Speaker
Dr
Evan Weinberg
(NVIDIA)
Description
The era of exascale computing enables the generation of ever-finer gauge configurations, capturing gauge-fermion physics with unprecedented accuracy. This approach to the continuum comes with a super-linear increase in the cost of the iterative Krylov solve of the Dirac fermion operator, the phenomena of critical slowing down. Multi-grid methods are the optimal approach to addressing this crisis of cost. In this talk we describe the state-of-the-art implementations of multigrid algorithms for Wilson-clover, HISQ, and domain wall fermions in the QUDA library for GPUs. This will include a discussion of the unique approaches to multigrid for each fermion formulation.
Primary author
Dr
Evan Weinberg
(NVIDIA)
Co-authors
Richard Brower
(Boston University)
Kate Clark
(NVIDIA)
Jiqun Tu
(NVIDIA Corporation)
Mathias Wagner
(NVIDIA)