BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Exploiting Mixed Precision in Numerical Linear Algebra
DTSTART:20211102T161500
DTEND:20211102T171500
DTSTAMP:20260407T091718Z
UID:383bcc1f3505dcecd0d4ffe262940d18583f5c3ddc2bf205204b6fe2
CATEGORIES:Conferences - Seminars
DESCRIPTION:Dr. Erin C. Carson\, Charles University Prague\nSupport for fl
 oating point arithmetic in multiple precisions is becoming increasingly co
 mmon in emerging architectures. Mixed precision capabilities are already i
 ncluded in any machines on the TOP500 list and are expected to be a crucia
 l hardware feature in coming exascale machines. From a computational scien
 tists perspective\, our goal is to determine how and where we can exploit 
 mixed precision computation in our codes. This requires both an understand
 ing of performance characteristics as well as an understanding of the nume
 rical behavior of algorithms in finite precision arithmetic.\n\nIn this ta
 lk\, we discuss recent and ongoing efforts in this area. In particular\, w
 e present and analyze a general algorithm for solving nxn nonsingular line
 ar systems Ax = b based on iterative refinement in three precisions. From 
 this\, we develop GMRES-IR\, a three-precision GMRES-based iterative refin
 ement scheme that works for even ill-conditioned systems. We discuss perfo
 rmance results on modern GPU architectures and present the HPL-AI benchmar
 k\, based on our mixed precision iterative refinement algorithms. The worl
 d's top supercomputers already exceed exaflop performance on HPL-AI\, achi
 eving over 4x higher performance than on the standard HPL benchmark.
LOCATION:GA 3 21 https://plan.epfl.ch/?room==GA%203%2021
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR