Speaker
Anthony Grebe
(University of Maryland, College Park)
Description
Historically, the cost of lattice-QCD calculations was dominated by applications of the Dslash matrix on a single right-hand side, which is constrained by memory-bandwidth limits to run at a small fraction (10-20%) of the available processing power. The advent of multigrid methods and multi-right-hand-side solvers mean that these limits are no longer the binding constraints they once were. Additionally, for nuclear physics calculations, a substantial fraction of the total computational cost is in the contractions of quark propagators into correlation functions. I will present software developments that I have made to help accelerate these, and I will also discuss the role of generative AI in this code development.
Author
Anthony Grebe
(University of Maryland, College Park)