A model of architecture for estimating GPU processing performance and power

Efficient usage of heterogeneous computing architectures requires distribution of the workload on available processing elements. Traditionally, the mapping is based on information acquired from application profiling and utilized in architecture exploration. To reduce the amount of manual work required, statistical application modeling and architecture modeling can be combined with exploration heuristics. While the application modeling side of the problem has been studied extensively, architecture modeling has received less attention. Linear System Level Architecture (LSLA) is a Model of Architecture that aims at separating the architectural concerns from algorithmic ones when predicting performance. This work builds on the LSLA model and introduces non-linear semantics, specifically to support GPU performance and power modeling, by modeling also the degree of parallelism. The model is evaluated with three signal processing applications with various workload distributions on a desktop GPU and mobile GPU. The measured average fidelity of the new model is 93% for performance, and 84% for power, which can fit design space exploration purposes.