A Comparative Study of Recursive Decision Tree Performance on CPU vs. CUDA

Recursive decision trees are traditionally executed on CPUs due to their sequential branching structure. Recent tensor libraries such as PyTorch, however, enable GPU acceleration of the numerical operations performed within each tree node. In this work, we compare CPU and CUDA implementations of a recursive decision tree algorithm using an expanded Stroke Prediction dataset with 500,000 rows. Both implementations share identical recursive logic and differ only in their internal execution backend. Our experiments show that the CUDA implementation achieves an approximate 6.26× speedup over the CPU version. Despite the dataset’s limited feature dimensionality, the results indicate that GPU acceleration can significantly improve the performance of recursive algorithms, with even greater gains expected for larger datasets.