Is it possible for milti-GPU cards? Copying data form 1st to 2nd GPU is slow - it will be nice to just send pointer etc. instead of copying all data. It will be possible to use for example 80% of memory by 1st GPU and 20% by 2nd. Maybe for typical computing better to have two (or more) graphic cards with independent memory or one 2x faster GPU but when 2nd card prepares the data which is using by 1st in rendering - I think that one big and shared memory will be useful.
But I don’t see it exists. So - it’s useless, too expensive to implement or only some people could buy it?