# Meeting Details

Title: Parallel Auxiliary Grid AMG Method for GPU CCMA Luncheon Seminar Lu Wang, Penn State Mathematics We develop a new parallel auxiliary grid algebraic multigrid (AMG) method for graphic processing units (GPU). The new method uses the information from the finest grid to construct an auxiliary structured grid if necessary and select a simple and fixed coarsening based on the auxiliary grid that allows explicit control of the overall grid and operator complexities of the AMG solver. These features allow (nearly) optimal load balancing and predictable communication patterns, which makes our new algorithm suits parallel computing on GPU well. We also design parallel smoother based on the special coloring of the auxiliary structured grid to accelerate the solver phase of the AMG method. We implemented the new parallel auxiliary grid AMG method on GPU and the numerical results show about $4$ times speed up comparing with the CUSP library.