Optimizing Particle Simulation for Kepler GPU

We optimized Moving Particle Simulation (MPS) method for Kepler GPU. Solving sparse matrix occupies a large portion of particle simulation because it is a casus of performance bottlenecks. We reported at NVIDIA's GPU Technology Conference 2012 that we achieved to accelerate MPS method about 7x faster on NVIDIA Tesla C2070 compared with on Intel Core i7 920 (4 cores). Last year, NVIDIA released the latest GPU, Kepler. We optimized and accelerated particle simulation on Kepler GPU by using new shuffle instruction and read-only data cache. We obtained to accelerate Sparse Matrix-Vector multiplication operation 1.48x faster on Kepler (Tesla K20c) compared with on Fermi (Tesla C2075).