Nondeterministic failure for AccFFT on GPU
@4pf has noted that some of the larger benchmark problems sometimes fail for AccFFT on GPUs. MEUMAPPS-SS runs fine using AccFFT on GPUs for a single node, but fails on two nodes (in my limited experience, it always fails).
At least for MEUMAPPS-SS, AccFFT on CPUs seems fine (@4pf, is that true in your experience too?). The heFFTe runs are also ok.
What happens is that the nonlinear solver doesn't converge, although it isn't yet clear where the discrepancy is.