Seeing that each 1GB stack of HBM offers 1024bit/1Ghz/128GB/s sisoft could easily read it as 4096bit, because it is.
Likewise, people keep ignoring despite recent proof(Tonga, Maxwell), memory efficiency isn't a static thing, memory usage isn't only increasing. Tonga/Maxwell have more efficient bandwidth because it is compressing textures, and very well, this means less data is transferred over the memory bus. Call it what used to be 1GB of textures is now compressed and only uses 700MB of space, so needs 30% less bandwidth to transfer that in the same amount of time. Memory efficiency means for the same resolution that Maxwell/Tonga would use less memory than Hawaii/Kepler.
This will change the goal posts on higher resolution requirements for memory, but then most people don't use 4k.
ultimately if it did use HBM you're limited by the numbers of stacks you can use, each extra stack decreases yields significantly. You stick a fully known working chip and 4 fully working stacks of memory on another basically very thin silicon chip and connect them together, if one connection fails(of thousands) then you have multiple dead chips. Yields are effectively exponential scale where every new chip decreases yields at an exponential rate, so putting 8 x 1GB chips is going to get incredibly expensive very very quickly.
We likely won't see 8 stacks in the future but 4 stacks with higher memory capacity per stack.
It depends, if they are HBM then 8GB probably would be read as 4GB.
given that they are dual stacks and all...
HBM are dual stacks... what, they aren't.
HBM is a stack of four memory chips, each 256gbit capacity, or 256MB or 1GB per stack, there is a move to double density but it wasn't likely to be the first round of products and it would increase cost obviously.
LIkewise each stack offers 128GB/s bandwidth though that is listed with 1Ghz memory clock speeds. There would/will be future increases in clock speed also.
The difficulty in knowing precisely what we'll see from HBM is time to market.
If there was a time scale release 2 years ago that say first iteration HBM is 1gb stacks 128GB/s using 1GHz clock speeds, the second step is up clock speeds, then up density, then second gen will be 2GB stacks with 256GB/s bandwidth per stack.... we don't know how the memory availability dates stack up against products. The first GPU to use them may come after the higher density stacks are available or they might use the first iteration of HBM. It's most likely to start off conservatively because in general memory speeds/capacity/speed don't move on till volume production is well under way and volume production needs a volume product to go with it.
