Good question…
My 3d workflow starts from Rhino then it’s translated to 3ds Max.
The 1st gpu (the AMD) is used only for the display, it’s not touched by the Cuda resources.
Then I can use full 12GB of vram dedicated to Vray GPU kernel for all gpus.
This is valid also for Rhinoceros 7 or 8 using Cycles or other gpu based renderers.
If I used the Windows resources for the display and the directx/opengl framebuffer I would have only 5/6GB free (or less if the visualized 3d meshes are millions) for the Cuda kernel.