Speculation: AMD's APU with HBM

[Vattila]

Due to the popularity of my speculation series, and spurred by Mockingbird's recent promotion of the series — amusingly speculating on the series itself — I have scanned my brain for upcoming and exciting technology about which I could poll the knowledgeable forum members here. My previous threads in the series have generated a lot of interesting discussion.

So here goes: What will AMD's "super APU" look like? This has intrigued me for a while now. By "super APU", I mean an APU with high-bandwidth memory (HBM), thus fully enabling the processing potential of an APU. Current APU performance is severely limited by memory bandwidth.

A super APU would be great for many markets, including notebook, small-form-factor and thin-client desktop, server/supercomputer, and embedded (consoles, in particular).​

We have seen what a super APU may look like and how it may perform through AMD's partnership with Intel on the Kaby Lake-G chip. This chip has a H-series Intel CPU connected via PCI-Express to an AMD GPU with 4 GB HBM connected via Intel's EMIB, all integrated on a single package. That AMD agreed to this partnership probably means that their own all-AMD solution will not be ready until 7nm at the earliest. But we know they have all the ingredients for a successful product, and we know high-performance APU is a big part of their exa-scale research.

I have previously speculated about what AMD's server APU may look like in an earlier thread (here). If AMD can crack the problem of making multiple GPU complexes ("GCXs") function as a single GPU, such a chip could also be very attractive to the gamers in the HEDT market, as a 4-die ThreadRipper APU based on it would pack quite a punch.

My full speculation series.

 

[Obvcop]

Surely If AMD released a product like this it would be in the far future and not on AM4. Anything with remotely decent performance is going to have far too high a power budget for most Am4 boards. If you build it on threadripper it's starting to turn into a VERY niche product and again you need new motherboard with Video out paths. If anything like this happens then surely it will be on a soldered socket and exclusive to something like the PS5 or new XBOX

 

[Vattila]

If anything like this happens then surely it will be on a soldered socket and exclusive to something like the PS5 or new XBOX

Do you not even believe they will introduce an all-AMD direct competitor to Intel's Kaby Lake-G?

 

[Obvcop]

hence why I said 'soldered socket', but it doesn't look like there is any sign that something like this is in the works so far. Intel are struggling to penetrate the market with kaby lake G and nvidea are able to keep them shut out of gaming brands. It makes zero financial sense to create a product with such a high risk like this

 

[whm1974]

hence why I said 'soldered socket', but it doesn't look like there is any sign that something like this is in the works so far. Intel are struggling to penetrate the market with kaby lake G and nvidea are able to keep them shut out of gaming brands. It makes zero financial sense to create a product with such a high risk like this

I agree as such a product like this would only meet a niche in high end laptops and SFF desktops smaller then mITX.

 

[Vattila]

Intel are struggling to penetrate the market with kaby lake G

So you think Kaby Lake-G will fail as a product?

I've read great reviews of Ultrabooks based on it, e.g. this PCWorld review.

"Today though, there are few takers of Kaby Lake G. In fact, only two vendors have shipped it: HP with the Spectre x360 15, and Dell with its XPS 15 2-in-1. Some we’ve spoken to have painted that as a failure of Kaby Lake G to catch on, while others have speculated politics to be the cause. Whatever the truth, it’s a shame, because applied the right way in the right laptop, Kaby Lake G is a road worth taking."

 

[Vattila]

it doesn't look like there is any sign that something like this is in the works so far

We do have the persistent Fenghuang rumour, though.

PS. June 2018 update.

"AMD is working on a lot of internal projects that have yet to see the light of day, one such project is the Fenghuang 15FF graphics chip for accelerated processing units. The Fenghuang 15FF APU was spotted last year in an entry within SiSoftware Sandra database but since then, no other info made its way on the web until today."

 

[IntelUser2000]

So you think Kaby Lake-G will fail as a product?

I don't think it'll fail, it just won't be successful as people think.

The point of integration(until now) has been about low cost, and lowering power. It has never been about performance.

If you go even higher end like the Threadripper replacement, then for client it would only be for high end desktops. Then it must perform. Such a high end device only makes sense for workstation because the TAM is larger.

Kabylake-G successor must solve:
-Battery life issues. Heck it has the HD 630 die. The battery life is closer to a discrete GPU one than an iGPU one
-CPU/GPU generational deficiency. CPU part came slightly before 6 core Coffeelake. A 6 core Coffeelake-G would have been far better. GPU is barely satisfactory. About half a gen behind

The execution is far better, but has much the same problem as Iris Pro did. Iris Pro died not because of politics, but because Nvidia Maxwell-based parts performed better, cost about the same, and had better battery life(due to switching).

Maybe Nvidia has to do little with it, but I assume the faults are 15% Nvidia, 85% technical.

 

[ericlp]

I wonder... Think there could be any room for a dedicated NPU/TPU? Is it time for every chip to have one? Kinda like an FPU that is common in every modern chip since the 486.

 

[LightningZ71]

I wonder about AMD maybe introducing a platform that can be common between consoles and notebooks/convertibles/uSFF desktops/AIOs. With TVs pushing past computer monitors into 8k resolution going forward, future consoles will have to look for greater performance than what they had in previous generations with respect to the GPU side of things. They will also look to continue the trend towards making their devices smaller and more energy efficient.

I think that products in the same format as Kady Lake G will offer the best road forward. It allows the entire thing to have a constant form factor to keep platform costs down while allowing AMD to mix and match components as needed by changing the GPU chip as needed, changing the hbm stack as needed and changing the processor itself as needed without having to respin the entire monolithic die. Now, will it be an mcm, or emib, or something in between? No idea.

