CFD and other high-fidelity multiphysics simulations can be efficient when GPU-bound. It all comes down to being able to write a matrix and stream a bunch of ve

CFD and other high-fidelity multiphysics simulations can be efficient when GPU-bound. It all comes down to being able to write a matrix and stream a bunch of vectors to multiply it by. When considering Krylov subspace methods, this extends to covering a large part of the problem space of linear algebra. I think the Tesla cards in particular are double precision, which is not necessarily the best for deep learning applications.