Hardware Based LDLT Decomposition For Large Matrices

Abstract

Market demands forces designers and engineers to develop more advanced products quicker and at lower prices. Computer simulations is one way of increasing development speed, thereby lowering the time spent verifying and testing the product. The most common way to do numerical analysis (computer simulations) on a variety of problems (exmaples include, but are not limited to strengt, thermodynimcal and/or fluid problems) is the Finite element method. The most computationally intensive part of the finite element analysis is the solving of the linear equation Ax = b. One of the most effective ways to solve this equation is by using a decomposing algorithm called LDLT decomposition.In this thesis a hardware LDLT decomposer is studied. It is shown how the number of memory reads needed is able to be roughly halved if implemented parallelly, however the memory bandwidth still proves to be the bottleneck of this operation.