Speculation: AMD's APU with HBM

[GodisanAtheist]

I think PeterScott is on the right track here stating the first "Super APU" will come to the consoles.

In some form, it already has with the PS4 using GDDR5 as both system and GPU ram. Not a leap at all to assume HBM taking the place of the GDDR5 and everything getting crammed into one SOC package to keep the console's size to a minimum.

For desktops, it would be nice to see AMD pursue a large L3 cache strategy like Intel did with (I think) their Iris Pro line. AMD's GPU IP has a lot more headroom than Intel's and it would be great to see what they're really capable of in non-bandwidth constrained scenarios.

 

[William Gaatjes]

I read this interview with Jim Keller, and i found this excerpt interesting :
https://www.anandtech.com/show/13048/an-anandtech-exclusive-the-jim-keller-interview

Dr. Murthy Renduchintala – Intel’s Chief Engineering Officer – said that the company has “embarked on exciting initiatives to fundamentally change the way we build the silicon as we enter the world of heterogeneous process and architectures,” which may been seen as a hint of Intel’s future direction.

Heterogeneous architectures can mean a lot of things.
But it is interesting that now Raja Koduri is working at Intel for the GPU division and Jim Keller for the SOC.
For fast x86 cores, Intel already has good people.
Makes me wonder if Jim Keller when at AMD was also a driving force behind HSA next to Phil Rogers who was the HSA guru at AMD but went to work for Nvidia.
But then again, having for example 4 fast power consuming cpu cores and 4 slow but very low power cores on the same die that function as one can also be called heterogeneous computing.

 

[IntelUser2000]

I think PeterScott is on the right track here stating the first "Super APU" will come to the consoles.

In some form, it already has with the PS4 using GDDR5 as both system and GPU ram. Not a leap at all to assume HBM taking the place of the GDDR5 and everything getting crammed into one SOC package to keep the console's size to a minimum.

PS4 chip is already a Super APU. Doesn't need HBM to be classified as such.

 

[GodisanAtheist]

PS4 chip is already a Super APU. Doesn't need HBM to be classified as such.

True, I guess I figured that in order to be a "Super APU" the RAM had to be on package for some reason.

But I guess having GDDR5 on the board would eliminate all the problems putting RAM on package is supposed to solve...

End of discussion then? AMD's first Super APU will be the PS4 APU, which has been out since 2013.

 

[Batboy88]

I think PeterScott is on the right track here stating the first "Super APU" will come to the consoles.

In some form, it already has with the PS4 using GDDR5 as both system and GPU ram. Not a leap at all to assume HBM taking the place of the GDDR5 and everything getting crammed into one SOC package to keep the console's size to a minimum.

For desktops, it would be nice to see AMD pursue a large L3 cache strategy like Intel did with (I think) their Iris Pro line. AMD's GPU IP has a lot more headroom than Intel's and it would be great to see what they're really capable of in non-bandwidth constrained scenarios.

Do it...That's their stuff the Big Cache.

 

[Vattila]

True, I guess I figured that in order to be a "Super APU" the RAM had to be on package for some reason.

In the context of this discussion, I defined "super APU" to mean an APU chip package with HBM aboard. See OP.

End of discussion then? AMD's first Super APU will be the PS4 APU, which has been out since 2013.

No, for this discussion, let's only consider HBM on package. Your earlier replies opine that this would likely be a console chip, and I think that is a pretty good prediction. Now, what about mainstream PC? I think a direct competitor to Kaby Lake-G, primarily for mobile and small-form-factor (SFF desktop, all-in-one, thin client, embedded, etc.), is likely to be first out of the gate.

 

[PeterScott]

No, for this discussion, let's only consider HBM on package. Your earlier replies opine that this would likely be a console chip, and I think that is a pretty good prediction. Now, what about mainstream PC? I think a direct competitor to Kaby Lake-G, primarily for mobile and small-form-factor (SFF desktop, all-in-one, thin client, embedded, etc.), is likely to be first out of the gate.

Then we are back to what you consider an APU. IMO Calling Kaby-G an APU waters down the term to the point of meaningless. Since it is just a normal CPU attached to a dGPU via PCIe. Putting them together on their own mini PCB before attaching it to the MB is rather inconsequential.

AMD could do something like this, but if they are sitting back and letting Intel test the waters, they can't be impressed with the lack of design wins for Kaby-G, and AMD really couldn't be expected to do better at getting design wins given the disparity in their marketing presence.

Bottom line AMD could do Kaby-G type thing, but it wouldn't be an APU, and given Kaby-G's failure to get design wins, I am thinking this is getting less likely as time goes by.

 

[Vattila]

Then we are back to what you consider an APU.

I hear you, but bear with me. For this discussion, I wanted to focus on the integration benefits — primarily the bandwidth issue — hence HBM on the package, which seems to be the obvious path to attack this. For the question and poll in the OP, consider the package as a black box. And for simplicity sake, let's call this (CPU, GPU + HBM in a package) a "super APU" in this context.

That said, I do agree with you, in that the design space in packaging is immensely interesting, with upcoming "chiplet" design methodology and packaging technologies such as 2.5D (silicon interposer), 2.1D (organic low-cost interposer), and other promising technologies (3D, bridge/EMIB, fan-out, etc.). A good source of information on these developments is Semiconductor Engineering:

https://semiengineering.com/category-main-page-packaging-test-electronic-systems/

Feel free to discuss all of this, but for the purpose of the question and poll in the OP, let's stick to my definitions here.

