BRAT: Branch Prediction Via Adaptive Training
Author(s)
Lafiandra, Jonathan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
In this thesis, BRAT is researched as a new hardware structure for cost-efficient branch prediction. Relying on the fundamentals of machine learning, BRAT computes a branch decision through a multi-layer neural network. To demonstrate the merits of BRAT, it is used to predict branches in a typical pipeline and evaluate its accuracy. By utilizing a hidden layer and activation functions, BRAT is able to introduce non-linearity and enable more accurate prediction of branch outcomes because this structure exposes relationships that may not be easily captured by a perceptron based approach or other popular methods. The memory utilized by BRAT scales linearly with the number of inputs in the decision process. At most memory footprints, BRAT is competitive with state-of-the-art branch predictors of equivalent memory budgets. Additionally, as the memory footprint is increased, it is shown how BRAT scales and how larger predictors in the future may perform.
Sponsor
Date
2021-08-03
Extent
Resource Type
Text
Resource Subtype
Thesis