Upcoming GPUs (2016)

[MiRai]

I think it's going to be 1200 for the 12gb ver and 1600 for the 16gb ver for titan.

Two reasons why I don't believe this to be true...

1) The market segment for the Titan GPUs is already incredibly small, and splitting them up into two, even smaller segments is probably going to hurt profits, more than help. I doubt nVidia is going to bother to spend extra money in order to manufacture two different types of ultra-top-end GPUs just to try and price gouge people at $100 per GB of VRAM.

2) The amount of VRAM on the PCB is determined by the GPU's memory interface:

GDDR5/X 256-bit = 2/4/ 8/16/32GB
GDDR5/X 384-bit = 3/6/12/24/48GB
HBM2 4096-bit = 4/8/12/16/20/24GB? (I'm guessing here, but it could also be 4/8/16/32GB like we see with 256-bit)

Now, all of the articles say that nVidia is going to "unveil" this new GPU, but that could mean anything. It could mean only spoken words with a picture on screen with no model or solid release date in sight, or a non-working mock-up model of the GPU, or a fully working GPU... we have no idea.

However, if nVidia is going to actually unveil a new Titan GPU at GamesCom and announce a release date in the near future, then my guess is that it's going to be G5X memory, and not HBM2. HBM2 isn't even available in the wild at this time, so for nVidia to offer it on a consumer-level GPU is absurd, because they release their GPUs like this: Tesla (Supercomputing/Deep Learning) > Quadro (Workstation) > GeForce (Gaming). I highly doubt that nVidia is going to cut into the super-scarce supply of HBM2 just so they can charge pennies on the dollar for it, when they could easily be making thousands of dollars per GPU in the Tesla or Quadro market.

So, with that said, if the GPU is going to be using the standard G5X VRAM, then it's either going to be 256-bit or 384-bit, and you can't have both 12GB and 16GB models on the same interface.

[mbox_bob]

2) The amount of VRAM on the PCB is determined by the GPU's memory interface:

GDDR5/X 256-bit = 2/4/ 8/16/32GB
GDDR5/X 384-bit = 3/6/12/24/48GB
HBM2 4096-bit = 4/8/12/16/20/24GB? (I'm guessing here, but it could also be 4/8/16/32GB like we see with 256-bit)

Err, the GB is determined by the density of the VRAM chips. The interface determines the number of chips.
GDDR5(X) chips have a 32bit interface, so you take you GPU interface, divide by 32 and that is the number of VRAM modules you will have on the PCB.
HBM1 is 1024 bits wide.

It then becomes a simple bit of math. gpu interface / VRAM chip interface * density / 8.

Funnily enough, the top end graphics cards tend to use the highest density chips available at the time, which is currently 8Gb (gigabit, not bytes), but it is not mandatory.

/e: it is possible to put more vram on the board, but this would then require the memory controller to switch between modules, so you would not gain any extra "bandwidth" out of doing so, and would probably then generate extra latency (in fact this is what the GTX 970 does, so it's 256bit interface is just marketing gumpf because it is not a true 256bit pipeline throughout the whole shebang).

/e2: The Titan-X (because someone will bring it up), does the memory controller switch to access the 24 VRAM modules (it has 4Gb modules), the internals of how it does it without impacting significantly on performance, who knows, and why the 970 does it so inefficiently in comparison, who knows. My guess is that the Titan-X does not switch to accessing a single module at a time, it is always addressing 12 modules simultaneously, whereas the 970 goes from addressing 7 modules to 1 module, then back to 7 etc (although only when it really needs that extra 512MB).