A Journey Through Parallel Dense Eigensolvers

Computing the eigenvalues and eigenvectors of dense matrices is a keystone component of applications ranging from materials science simulation to data science. Taking the Bethe-Salpeter Equation as an application case — which leads to large, structured dense eigenvalue problems in ab initio optical property calculations —, this seminar will survey the state-of-the-art of dense eigensolvers in parallel with the evolution of supercomputers through time. The conclusion will open a discussion on anticipated changes in modern numerical libraries toward low arithmetic precisions, driven by the growing dominance of AI-optimized hardware.