High-Performance Computation of Voronoi Diagrams using a Graphics Processing Unit

Document Type



Dr. Greg Wolff, wolffe@gvsu.edu; Dr. Christian Trefftz, trefftzc@gvsu.edu

Embargo Period



A new trend in high-performance computing involves the increasing performance and increasing programmability of commodity Graphics Processing Units (GPU). The focus is to implement algorithms specifically for these highly parallel and efficient processors. The algorithms are designed to execute in Single Instruction Multiple Data (SIMD) mode. For this project, a computational geometry algorithm, the calculation of an approximation to the Voronoi diagram, was successfully implemented to test the feasibility of performing general purpose computations on GPUs. Nvidia’s Compute Unified Device Architecture (CUDA) technology was used as the development platform; the GeForce 8 Series GPU was used for execution. The performance of the algorithm on the GPU under varying parameters was compared with serial code executing on a traditional CPU. Hardware and software profiling was used to develop a model for characterizing the behavior of the code on the experimental architecture. Extensive testing demonstrated the ability of the model to predict the performance and scalability of the algorithm.

This document is currently not available here.