GTX 1080 unveiled; 9 teraflops

[JEMC]

Pemalite said:

JEMC said:

The big problem of HBM is that using the 1024-bit bus per stack take so much traces that AMD and Nvidia are only able to use 4 modules. Even the GP100 was only using 4 stacks, but of HBM2.

I wonder what that HBM successor might be. Has anyone heard anything about it?

That APU you talk about... I find it hard to believe it, to be honest. Not only because the first mention of it came from Fudzilla (at least, if you're talking about that 16 Zen cores + Greenland + HBM rumor) but because not even AMD would launch such a power hungry APU. It would need to come bundled with a CLC unit and that would make it too much expensive to make it a worth buy over a separate CPU+GPU, that could use less power and be even faster.

The traces aren't really the problem. That is what the interposers are for.
However... Interposers are also built at 65nm and they do cost a fair chunk of coin, you could technically have two stacks of HBM to each interposer for a total of 8 if the Interposer was designed to accomodate it, but they aren't. It's one thing to do HBM version 2, doing an Interposer version 2 on a different fabrication process in order to accomidate multiple stacks is a different game entirely.

But then Costs will also blow out, only so much you can do per market segment, HBM is already expensive.

As for HBM's successor, AMD has just labelled it as "Next Gen memory". I would hazard a guess that as HBM is seen as "wide" AMD/JEDEC/Memory companies might go for a faster approach and take advantage of HBM's inherent advantages and drive up the speed.

As for the APU.
Anandtech at one point asked it's reader base what users wanted to see out of AMD and AMD was actually reading and responding to some of the comments... And without question everyone wanted a beefy APU with built-in GDDR5 or better Ram, AMD had actually done something similar with the 780 chipsets where they would bundle 64Mb-128Mb of DDR3 Ram for the motherboards graphics.
So AMD knows there is demand, at-least in the enthusiast community.

Now the APU in question isn't for your every day consumer, it is for the HPC market, Aka. Server grade stuff that already uses 16+ CPU cores and beefy GPU's for compute, it just happens to be that AMD sells both seperate one as Opteron the other as FireGL.
AMD themselves have already confirmed the existence of such a product, but what kind of hardware still remains to be seen, but keep in mind it is for the HPC market, it's likely to be extremely chunky, they don't typically do low-end stuff. ;)

Thus... Cooling isn't going to be an issue, these will likely come complete as a "compute add-in board" or an all-in-on motherboard with tons of ports for drives.

And even then, you don't need to go liquid cooling to cool 200-300w, a motherboard allows for a larger surface area than a GPU and GPU's have been cooling such wattages like that for years thanks to heat pipes, vapour chamber etc' designs.

Cooling is an issue when all those 200 or 300W of heat come from a single chip the size of an APU. There are no CPU heatsinks designed to tame such a monster, which is why I talked about AMD having to go with a CLC unit to cool such thing.

But it doesn't really matter, because such an APU for the HTC market won't go into retail for us to buy it.

That said, there was a rumor a couple of months ago of AMD working on a Bristol Ridge APU with 16CUs, that would put it almost on the same level as an XboxOne (because with DDR4, the bandwidth would be quite lower). For something more powerful, we'll have to wait for Zen and Pascal, well into next year.

*Edit*

I found the rumor: http://wccftech.com/amd-bristol-ridge-16-cu-apu/