Perceptron branch predictor

Overview

Predicting branches with perceptrons instead of saturating counters.

My senior thesis explores the perceptron branch predictor: a neural approach to branch prediction where each branch is predicted by a small perceptron trained on global branch history, instead of the classic two-bit saturating counters.

The accompanying video walks through the predictor's behaviour cycle by cycle. The hardware side of the work includes a Wallace tree adder, used to sum the perceptron's weighted inputs quickly enough to fit in a single clock cycle.

Downloads

The full thesis and the defense slides.

The thesis is written in Montenegrin. The presentation contains the diagrams and results used during the defense.

Stack

Links