MultiObjective GPU design space exploration optimization
Ali Jooya, N.J. Dimopoulos, Amirali Baniasadi · 2016
Obtainable power and performance for GPGPU applications on a GPU depend on many architectural and software parameters. Therefore, it is crucial to have a model to explore the design space and highlight a smaller subset of configurations that meet a given system goal. In this study, we present an application specific, MultiObjective Optimizer that explores the design space of GPUs and finds close to optimum configurations with respect to multiple objectives. The proposed model is composed of three steps to a) find the effective range for configuration parameters, b) predict power and performance of the application by utilizing a Neural Network based predictor and c) analyze the model's predictions and perform Pareto Optimal multiobjective optimization to produce a small subset of configurations which are optimized with respect to both power and performance. We compare the model produced Pareto Optimal configurations to actual Pareto Optimal configurations obtained from simulations and show that the Pareto Optimal configurations obtained from the model is very close to the actual ones.