Chapter 7 - Compute Unified Device Architecture

Abstract

A Graphics Processing Unit (GPU) is an integral piece of hardware in many electronic devices. It is used to expose graphical interfaces to the user and can be found in almost every workstation or mobile device. Starting from the mid-nineties of the last century, the increasing popularity of 3D-accelerated games (such as the Quake and Unreal series) caused a rapid growth of computational capabilities of modern GPUs allowing for the rendering of more and more complex scenes. In the late nineties, manufacturers extended the GPU's core functionality, the efficient rendering of 3D scenes, to additionally support the processing of geometric operations, notably the pioneering NVIDIA GeForce 256 [5] and its dedicated Transform and Lighting (T&L) unit. This trend continued in the new millennium with the introduction of pixel and vertex shader languages which allowed for the manipulation of the prior to this hard-wired rendering pipeline. Consequently, computer scientists used the increasing computational power of GPUs to implement more general algorithms by expressing them in the aforementioned shader languages. This was the birth of the so-called general-purpose computing on GPUs (GPGPU). A major drawback of GPGPU programming was its lack of abstraction: the processed data had to be encoded in textures and manipulated with the instruction set of the shader language which substantially limited the design of complex programs.11An exhaustive overview of the history of GPGPU can be found in [13].