The video does not say just how much of it could be used in the same way as Vram, but you could potentially render even larger scenes with this thing than what your CPU can handle (though it seems the main point is to every piece of data you need on the card itself, including your scene files and applications).
Sure, the price will likely be an arm and a leg, but it looks like an incredible piece of tech. nonetheless.
I recall from earlier announcements that this GPU should have an on-board NAND-based SSD cache, so it would not exactly be the same as having 1TB of ordinary GDDR VRAM.
AIii, maan… 
smarta n’urons
& a quicky from web - source -
… Measuring the memory use of ResNet-50 training with a mini-batch of 32 on a typical high performance GPU shows that it needs over 7.5 GB of local DRAM.
…
That card was announced at last year’s SIGGRAPH. Is it shipping yet?
I think this is actually pretty deceiving. This card does not have 1TB of GDDR memory. It only has 32GB (not that that’s anything to sneeze at) but it has slots for 2 M.2 SSDs that can be configured to hold 1TB of quick access storage. It’s important to note that this is intended to be used with editing 8K footage. So, the frames are loaded onto the cards dedicated SSD so it can them play back in real-time in say, Premier Pro or DaVinci Resolve.
While I’m sure you could render on this card with Blender just fine, You’re not going to get any benefit of the TB of storage in Blender. I suppose someone could come up with a way of storing and retrieving large textures from it but honestly, I have a feeling that texture retreiveal from system RAM to video card RAM would still be faster and way cheaper (something I think they are already working on). Either way, it’s not very optimal nor, I think, necessary. The 32GB of RAM this thing has is a crazy huge amount of ram. Even 12 to 16GB is enough for most peoples work unless you are really bad at optimizing your scenes. 
Heck, until the nvidia 10 series we were all rendering with no more than 6GB (usually it was more like 3GB!). Now the standard is 8GB and I have yet to top out my two 1070’s.
But here’s the real kicker: $9,999!!! For that money, you could buy a nice beefy machine with 4 Titan X’s in it. Seriously, This card is for video editors working on 8K features films or new media/8k early adopter types.
source: http://postperspective.com/new-amd-radeon-pro-duo-graphics-card-pro-workflows/
Wow - big words, big numbers, big price… But little of actual value. Well for video editing there might be, but we are all here in Blender forum, so who cares?..
All Vega have a new HBCC memory controller… over the HBM ones, it permit direct access to other memory pool. The SSG version come with 2 TB of Nvme storage directly onboard ( need to be confirmed ), the thing is contrary of standard memory access as it have a dedicated cache controller in the gpu cores ( meaning extremely low latency access.
The thing is it allow to cache fast data on HBM2 and store long live data ( movie frame, textures ) directly on the Nvme pool ( note it can access directly a tons of memory pool ( can be SSD on your system etc ).
Well this gpu is more for edit 8K video content smoothly than render, or aimed at professional VFX studio or animation ( for extreme large scene content )
Now the short presentation was more about Vega Frontier, it was a financial analyst day, so not much information before the 31th but: 13Tflops FP32, 16GB HBM2…
I will wait the gaming version anyway.
http://pro.radeon.com/en-us/vega-frontier-edition/
Yeah, but there’s a difference between accessing and loading textures from on board SSD for use in games and accessing them from memory while rendering.
In the case of games, the texture is loaded at the beginning of the level or when first seen into GPU memory. It can be manipulated and rendered in large chunks. At some point it can be offloaded when it’s not being used.
In the case of ray/path tracing, the texture needs to be in ram the whole time. when a ray hits a surface with that texture on it, the engine needs to look into that texture to see what color the pixel should be. Then it may hit another surface and find a different texture and the same happens again. This happens thousands of times a second. If you tried to load the texture from an SSD, every time you hit that texture… Or even tried to look into that texture off the SSD, it would slow you down immensely. If your SSD is as fast as system or GDDR ram then yes, that would work just fine. But so far, that technology doesn’t exists.
