Only 256Mb per desktop will be used in worst case, should be negligible compared to all the other overhead.
It does work as intended and will not change anytime soon, another option is to use Directpath IO and dedicate a GPU to a virtual machine, but then you will be limited to the amount of GPU:s on the card.
For example the K1 have 4 GPU:s so then you can have 4 virtual machines with GPU, the K2 have 2 GPU:s.
This is called "vDGA" and is currently in Tech Preview.
// Linjo