The Vega FE is already in stores. (not the cheaper RX)
AMD has been advertising their High Bandwith Cache Controller on their new Vega cards. It is supposed to allow for VRAM objects to be offloaded to slower system memory or storage so that the device doesn't hit a memory limit.
Will this be integrated into Blender? It would be nice to not run out of VRAM. I understand that a "pagefile" style memory dump would be slower than VRAM, but wouldn't it still be faster than using a CPU? If in theory the card can handle memory requests wouldn't blender only need to not stop at the physical VRAM limit?
This is a question diving into how the main release of blender handles memory objects. Also, the FE Professional cards with HBCC have been for sale for weeks. (They are just really expensive) No one has probably looked into this yet. I can see that https://developer.blender.org/diffusion/B/ has the source code. I just did a git clone of the master project. Opened the GPU_framebuffer header. It looks like it uses the OpenGL framebuffer framework. I don't think it would take long to adapt to the HBCC. Anyone look at this yet?