Speeding Up LS-DYNA Implicit with Mixed Precision, Low Rank Approximations, and Accelerators
The multifrontal method of Duff and Reid [1] dominates the runtime of most LS-DYNA im-plicit analyses. Its complexity will range from O(N1.5) to O(N2), depending on the model. This paper will give an overview of attempts to reduce the run time of solving large systems of linear equations, both on the host processor as well as with accelerators. Most of what is discussed herein is available today in the development version of LS-DYNA and should be released with R15. Everything discussed herein only applies to our symmetric indefinite solver.
https://www.dynalook.com/conferences/14th-european-ls-dyna-conference-2023/nvh-implicit/lucas_ansys.pdf/view
https://www.dynalook.com/@@site-logo/DYNAlook-Logo480x80.png
Speeding Up LS-DYNA Implicit with Mixed Precision, Low Rank Approximations, and Accelerators
The multifrontal method of Duff and Reid [1] dominates the runtime of most LS-DYNA im-plicit analyses. Its complexity will range from O(N1.5) to O(N2), depending on the model. This paper will give an overview of attempts to reduce the run time of solving large systems of linear equations, both on the host processor as well as with accelerators. Most of what is discussed herein is available today in the development version of LS-DYNA and should be released with R15. Everything discussed herein only applies to our symmetric indefinite solver.
Lucas_Ansys.pdf
— 561.2 KB