Memory Dependence Prediction for Energy Constrained Devices

Historically, energy constrained devices (ECDs) have favored simple in-order pipelines over out-of-order (OoO) ones due to lower energy consumption. While prior research has focused on optimizing in-order compute architectures, not a lot of work exists on adapting OoO pipelines for ECDs. This thesis aims to help enable speculative execution and OoO processing to increase performance on such devices. Memory dependence prediction (MDP) is one such mechanism used for achieving these goals. This thesis leverages prior work and applies machine learning to uncover new MDP designs. A novel perceptron MDP is first developed that almost matches the performance of Store Sets, a state of the art MDP. The insights gained during the process lead to a much simpler one bit counter (OBC) technique suitable for use on low powered devices. The OBC MDP design matches the performance of Store Sets while saving 95% area and 96% energy.M.A.S