Should AMD make a "console killer" APU?

[cytg111]

Yes,technically.
But inside a console not so much,they start with a TDP in mind and figure everything else out later,main concern is the power supply that often is an external unit because it won't fit or would produce too much heat inside the console.

Got it, the "console" of course would have to grow in size.

 

[tential]

It makes no sense to do so. When a person needs to replace their APU when the GPU power isn't enough, theye also have to replace the CPU. It just doesn't make sense on so many levels.
There is a reason APUs haven't taken off on the PC for gaming. GPU power is needed at a faster rate than CPU power, thus you waste your money getting an APU in which the GPU portion gets outdated while the CPU portion is still viable to use. Hence why people split them for PC gaming.

APUs need to be fast enough for basic gaming and games like LoL, Dota 2, CS, etc. That's it.

There is no business rationale for building such a part.

 

[laamanaator]

Intel's solution to this is eDRAM, which seems to work decently (although it's expensive, and Intel's GPU drivers are awful). There are other options for on-package memory too, including HBM. Adding a single stack of, say, 2GB HBM would more than make up for this. It would increase the price noticeably, though, with the interposer and all that. Even 1GB of some sort of high-bandwith memory would significantly improve this situation.

HBM and eDRAM could be a quite good solution for the bandwith starvation of iGPUs, but as you stated your self, they cost too much. And because of that high price we probably won't see that kind of memory configuration for some time, if ever.

 

[Valantar]

It makes no sense to do so. When a person needs to replace their APU when the GPU power isn't enough, theye also have to replace the CPU. It just doesn't make sense on so many levels.
There is a reason APUs haven't taken off on the PC for gaming. GPU power is needed at a faster rate than CPU power, thus you waste your money getting an APU in which the GPU portion gets outdated while the CPU portion is still viable to use. Hence why people split them for PC gaming.

APUs need to be fast enough for basic gaming and games like LoL, Dota 2, CS, etc. That's it.

I don't necessarily agree. If a solution like this was cost effective enough, it could open the door to relatively simple and inexpensive (decent performance) pre-built desktops and simple (few parts/little assembly) DIY solutions. Also, there's the obvious upgrade path of adding a dGPU, which would let you keep a high performance 4C8T CPU (now even less thermally constrained when the iGPU is less used), and keep the iGPU for DX12/Vulkan offloading or other compute tasks.

Also, the idea of cooling your entire computer with a relatively inexpensive tower heatsink or 240mm AIO is wry appealing - to me, at least.

Also, a solution like this could open the door to trying out AAA gaming with entry level PCs, without having to invest in a dedicated GPU. Current iGPUs simply can't do this. Heck, my A8-7600 - which has what is a relatively good i GPU by today's standards - can't even play Rocket League at decent settings. This just doesn't cut it in 2016, let alone 2017.

 

[ELopes580]

Not sure if anyone has mentioned yet, but isn't this what the whole Steam box is? How has that performed in the marketplace?

 

[DrMrLordX]

Okay, you're clearly not thinking of what I intended with this thread. 8c16t Zen + Vega = at least 200W TDP, and at the very least $6-700. An APU to compete with this would be DOA, no matter how compelling its performance - less upgradeability, no sensible way of cooling it ... No thanks.

Are you sure about that? A "console killer" is basically going to be AMD's best GPU tech crammed onto the APU die, and in a few weeks/months that's going to be Vega. It doesn't have to be the full monty but it does have to be more shaders than what's in the PS4 Pro/XBox One APUs today. PS4 Pro already has 36 CUs on it.

And to build off what The Stilt said it would need something like HBM or one hell of a fat l4.

I am not exactly sure what would be the most marketable option for AMD to approach something like this, but with so many of the motherboard's functions going onto SoCs already, it might almost make sense to sell APUs like that soldered in BGA format onto a board (or daughtercard) with high-speed memory soldered right on there. The only thing you would configure yourself would be a bank of DDR4. Maybe have HBM2 soldered onto the board with RAM slots elsewhere? And you just buy the whole thing wholesale. If the rumours are true about the UEFI being stored on the SoC itself . . . which granted, is unproven at this point. Anyway, it would make sense. If you want to upgrade, you upgrade the board/daughtercard all at once, getting a new APU with more/better CUs and different configuration of HBM2 (or whatever).

What I'm asking for is something (significantly) better than the current high-end iGPUs. The CU stagnation in AMDs APUs in recent years is getting to look pretty bad, and the move to 14nm should prompt SKUs with a significant increase in CUs. The current state of iGPUs is a rather sad state of affairs, and leaves a rather large performance gap between $100 GPUs and iGPUs.

Okay, look at Snowy Owl and ask yourself: how can consumers leverage something like that?

AMD's challenge to market a "fat" APU is that - again, to go back to what The Stilt said and to return to my original reply to this thread - consumers can just choose a standard AM4 CPU + dGPU. If you look at the cost/benefit analysis the APU is probably never going to be there unless devs start really utilizing OpenCL2 (HSA, hah!) for low-latency iGPU-based compute functions. Then the "serious" enthusiast can choose fat APU + Vega over Summit Ridge + Vega for gaming or . . . whatever.

Now granted the whole BGA motherboard/daughtercard setup would be hella cool if I could still set up the DDR4 the way I wanted and if the added cost of the PCB for the APU + HBM2 would be low enough that generational upgrades would be possible/plausible. It would be cost-prohibitive if I had to swap to an entirely new board every time (think cost of eternal connectors, SATA/M.2 connectors, stuff like that). But for most PC users, swallowing that would be difficult. As your reaction indicated.

Also, I'm obviously not imagining this as a high-end gaming SKU - hence the term "console killer."

