Exploração do Espaço de Projetos de Sistemas GP-GPU ciente de Dark Silicon

Abstract

The high potential for parallelism and bandwidth offered by GPUs and the popularization of the CUDA and OpenCL programming languages made GPUs useful in different applications from those for which they were originally designed to. Then, these facts consolidate the concept of GP-GPU or Graphics Processing Units for General-Purpose computing. With the use of systems that join CPUs and GPUs for collaborative processing, tools were developed to explore the performance and consumption of the various architectural parameters in heterogeneous computing designs. However, these tools are scarce, limited, computationally expensive, and need architectural parameters that are difficult to obtain. This work proposes the design of performance prediction models for GP-GPU systems from Machine Learning techniques. We evaluate the predictors in a design space exploration tool. MultiExplorer meets pre-defined goals such as performance maximization and dark-silicon area minimization, subject to constrains as circuit area and energy consumption bounds. Depending on the design space, this tool evaluates hundreds of thousands of architectural alternatives and, therefore, performance predictors with low delay and high accuracy are essential.