 

[Vattila]

It allows the entire thing to have a constant form factor to keep platform costs down while allowing AMD to mix and match components as needed by changing the GPU chip as needed, changing the hbm stack as needed and changing the processor itself as needed without having to respin the entire monolithic die.

Good points. And since everything is on the package, it allows rapid improvements of the interconnect, as needed for optimal efficiency and elimination of bottlenecks.

 

[Vattila]

Think there could be any room for a dedicated NPU/TPU? Is it time for every chip to have one?

Too soon, I guess. We need killer applications and the best hardware to run them to be established first. Machine learning is still in the early days with lot of things yet to be tried, I feel.

It may make sense for an HPC APU in the server market, though.

 

[frozentundra123456]

Where is the option for "It will be a long time, if ever, before we see a consumer APU with HBM"?

 

[Vattila]

Where is the option for "It will be a long time, if ever, before we see a consumer APU with HBM"?

Timing is not considered in the poll. Just go with what you think will be the first available product, whenever that may be.

 

[whm1974]

I wonder about AMD maybe introducing a platform that can be common between consoles and notebooks/convertibles/uSFF desktops/AIOs. With TVs pushing past computer monitors into 8k resolution going forward, future consoles will have to look for greater performance than what they had in previous generations with respect to the GPU side of things. They will also look to continue the trend towards making their devices smaller and more energy efficient.

I think that products in the same format as Kady Lake G will offer the best road forward. It allows the entire thing to have a constant form factor to keep platform costs down while allowing AMD to mix and match components as needed by changing the GPU chip as needed, changing the hbm stack as needed and changing the processor itself as needed without having to respin the entire monolithic die. Now, will it be an mcm, or emib, or something in between? No idea.

Sounds like this would be a high end niche product that only appeal to folks who want the most performance in small packages, and willing to pay for it.

 

[whm1974]

Where is the option for "It will be a long time, if ever, before we see a consumer APU with HBM"?

Maybe never?

 

[Vattila]

Maybe never?

Don't be such a pessimist.

 

[whm1974]

Don't be such a pessimist.

What is the use case? Well designed small mITX cases exist that will still take a decent video card which will make such a product almost pointless.

 

[Vattila]

What is the use case? Well designed small mITX cases exist that will still take a decent video card which will make such a product almost pointless.

Do I detect some iGPU disgust?

 

[whm1974]

Do I detect some iGPU disgust?

iGPUs have their place, but the best performance will always be offered by video cards.

 

[Vattila]

iGPUs have their place, but the best performance will always be offered by video cards.

Well, if iGPUs have a place, you would, from a design perspective, try to implement them as efficiently as possible, which means eliminating the footprint, bandwidth bottleneck, access latency and power drain of an external memory interface — which leads you to the obvious plans for AMD's APU design.

So I presume your "maybe never" remark just is a declaration of indifference to this product category. Which is fine, by the way. Each to their own.

 

[whm1974]

Well, if iGPUs have a place, you would, from a design perspective, try to implement them as efficiently as possible, which means eliminating the footprint, bandwidth bottleneck, access latency and power drain of an external memory interface — which leads you to the obvious plans for AMD's APU design.

So I presume your "maybe never" remark just is a declaration of indifference to this product category. Which is fine, by the way. Each to their own.

Yes but video cards can designed to be more efficient as well and have higher performance and are able to have more video memory as well.

 

[IntelUser2000]

Maybe never?

Since Kaby G is out, I wouldn't say that.

But if you want to put it into categories, lets see.

1) AMD APU with HBM
2) APU with HBM that's common, like Raven Ridge

#1 is probably likely in the near future, but similar to Kaby G it should remain high end. #2 is far less likely. Separate memory stacks will always cost more. The problem in the value space is always that extra packaging adds cost, even if the memory stack is very cheap.

Regular process updates with big gains(unlike today) is what enabled CPUs to integrate everything and we got decades of better performance, lower power, and lower cost, but that's changing. That's why we're moving into specialized circuits for some segments.

Thus, the trade-off between performance and cost has to be considered. The distinction will always exist.

 

[whm1974]

Since Kaby G is out, I wouldn't say that.

But if you want to put it into categories, lets see.

1) AMD APU with HBM
2) APU with HBM that's common, like Raven Ridge

#1 is probably likely in the near future, but similar to Kaby G it should remain high end. #2 is far less likely. Separate memory stacks will always cost more. The problem in the value space is always that extra packaging adds cost, even if the memory stack is very cheap.

Regular process updates with big gains(unlike today) is what enabled CPUs to integrate everything and we got decades of better performance, lower power, and lower cost, but that's changing. That's why we're moving into specialized circuits for some segments.

Thus, the trade-off between performance and cost has to be considered. The distinction will always exist.

Thanks, the last sentence is exectly my point. Which is why I started out with a decent performing CPU and a mid-range GPU. That way If I need more graphics performance later I will just need to replace the card and not my CPU, motherboard, and memory. With something like Kaby Lake-G I will have to do the later.

 

[PeterScott]

I think most of the polling options are inherently wrong in their definition.

Small monolithic dies aren't going to be "Super" APUs.
Separate CPU and GPU dies are not an APU at all.

By definition, a Super APU would be a fairly large monolithic die, mostly for the large GPU needed to make it "Super".

If you want to see a Super APU, look at the Xbox 1 X:
Transistors: 7,000 million
Die Size: 359 mm²
That's what a super APU looks like.

I am dubious that AMD will build a "Super" APU for PCs at all.

I think they will go the Intel Kaby G route for the GPU intensive niche. But that isn't an APU. It's just a discrete CPU and discrete GPU packaged on a small PCB together.