Forbes did a good piece on this 1500 dollar monster.
They explained why its $1,500.
AMD is targeting the card at developers who want to develop (VR) games and play them too.
The card will have three sets of drivers, Beta Crimson, WHQL Crimson, and FirePRO.
So I guess its not really for most of us and is more like Nvidias Titan stuff, yeah it play games but it does other stuff better and its a billion dollars. 

I think anyone here would rather use two Nanos over this for gaming for just under $1,000.

Vlada011
I would like to price drop on 900$ when New Pascal show up.
But I don't know... It's scary to invest in some dual monster extra optimized for DX12 with only 4GB.
I mean better 30% less power and 8GB memory. Somehow I'm not convinced that HBM 4GB is same as 6-8GB GDDR5.
But card is great. 

I mean it's beauty compare to New GeForce reference Freak.
 

Not likely since it is an all in one card, with three sets of drivers, nvidia does not even offer that anymore... The Titan X does not do what the Titan/Black or Z nor offered the same options as the first Titans. If the card had a bigger radiator they would had been able to bump the clocks more. But most developers dont have room for a rad over 120. normally workstations have single 120mm fan slots front and rear. In DX12 games the 2x 4gb HBM giving you 8gb in all will not be an issue it will treat the gpu as on card and use all 8gb of ram, any game that is not coded correctly for DX12 then yes you will only have 4gb of vram. But I will say if a game is coded correctly for DX12, has a good LOD manager 4gb of hbm vram is not even an issue even at 4k but yes it does drop performance but not as much as those gpu's unable to use DX12 correctly and or have slower GDDR5 vram. I hope they open up the FirePro drivers for the Fury X, or someone makes a modified FirePro driver since the Fury X is the same chip as the Radeon Pro Duo. Right now though have had zero issues with Maya or my games with the Fury X. Best card I have owned since my GTX Titan. Best AMD card I have owned since my x800m.

ty_ger07

In DX12 games the 2x 4gb HBM giving you 8gb in all will not be an issue it will treat the gpu as on card and use all 8gb of ram, any game that is not coded correctly for DX12 then yes you will only have 4gb of vram.

Is there any evidence of that being true in the real-world, yet? The vram stacking seems like vaporware. I have heard this claim for a year.

Will see as the year presses on as we get true, correctly coded DX12 games, will be interesting as well want to see amd paired up with nvidia if both companies allow it to happen in a multi gpu config. Its not really vaporware just have to wait, it all falls on the devs and studios if it will be a thing or not its up to them to implement it into their games.

donta1979

ty_ger07

In DX12 games the 2x 4gb HBM giving you 8gb in all will not be an issue it will treat the gpu as on card and use all 8gb of ram, any game that is not coded correctly for DX12 then yes you will only have 4gb of vram.

Is there any evidence of that being true in the real-world, yet? The vram stacking seems like vaporware. I have heard this claim for a year.

Will see as the year presses on as we get true, correctly coded DX12 games, will be interesting as well want to see amd paired up with nvidia if both companies allow it to happen in a multi gpu config. Its not really vaporware just have to wait, it all falls on the devs and studios if it will be a thing or not its up to them to implement it into their games.

I don't think it will happen. At least not unless a new crossfire/sli bridge is released. The PCI-E and bridge bandwidth is way too low for cards to share memory resources between eachother. We are talking bandwidth between the cards being 10x or more slower compared to the memory bandwidth of the card. Performance would plummet as soon as they attempted it. If we are going to avoid the bandwidth problem, the cards will have to share duplicated texture resources (not stacked), a bridge between the cards of 200 GB/s or higher will need to be developed, or crossfire/sli memory stacking will have to be limited to dual GPU cards where a high speed interconnect is designed which doesn't leave a shared PCB.

Whatever the outcome is, I am fairly confident that we will never see present models of GPUs with fully-working stacked memory abilities in the future. If anything, the present models may be able to stack a few hundred MB of memory, but the majority of the memory will still be duplicated (not stacked).

I have seen people say "you're missing the point" when making the argument I just made. They say "the cards will work seperately on their own task and won't need to transfer lots of data back and forth". The problem is that if the game has 3 GB of textures, each card is going to need to waste 3 GB of space saving its own duplicated copy (not "stacked") of that texture data since it can't store the 3 GB textures on only card A and expect card B to be able to access a texture on card A fast enough through the existing PCI-E system or bridge. So, the only information which absolutely cannot be stacked at the moment is the information which happens to take up the most space.

If stacking is incorporated using current hardware, stacking will save hundreds of MBs, not GBs.