Influence of Stacked 3D Memory/Cache Architectures on GPUs

Abstract

Graphics Processing Units (GPUs) are highly parallel processing units that offload graphics rendering from microprocessors. For over 20 years, these units were exclusively utilized for graphics processing and over that period, more and more technological breakthroughs have been accomplished in improving the computational power of GPUs. One of the latest advances is the ability to program the GPU pipeline, allowing non-graphics applications and algorithms to run on top of the GPU. This has opened a whole new research area that is concerned with utilizing GPUs for running general purpose applications. Consequently, this has put more pressure on manufacturers in seeking innovative ways of improving GPUs even more. For instance, 3D die-stacking-another emerging technology-can be considered in improving GPU's performance.In this chapter, we discuss how 3D technology can be implemented in GPUs. We also investigate the problems and constraints of implementing such a technology and propose and assess solutions to these problems. Moreover, we propose architectural designs for the GPU that implements 3D technology and evaluate these designs in terms of cost, power consumption and thermal profile.However, before we delve into that discussion, it is very important for the reader to understand how GPU works and what its architecture looks like. Therefore, the next section introduces GPU technology and its hardware architecture. Also, we shortly discuss the 3D technology and the benefits that it offers for this demanding application.