Home · GPUs in the FEM pipeline

Not every stage of FEM wants a GPU

A finite element analysis is a pipeline: mesh, local kinematics, constitutive evaluation, assembly, linear algebra, and postprocessing. A GPU pays when many identical arithmetic operations run with the data already on the device. Element loops, explicit internal forces, and sparse matrix–vector products do. Meshing, small direct factorisations, and a solve that copies the matrix every iteration do not. Click a stage. Switch the analysis type: the bottleneck moves, and so does the case for the device.

Wall time · CPU vs after GPU

S = 1 / (1 − p + p/s)

Navy is GPU usefulness of the stage. Click a bar. Orange in the time plot is the same analysis after moving the parallel stages, including a transfer tax that dominates at small DOF.