next up previous
Next: About this document ...

Jonathan Cohen
Domain decomposition and multilevel methods on GPUs

Nvidia Corporation
jocohen@nvidia.com

Modern GPUs are highly efficient massively parallel general purposes processors. A single high-end GPU from NVIDIA can achieve several hundred GFLOPS/s in double precision arithmetic. Because of their inherently parallel and high-throughput nature, they are a natural fit for domain decomposition and multilevel methods. I will give a brief technical overview of NVIDIA's CUDA platform for GPU computing, and present initial work to develop sparse linear solvers using CUDA.





Copper Mountain Conference 2011-02-20