Also that gpu have little bandwith toward main memory and most cuda patterns involve loading, computing and retrieving which works as well remotely as locally,

Also that gpu have little bandwith toward main memory and most cuda patterns involve loading, computing and retrieving which works as well remotely as locally, as opposed to programs accessing memory at random time all over their memory space the gpu memory is neatly organized in texture areas and the programming paradigm already entails moving them in and out the device as few times as possible