Research Seminar with Eda Oktay (MPI Magdeburg)

Eda Oktay, MPI Magdeburg Golm, House 9, Room 0.1312:00-13:00

 Reduced- and Mixed-Precision Block Householder QR Decomposition

Recent advances in hardware architectures have highlighted the importance of reduced-precision arithmetic for improving performance and lowering energy consumption in large-scale scientific applications. However, uniformly low-precision methods can suffer from numerical instability, especially in fundamental linear algebra routines. Mixed-precision strategies, therefore, play a crucial role in balancing efficiency, stability, and accuracy. In this talk, after providing background on mixed-precision arithmetic, we focus on the block Householder QR decomposition using the WY representation and explore its behavior under reduced- and mixed-precision settings. We first present a uniform half-precision implementation and evaluate its potential performance benefits and significant numerical-accuracy limitations. We then show how increasing the precision of selected computations within the 
algorithm can substantially improve stability, and we analyze how these precision choices affect overall performance.