Title:Towards CPU Performance Prediction: New Challenge Benchmark Dataset and Novel Approach

Abstract:The server central processing unit (CPU) market continues to exhibit robust demand due to the rising global need for computing power. Against this backdrop, CPU benchmark performance prediction is crucial for architecture designers. It offers profound insights for optimizing system designs and significantly reduces the time required for benchmark testing. However, the current research suffers from a lack of a unified, standard and a comprehensive dataset covering various CPU benchmark suites on real machines. Additionally, the traditional simulation-based methods suffer from slow simulation speeds. Furthermore, traditional machine learning approaches not only struggle to process complex features across various hardware configurations but also fall short in achieving sufficient accuracy.
To bridge these gaps, we firstly perform a streamlined data preprocessing and reorganize our in-house datasets gathered from a variety CPU models of 4th Generation Intel Xeon Scalable Processors on various benchmark suites. We then propose Nova CPU Performance Predictor (NCPP), a deep learning model with attention mechanisms, specifically designed to predict CPU performance across various benchmarks. Our model effectively captures key hardware configurations affecting performance in across various benchmarks. Moreover, we compare eight mainstream machine learning methods, demonstrating the significant advantages of our model in terms of accuracy and explainability over existing approaches. Finally, our results provide new perspectives and practical strategies for hardware designers. To foster further research and collaboration, we \textit{\textbf{open-source}} the model \url{this https URL}