Publication Details
Abstract
Predicting CPU performance in the absence of physical benchmarking is a practical problem facing hardware engineers, system designers, and researchers with limited resources. This research proposes an open-source and reproducible hardware-software workflow that is capable of predicting Cinebench R23 multi-core scores based on hardware specifications alone, without any benchmark measurements as an input to the prediction models. This workflow is applied to a dataset of 215 state-of-the-art CPUs from various hardware vendors, including Intel, AMD, and Apple, with CPUs classified into Desktop and Laptop categories. Four different machine learning models, namely Linear Regression, Ridge Regression, Random Forest, and Gradient Boosting, are utilized in this workflow. All features, such as core count, thread count, base clock, and turbo clock, are actual hardware specifications available from datasheets prior to any physical benchmarking or testing. A log transformation was applied to the target variable to reduce skewness in data, resulting in a reduction from 4.035 to 1.127. It is found that Gradient Boosting performs better, with R2 = 0.845, RMSE = 0.149, and MAE = 0.106 on the test data, validated by 5-fold cross-validation with a high R2 of 0.853 ± 0.037. Feature importance analysis of the models shows that thread count is the primary feature contributing to multi-core performance, followed by core count, while clock speed contributes to a much lesser extent, about 6% each, and manufacturer is of negligible importance, less than 1%. This workflow is reproducible, platform-independent, and executable on any standard laptop without any additional hardware, making it accessible to researchers in constrained environments.