Project Proposal: Hybrid GPU/CPU ray tracing for massive models

The speed and high exponential growth in performance of modern programmable GPUs makes them an attractive platform for the implementation of many compute intensive algorithms previously run on CPUs. One of the major techniques in computer graphics is ray tracing, which is an essential building block for many algorithms, especially for global illumination techniques. It is also embarrassingly simple to parallelize and as such would be a welcome candidate for implementation on GPUs. However, the basic ray tracing algorithm is not amenable to the stream computation model of GPUs: except for brute force approaches, ray tracing is very branch intensive and involves mostly random memory accesses due to hierarchical data structures, all of which are slow on GPUs. Implementations have therefore lagged far behind fast CPU systems in speed, despite vastly higher theoretical speed of GPUs.