We show recent advancements in the automated steady-state simulation of incompressible, highly viscous fluids with complex constitutive laws, in combination with fluid domains implicitly described by surface meshes and the transition to general purpose GPUs for scalable multi-node computations, using the example of a steady-state flow through an extrusion die. Generally, fluid domain adapted meshes of sufficient quality cannot be generated without some human intervention. To fully automate this process, we instead use fictitious boundary methods in combination with h and r adaptive meshes. This approach generally requires higher resolutions; hence, the need for scalable and efficient methods for large-scale simulations. To accomplish this, we use a hierarchical monolithic multigrid approach and show its scalability on current HPC clusters for complex, industry relevant 3D geometries. We also provide insights into challenges for the adaption to multinode GPU systems and the ways we overcome these. All of this is accomplished in the FEAT3 software, which provides the required combination of hierarchical mesh generation, efficient and robust multigrid solvers, as well as flexible backends for linear algebra and finite element assembly kernels, capable of making use of the steadily increasing portion of GPU compute resources on recent HPC clusters.