Computational Science seminar series: Dr Srikara Pranesh

Mixed Precision Algorithms in Numerical Linear Algebra : Techniques, Analysis, and Software.

Speaker: Dr Srikara Pranesh

With constant increase in computational power, the role of numerical simulations is becoming more pronounced in science and engineering. Since numerical linear algebra (NLA) lies at the heart of many scientific computing applications, developing fast and robust algorithms with reliable software can facilitate in solving more challenging problems. In this talk I will discuss some recent advances in using low precision floating-point arithmetic in NLA algorithms. Specifically some of my recent work in the solution of dense linear system of equations. I will begin by discussing GMRES(Generalized Minimal Residual)-based iterative refinement (GMRES-IR) method for the solution of linear system of equations, followed by a demonstration of four times speedup against state-of-the-art solver in LAPACK. Next I enumerate some mathematical issues that must be addressed to ensure robust implementation of GMRES-IR in software, and discuss possible remedies for these issues. Extension of these ideas to least squares problems will also be discussed.

Connection via Zoom. Contact Tom Ranner (T.Ranner at leeds.ac.uk) for connection details or to subscribe to mailing list.

Details of seminar TBC.