It is how it is done in Octane and it is the cause of the memory limitation, because GPU memory is limited to 1.5GB on most powerful affordable cards (I doubt that we will have more memory on these Geforce cards in the future, because Nvidia has also to sell its very expensive Tesla and Quadro with 4GB).1) Wouldn't it be more prudent to store the scene geometry on the gpu memory?
So storing the geometry in the Motherboard memory would allow much more memory (7GB available on my 8GB amount of system RAM).
For what I have understood about Octane, Storing the geometry on the GPU memory allows much faster acces to the informations for computing, as data are close to the computing cores (once the scene is loaded, there is few transfer activity on the PCIE data bus), but the drawback is the lack of available memory for big scenes or high resolution rendering.