[bauta-steen]

Perhaps the main issue is that even if a PDE solver supports distributed CPU parallelism, the distributed block solvers do not typically allow for decoupling to thousands of independent threads that GPUs are good at. Therefore as the PDE problems and solvers are tightly coupled they do not easily parallelize to GPUs and don't allow for simple recompilation with GPU targets. Most often an existing code would require a complete rewrite/redesign (man years of work for big code bases), at least if there are to be any gains to be had. There are new codes coming, particularly in academia, utilizing GPUs. From what I've seen one can expect around 10x improvement switching to GPUs, so it's good, but not magnitudes better considering the work involved.