Parallelism is the basis of virtually all modern high-performance graphics systems. Over the past few years, many advances have been made on hardware accelerators as well as software-based parallel rendering algorithms. Our lectures will cover four selected topics in state-of-the-art parallel graphics and visualization technology:
Multiprocessor workstations and massively parallel supercomputers will continue to be relevant to large-scale graphics applications. However, as silicon technology continues its inexorable advance we are beginning to see graphics accelerator chips with dozens, even hundreds of processing elements. A look at current trends and the forces behind them suggests that it is now appropriate for the parallel graphics community to pay attention to issues concerning very small machines. Frank Crow will give an overview of the development parallel graphics technology and identify the trend of high performance graphics with a focus on commodity PC platforms.
InfiniteReality system is a near-massively parallel machine, employing over 300 independent processors in its fullest configuration. Kurt Akeley will provide an overview of the architecture, and its relationship to other possible architectures, using the Sort-First/Sort-Middle/Sort-Last classification developed by Steve Molnar. He will spend some time on selected details, chosen to emphasize design choices that increased computation requirements in the interest of reduced bandwidth, improved efficiency of parallelism, or simplicity of implementation. He will also discuss the texture system, which was central to the design constraints of the system. Finally, he will identify both opportunities and constraints of OpenGL as an architecture for parallel, high-performance implementations.
Current high-performance (and that means parallel!) rendering systems suffer from two serial bottlenecks: serial application code and serial transmission of data from geometry processors to pixel processors. Unfortunately, fixing these problems means breaking a number of other aspects of current designs including their APIs and their techniques for state management, ordering, load balancing, and texture access. Gordon Stoll will describe the results of an ongoing project at Stanford to determine how systems will need to look from the outside in order to fix bottleneck #1 and how they will be designed on the inside to fix bottleneck #2 (and to support the fix for #1!). He will describe their parallel rendering system which avoids these bottlenecks, including approaches to state management, ordering and load balancing. He will also describe their experiences in designing and using the API for the system, and the implications of the API design on the internals of the system. Finally, he will describe their initial thoughts on the parallel texture access problem.
Increasingly, leading-edge scientific computations with demanding memory and processing requirements are being performed on massively parallel supercomputers such as T3E, SP2 and Origin 2000. An example is DOE's ASCI program plan. To support applications which use these MPP supercomputers, visualization tools appropriate to the parallel architecture must be developed because most of the application data sets easily overwhelm the processing power and storage space of an average graphics workstation. Using volume visualization applications as examples, Kwan-Liu Ma will illustrate the principles of parallel software rendering on MPP supercomputers. The focus of his talk will be on load balancing and parallel image composition which affect the overall scalability of the rendering process.
We propose a half-day course (3.5 hours) on the above topics which we believe are most relevant based on recent technology advances and current demands from science and engineering applications. Each presenter will be given 40-50 minutes to cover enough background materials for the audience to explore each topic further through reading the collections of papers included in the course notes.
A more concise syllabus with estimated timeline is given as follows: