[gnwt4a]

never compare per-iteration timings in black-box codes. u dont know if the algorithm has been changed, the precision lowered or the arithmetic (usually divs/logs/trigs) done at reduced (but faster) accuracy. always compare timings of what is considered a converged solution at (very) high resolutions - (eg) 5 digits in the mean velocity profile.


Edit: Furthermore, do not use simple geometries (e.g. no plane channels nor straight pipes) which prove nothing and are likely to use specially simplified algorithms, and never uniform grids that converge fast and can be done more efficiently with trig expansions.

[arjun]

Originally Posted by CFDfan

I observed the same weird behavior (different answers, or solver divergence when using the GPU) with FEA codes for electromagnetic simulations as well. Obviously not all bugs are cleared and "fast" doesn't necessarily mean "accurate"

Solver can diverge in GPU because the AMG is hugely affected by it. Both the coarsening and the smoothing part.





PS: I have recently developed a smoother that shall not be affected by this issue but still i do not have it in GPU version. At some point of time, will get it. In other solvers, they suffer.