Mixed-precision AMG as linear equation solver for definite systems
Maximilian Emans, Albert van der Meer · Procedia Computer Science · 2010
The performance of algebraic multigrid (AMG) algorithms, implemented in 4-byte floating point arithmetic, is investigated on modern cluster architecture with multi-core CPUs. The algorithmic considerations comprise the effect of preconditioning in 4-byte floating point arithmetic on Krylov solvers using standard 8-byte floating point arithmetic. The data of basic linear algebra benchmarks are used to interpret the performance of AMG algorithms employed as linear solvers in computational fluid dynamics simulation tools.