Quick question about HBM

[nj4ck]

OK guys, so I gotta ask what is probably a really stupid question, so forgive me it it is. 

 

So I saw this, related to the 380X:

 

 

Does 2.5D HBM mean that the memory and the die will be on the same unit, right next to eachother, or is the graphic misleading? 

[Guest]

As far as I remember, yes. They will be one "unit" so they don't have to lay long tracers from the GPU to the memory anymore.

[nj4ck]

Instead of memory chips strewn everywhere for ordinary GDDR, it will now be just one or two which greatly simplifies the circuitry needed.

 

As far as I remember, yes. They will be one "unit" so they don't have to lay long tracers from the GPU to the memory anymore.

 

OK, that's what I was hoping it meant. So if they have the GPU and VRAM on one unit, what's stopping them from slapping an additional CPU die on there, essentially creating a monster-APU? Sure, it would be a 300+W part and they would need a new socket, but apart from that...?

[Guest]

OK, that's what I was hoping it meant. So if they have the GPU and VRAM on one unit, what's stopping them from slapping an additional CPU die on there, essentially creating a monster-APU? Sure, it would be a 300+W part, but apart from that...?

 

The 300+W is the problem. Good luck dissipating all the heat that this will generate. Also, VRAM might have insane bandwidth but it's not as responsive as traditional RAM. There's a good reason why you're not using GDDR5 as your system memory after all. I don't know how HBM will do in comparison to GDDR5 in that regard, but I assume that it'll be similar.

[Briggsy]

OK, that's what I was hoping it meant. So if they have the GPU and VRAM on one unit, what's stopping them from slapping an additional CPU die on there, essentially creating a monster-APU? Sure, it would be a 300+W part, but apart from that...?

 

My guess is AMD are waiting for 14/16nm, or even 10nm to get the power consumption down. HUMA (heterogeneous unified memory access) has been AMD's goal since they acquired ATI (probably why they bought ATI) so they can ultimately build a high end SOC, and saturate the mobile market with low watt SOCs. But to have a monster SOC (system on a chip) like that would require a smaller process node then 28nm, or peoples faces would melt.

[nj4ck]

No you got it wrong. The GPU and memory are not on the same die.

 

If you're talking about HBM for graphics cards, who would want a CPU on the PCB?

 

If you're talking about HBM for the computer as a whole, it would come in the form of RAM sticks more than anything else.

 

Memory (as RAM) would never be put together with the main processor in the same die. It would be impossible as the processor manufacturer is not the memory manufacturer.

I'm talking about HBM VRAM, not system RAM. I know its not on the same die, but this wouldn't be the first time we'd see more than one die under the same IHS. And yes I know that my idea is stupid, I'm just wondering if bringing all the heat-emitting components as close together as possible and having them be cooled by a single massive cooler wouldn't be more efficient than the current solution. Dissipating 300+W of heat is after all not impossible with the types of coolers we have today.

[nj4ck]

The 300+W is the problem. Good luck dissipating all the heat that this will generate. Also, VRAM might have insane bandwidth but it's not as responsive as traditional RAM. There's a good reason why you're not using GDDR5 as your system memory after all. I don't know how HBM will do in comparison to GDDR5 in that regard, but I assume that it'll be similar.

No no no, the RAM would still be standard DDR3/DDR4, only the VRAM would be integrated onto the SoC. As to the heat output, that's obviously a problem. But my R9 290 also pulls easily 300W of power and would still stay at reasonable temperatures back when I was still air-cooling it. So I don't think the heat-problem would necessarily be the nail in the coffin of such a project, they'd just have to do better than the standard AMD boxed cooler  :lol:

[nj4ck]

I get your idea about a franken-APU with GPU, CPU and VRAM on the same die. But before we even get to the issue of heat, how do you suggest to have it manufactured?

 

HBM is memory that takes specialized manufacturer like Hynix to create, while the CPU/GPU would at best be made as an APU by processor specialist like GlobalFoundries.

 

Having everything on the same die would require all of them to be on the same process node and made by a single manufacturer.

 

And then the whole motivation for such a product would be space saving like for mobile devices. If it is going to output anything like 300W of heat it would incinerate any mobile device.

 

On a desktop it would make no sense to give up modular upgradeability just to have everything under one cooling solution.

Well I'm by no means an expert, so I have no clue how this would be manufactured. Its just an idea   But I wasn't talking about putting everything on the same die, I meant seperate dies on the same unit like the old Opterons. I'd guess that different manufacturers would make the seperate pieces of silicon which would then get fused together on a PCB to make a SoC. There are indications that AMD is working on an APU with HBM, so I think this will be the general direction things will be going in. I doubt that we'll be seing a 300W APU anytime soon, but as manufacturing processes shrink and power efficiency increases, soon a system that would now need 300W may only need 150W with the same performance. And that would be very managable, similar to a current Intel 8-Core.

I personally would certainly like that. It would allow for much more compact systems with less fans, less points of failure and higher performance. It would also reduce cost quite a bit, after all you wouldn't need the graphics card at all, just the GPU die.