilitates transitions between models.