Hardware-Aware Analysis and Optimization of Stable Fluids

by syoyo

Hardware-Aware Analysis and Optimization of Stable Fluids
Theodore Kim
I3D 2008.

(image from Hardware-Aware Analysis and Optimization of Stable Fluids)

I like what they are doing in this paper, i.e., first estimate theoretical performance of computation and bandwidth of the algorithm, compares measured performance and theoretical peak to find bottleneck, then derive new technique which fills gaps between measured performance and theoretical peak.

Stable Fluid is bandwidth-intensive

They find that classic “Stable Fluid” algorithm is bandwidth-intensive.
80% ~ 95% of computational core is idle while waiting input/output data.
Its a bit surprise because I’ve thought fluid simulation algorithm is still compute-intensive.

I think, in global illumination algorithm development case, estimating a theoretical peak performance as done in this paper is alsp important before starting to implement my/your new GI algoritm.