Title:
Exploiting Mixed Precision Arithmetic in the Solution of Linear Systems
Abstract:
The emergence of low precision arithmetic on modern hardware, such as the increasingly supported fp16 and bfloat16 floating-point arithmetics, provides new opportunities to accelerate linear algebra computations. A key challenge in doing so is to preserve the accuracy and stability of the computation. In this talk we will present mixed precision algorithms for the solution of linear systems, that combine low and high precisions in order to achieve both high performance and high accuracy. We will describe recent key advances in making these algorithms effective on modern hardware, in very low precision, and on a wide range of matrices arising in various applications.
