It seems like path dependent lock in to me and a risk/reward calculation.

What do you gain by not using CUDA vs what do you risk?