Just as a note, GPU supported code just now seems to be coming on line with like Quantum Espresso GPU version getting an alpha in 2019. I think a large part of that is due to the 64 bit precision making the speed up not as great on GPUs (and to a large extent eliminating the usage of consumer level cards which are popular at the classical molecular dynamics software like GROMACS). Plus Guassian, while having support since Kepler days note that earlier cards didn't have the memory requirements (which also is a huge problem for consumer cards) making adoption slower, because of the added costs.