Publication Type : Conference Paper
Publisher : IEEE
Source : 2026 International Conference on Computing Theory and Wireless Communications (ICCTWC)
Url : https://doi.org/10.1109/icctwc68241.2026.11583324
Campus : Amritapuri
School : School of Engineering
Center : Humanitarian Technology (HuT) Labs
Department : Electronics and Communication
Year : 2026
Abstract : Branch prediction is a major process in maintaining the performance of pipelined processors, especially when work-loads have a high density of conditional branches. Traditional predictors most used one be the 2-bit saturating counter are fast and inexpensive, but lose accuracy when multiple branches map to the same entry or when control flow becomes not regular. These drawbacks have motivated a change toward statistical and machine- learning-based alternatives. The following work studies two lightweight supervised models, namely Logistic Regression and Decision Trees, using real branch traces from the 3rd Championship Branch Prediction compe- tition. These traces capture program counters, global branch history, and short-term address patterns, providing a realistic view of runtime branch behavior. In this paper, both models are preprocessed by variance filtering and normalization before they are trained on the large subsets of the traces and assessed on held-out samples. Their performance is then compared against a software implementation of the traditional 2-bit predictor. For all work-loads studied in our paper, ML models perform better than our baseline predictor. Results confirm that even the most basic learning can identify structural relationships in branch behavior that other predictors do not exploit, and they reflect the practical utility of light ML solutions in branch prediction.
Cite this Research Publication : Gujjala Yakshith Reddy, Katikireddy Pranay Reddy, Rajesh Kannan Megalingam, Improving Branch Prediction Accuracy in Pipelined Processors Using Machine Learning Techniques, 2026 International Conference on Computing Theory and Wireless Communications (ICCTWC), IEEE, 2026, https://doi.org/10.1109/icctwc68241.2026.11583324