GPU Machine Learning Library (GPUMLib)

Abstract

The previous chapter accentuated the need for the understanding of large, complex, and distributed data sets generated from digital sources coming from sensors or other physical instruments as well as simulations, crowd sourcing, social networks or other internet transactions. The focus was on the difficulties posed to ML algorithms to extract knowledge with prohibitive computational requirements. In this chapter we introduce the GPU, which represents a novel and compelling solution for this problem, due to its inherent high-parallelism. Seldom ML algorithms have been implemented on the GPU and most are not openly shared. To mitigate this problem, this Chapter describes a new open-source library (GPUMLib), that aims to provide the building blocks for the development of efficient GPU ML software. In the first part of the chapter we cast arguments for the need of an open-source GPU ML library. Next, it presents an overview of the open-source and proprietary ML algorithms implemented on the GPU, prior to the development of GPUMLib. In addition we focus on the evolution of the GPU from a fixed-function device, designed to accelerate specific tasks, into a general-purpose computing device. The last part of the chapter details the CUDA programming model and architecture, which was used to developGPUMLib. Finally, the generalGPUMLib architecture is described.