Overview

Current commercial graphics processing units (GPUs) are capable of rendering large numbers of triangles at interactive rates, but consume a tremendous amount of power in order to do so. Even more problematic is that the rate at which GPU power has grown exceeds the CPU power growth rate. The problem is exacerbated by the continual demand for increased graphics quality. Each new process generation enables GPUs to reduce the energy required to render an image frame, but the in- crease in screen resolution, scene complexity, and image quality conspire to create the rapid rise in power consumption. The goal of the proposed research is to develop new architectures for high-performance graphics processing that will significantly enhance the ability to render visually realistic scenes, and to do so in a manner that consumes less power than current GPU power growth trends. Our proposed approach involves the development of novel chip microarchitecture, memory systems, and graphics algorithms.

Today’s general purpose GPUs (GPGPUs) are more programmable and less specialized than previous generations and have been shown to accelerate not only graphics processing but a number of high-performance scientific applications as well. The primary graphics algorithm used is the Z-buffer, and the acceleration mechanism (for both graphics and scientific apps) is to exploit parallelism through wide SIMD processing. Increasing demands for realism in rendered images requires an increasingly complex set of techniques using the Z-buffer algorithm. Most of these advanced lighting effects are more naturally achieved by an alternate rendering algorithm, namely ray-tracing. Movie production companies such as Disney/Pixar and Sony Imageworks use some form of ray-tracing to achieve global illumination effects. Ray tracing simulates light transport through a scene and naturally supports high quality composite global lighting effects such as shadows, transparency, reflections, refractions, and indirect illumination.

Our research agenda is to develop system architectures that will provide enough performance to render movie-quality images at interactive rates. This will require at least one to two orders of magnitude performance increase over even the most aggressive systems currently proposed in the literature. We believe that incremental improvement of existing commercial approaches will not achieve this goal at reasonable power levels. We propose a full system approach which examines the problem at all levels: fundamental algorithms, data structures, memory systems, circuits, execution pipelines, and system architectures.

Intellectual merit of the proposed activity

Because our plan is a top to bottom exploration of the problem, we expect significant advances in a variety of areas. We propose to focus on ray tracing as our fundamental algorithm because of the importance of global lighting effects in realistic rendering. Expected primary contributions of the overall project include: design of new algorithms for dynamic refit of the hierarchical bounding structures used to accelerate ray traversal so as to better support dynamic scenes; novel application of data streaming in many aspects of the ray tracing algorithm; structure-aware memory systems that react and respond autonomously to data structures; new techniques for dealing with procedural geometry and texturing to trade computation for memory bandwidth; support for reduced-precision floating point processing where appropriate; a system organization of lightweight MIMD thread processors that perform well with rays that are not efficiently executed in SIMD bundles; flexible system organizations that allow chained pipelined data-flow operations between thread processors for streaming phases of the algorithms.

 

Broader impacts of the proposed activity

The human computer interface has been significantly enhanced by improved graphics systems. We believe that increasing visual realism will enhance commodity system support of virtual realistic interfaces. Energy efficiency is a national priority and reduced energy consumption is critical to reducing the cost of computing systems. Cheap, yet high quality graphics will enable a broader user spectrum to efficiently interact with the information infrastructure. The educational experience at all levels can similarly be enhanced. In addition, our connections with commodity CPU and GPU manufacturers should allow an efficient technology transfer of our architectural results to industry as the project matures. In terms of outreach, visual images are a powerful motivator in attracting new and di- verse students to computing. We already do and will continue to work with the Entertainment Arts and Engineering program at the University of Utah, and with our summer outreach programs in the School of Computing, to use graphics processing as an introduction to computing for a broad variety of students including middle- and high-school students.