The MFEM team has created a helpful page of tips and tricks that explain how to make the most of GPUs when running finite element algorithms. This support documentation includes information about optimizing porting and performance. Learn more about these features and processes:

• MFEM’s internal memory manager to simplify the use of host/device memory
• MFEM_FORALL macro to enable performance portability
• Maximizing the main memory bandwidth
• Profiling on NVIDIA GPUs to improve the performance of a memory bound kernel
• Roofline model for predicting the peak performance achievable by a specific algorithm

Questions or suggestions can be posted to MFEM’s issue tracker.