 

[PeterScott]

Feel free to discuss all of this, but for the purpose of the question and poll in the OP, let's stick to my definitions here.

NO, your definition is flat out wrong, so I will continue to disagree with it.

Kaby Lake G type mini PCB combo doesn't give you anything you don't get from a statndard CPU and discrete GTX 1050ti. Except I think the latter costs less and performs better.

There is NO integration at all between CPU and GPU except that they ride on the same PCB before they get attached to the laptop MB. Which is completely irrelevant.

IMO if you want to stretch the definition of APU to separate chips riding on the same package, they should at minimum be sharing memory bandwidth, which Kaby-G type designs don't. It's just a tradition dCPU and dGPU, each with independent memory systems, connected via PCIe.

Functionally Kaby-G type designs are not really different that GPU card in a tower PC, and is not by any reasonable definition, and APU.

 

[whm1974]

NO, your definition is flat out wrong, so I will continue to disagree with it.

Kaby Lake G type mini PCB combo doesn't give you anything you don't get from a statndard CPU and discrete GTX 1050ti. Except I think the latter costs less and performs better.

There is NO integration at all between CPU and GPU except that they ride on the same PCB before they get attached to the laptop MB. Which is completely irrelevant.

IMO if you want to stretch the definition of APU to separate chips riding on the same package, they should at minimum be sharing memory bandwidth, which Kaby-G type designs don't. It's just a tradition dCPU and dGPU, each with independent memory systems, connected via PCIe.

Functionally Kaby-G type designs are not really different that GPU card in a tower PC, and is not by any reasonable definition, and APU.

I find myself in strong agreement with you on this. And the major downside to Kaby-G type designs is the user is required to replace both the CPU and GPU at the same time, instead of only upgrading what he needs.

 

[Vattila]

Makes me wonder if Jim Keller when at AMD was also a driving force behind HSA next to Phil Rogers who was the HSA guru at AMD but went to work for Nvidia.

As a programmer, I'd love to see a return to the focus on HSA. Would be great if Keller signs up to the HSA Foundation, and Koduri signs up to GPUOpen/ROCm (part of which builds on HSA work, I think), so that we could establish open standards for heterogeneous designs and accelerated workloads.

 

[Vattila]

Design and Analysis of an APU for Exascale Computing

http://www.computermachines.org/joe/publications/pdfs/hpca2017_exascale_apu.pdf

 

[Vattila]

A paragraph from that paper, that I find particularly interesting, regarding HSA/ROCm and programmability:

"The CPU and GPU must be architected in a cohesive manner from an energy efficiency as well as programmability standpoint. To this end, Heterogeneous System Architecture (HSA) compatibility is one of the major design goals of the APU. HSA provides a system architecture where all computing elements (CPU, GPU, and possibly other accelerators) share a unified coherent virtual address space. This enables efficient programming and computation via several mechanisms like free exchange of pointers by both CPU and GPU code, eliminating expensive data copy operations, transparent management of CPU and GPU caches via cache coherence, task offloads by both CPU and GPU to each other or other CPU/GPU units [13], and efficient synchronization mechanisms. These features are supported by AMD’s Radeon Open Compute platform (ROCm) to improve the programmability of such heterogeneous systems."

 

[Gideon]

I wonder what happened to Vega Mobile? With 4 more CUs and Vega architecture It seems to be a way better product than the stupid combo that Kaby-Lake-G is. And it would work with any CPU. I see zero benefits of doing things the way that is done on the intel chip.

Also if heterogeneous computing ever does take of, IMO it would make much more sense (considering manufacturing and product lines) to have a smaller dedicated GPU on chip (e.g. Raven Ridge) for latency sensitive stuff and a separate beefier GPU (e.g. Vega M) for other tasks, if needed.

 

[William Gaatjes]

Then we are back to what you consider an APU. IMO Calling Kaby-G an APU waters down the term to the point of meaningless. Since it is just a normal CPU attached to a dGPU via PCIe. Putting them together on their own mini PCB before attaching it to the MB is rather inconsequential.

AMD could do something like this, but if they are sitting back and letting Intel test the waters, they can't be impressed with the lack of design wins for Kaby-G, and AMD really couldn't be expected to do better at getting design wins given the disparity in their marketing presence.

Bottom line AMD could do Kaby-G type thing, but it wouldn't be an APU, and given Kaby-G's failure to get design wins, I am thinking this is getting less likely as time goes by.

I am wondering, for how much is a kabylake-G mcm module sold ?
I think you have a point that it is more a demonstration from Intel to let the public know that EMIB works, especially for HBM and as an example, even with products from the competitor.
I wonder what the price would be for a single die AMD ryzen+vega APU with hbm on board on a emib mcm module technology in comparison to the same apu single die apu with hbm on silicon interposer technology.

If Intel is asking a very high price for kaby lake-G, then it makes sense not much people wants to design/build a system based on kaby lake -G.

If Intel is asking a very high price for building emib devices for other manufacturers, it is not going to take off except for Intels own use.
If Intel would want to become EMIB an competitve industry standard, they would have to keep the margins a bit lower in order to establish traction.