Graphics Architecture, Winter 2009 John Owens

In our final content lecture, we look at how to parallelize the graphics pipeline. What is challenging about parallelizing the GPU? What are the ways we could parallelize it? We discuss the sorting taxonomy of parallelism strategies, look at different ways to communicate within a multi-node system, and analyze the taxonomy using historical graphics architectures.

Jeremy Sugerman from Stanford describes GRAMPS, a programming model for graphics pipelines and heterogeneous parallelism.

We turn away from a fixed-function graphics pipeline and explore what we can do with a user-programmable pipeline, where not only pipeline stages but also the structure of the pipeline can be customized. We look at Reyes, delay streams, and the programmable culling unit.

This lecture contains the overflow from the 4 pipeline lectures, mostly the composition/display lecture.

The final stage of the graphics pipeline is composition/display. In this lecture we look at antialiasing algorithms, compositing, the depth buffer, and monitors. [Note: The beginning part of this lecture is the remainder of the rasterization lecture, and this lecture spills into the overflow lecture.]

John Nickolls, chief compute architect for NVIDIA's GPUs,
discusses NVIDIA GPU graphics and compute architecture.