Analysing Performance and Efficiency of Smoothed Particle Hydrodynamics
Rama Hoetzlein and Tobias Hollerer
Graduate Workshop in Computer Science, UC Santa Barbara, 2008.

Video #1: Zero gravity. 900 x 360, DivX avi., 18 MB
Video #2: Levy break. 676 x 270, DivX avi., 45 MB

With the rise in performance of modern GPUs, Smoothed Particle Hydrodynamics is an increasingly attractive solution
for real-time simulation of fluid flows in visual effects for film and games. Starting with simulations of 2000 particles
at 20 frames per second in 2003 [Muller], smoothed particle hydrodynamics has now been simulated with over 64,000
particles at 57 fps using GPUs [Zhang]. While performance has clearly increased, the terms performance and efficiency are
often used interchangeably. This makes comparisons of algorithm implementations difficult since authors must deduce algorithm efficiency for differing hardware. The development of concrete metrics of performance and efficiency will facilitate better comparison between results. Simple metrics are presented here with an analysis of real-time simulations over the past five years.

A GPU-based fluid simulation was developed using CUDA for direct comparison of CPU and GPU performance (above). Using our GPU implementation, we also render fluid dynamics with shadows, screen-space ambient occlusion, and depth-of-field at near interactive rates.