It has to at least be midrange then, which is where PCs have to go to really step out above and beyond consoles in power. PCs, by default, require superior computing performance on all ends to make up for bad ports/bloated software. You need at least a 480 in that APU.

Not sure if anyone has mentioned yet, but isn't this what the whole Steam box is? How has that performed in the marketplace?

No. There isn't a desktop APU that exists today to exceed the graphics power of the PS4 Pro APU, for example.

 

[prtskg]

A lot depends upon zen cpu, very few would be interested in an apu with weak cpu. I like the concept of apu though. It's good for htpc and very energy efficient.

 

[DrMrLordX]

A lot depends upon zen cpu, very few would be interested in an apu with weak cpu. I like the concept of apu though. It's good for htpc and very energy efficient.

The general idea here is something like Raven Ridge, which is already Zen 4c/8t + iGPU. There are rumours that the desktop version will have HBM2, while the mobile version won't.

 

[Valantar]

The general idea here is something like Raven Ridge, which is already Zen 4c/8t + iGPU. There are rumours that the desktop version will have HBM2, while the mobile version won't.

That sounds intriguing, although expensive. Would probably be better with Samsung's 'Low cost HBM' or something similar. Of course, if they could integrate the interposer into the CPU substrate, that would all but eliminate the cost increases. But now we're in "wild speculation" territory. Tread carefully, here be dragons!

 

[sirmo]

Hardware is just a piece of the console experience. Even if such a APU existed, it would still have to live with inefficiencies of PC ports, and poorly optimized games. Those who like consoles like them because they are plug and play.

PC gives you infinite customization which is exactly the opposite approach needed to replicate the console experience.

It really doesn't matter though, I think PCMR is on the rise, there is room for both.

 

[DrMrLordX]

That sounds intriguing, although expensive. Would probably be better with Samsung's 'Low cost HBM' or something similar. Of course, if they could integrate the interposer into the CPU substrate, that would all but eliminate the cost increases. But now we're in "wild speculation" territory. Tread carefully, here be dragons!

Indeed it could get messy. I'm interested in seeing if the HBM2 rumours for AM4 Raven Ridge prove true. Though with the wonky release schedule AMD has managed for Bristol Ridge and Raven Ridge, I'm far more likely to wind up with Summit Ridge. I'll just have to grab a spare 290 and use it instead of an iGPU. Which makes me kinda sad but . . . oh well.

 

[Ancalagon44]

It will happen eventually but probably not this generation. Next generation, when we get to 7nm manufacturing, it will probably start happening.

Integration is the future - it will happen more and more. It used to be that your motherboard needed to have a sound card and network card. You needed a 3D card to play any 3D games. In fact, a lot of the time you needed a 2D card too. The Voodoo 2 was a 3D daughterboard that required a 2D card in order to function. From the Voodoo 3 onwards, everything was integrated into one card.

Eventually, most of that functionality moved from separate cards to the motherboard itself. So, you could get a motherboard with 2D and 3D functionality, as well as some form of sound output and at least one network adapter. Interestingly, motherboards also used to have memory controllers on them - not anymore.

Now, even the 2D and 3D functionality has been moved to the CPU. Heck, AMD has TrueAudio, which will probably be available soon on CPUs if it isn't already.

The point is that integration is moving in one direction only. We do already have APUs that have decent 3D performance. The reason we don't have anything as powerful as the consoles, as others have said, is power. But, as things get more efficient, power will cease to be a constraint. APUs with powerful 3D graphics performance will be much cheaper than a dedicated GPU, and fast enough for most people. Not this year and not next year, but maybe in 5-10 years time.

 

[NTMBK]

Now, even the 2D and 3D functionality has been moved to the CPU. Heck, AMD has TrueAudio, which will probably be available soon on CPUs if it isn't already.

Actually, TrueAudio hardware isn't even on the latest AMD GPUs. "TrueAudio Next" is a purely software OpenCL library.

 

[Ancalagon44]

Actually, TrueAudio hardware isn't even on the latest AMD GPUs. "TrueAudio Next" is a purely software OpenCL library.

Umm no

It is a DSP. Present on every AMD GPU since Bonaire.

 

[NTMBK]

Umm no

It is a DSP. Present on every AMD GPU since Bonaire.

That is the original TrueAudio. It was a DSP. It is no longer present in Polaris. TrueAudio Next is software: http://gpuopen.com/amd-trueaudio-next-and-cu-reservation/

The first thing to make clear is that at its core, TrueAudio Next is an open-source, OpenCL™ audio math library that can run on a GPU or x86 CPU (thus providing a means for possible platform fallback).

 

[superstition]

It will happen eventually but probably not this generation. Next generation, when we get to 7nm manufacturing, it will probably start happening.

Actually, for a part like this AMD would likely go with a larger node, like the 22nm SOI, in order to be able to dedicate less of the die to CPU cores and raise clocks on the CPU cores and graphics without going way outside the efficiency window.

14nm LPP is hardly optimal for this sort of part unless one is going to go with a two chip in one die solution. In that case the low power efficiency window would be made up with an 8 core part, or 6/12. Even so, I don't think AMD would use 14nm LPP for the graphics portion of such a solution.

Integration is the future - it will happen more and more.

Die shrinks exacerbate the issue of concentrated heat, something APUs exacerbate by putting so much GPU power in close proximity to the CPU heat.

There are two other factors people haven't mentioned:

1) How much quality VR catches on in the short term— how high demand there is for a 90 FPS high-quality RIft/Vive console.

2) Whether or not console makers will be willing to return to the sell hardware at a loss model. They've been spoiled by cheap AMD cat cores and such.

As for Intel's use of EDRAM, some think that high-clocked DDR4 (rather than <3000) coupled with good caches makes the EDRAM fairly useless for a CPU. IBM is using it heavily in its Power chips, though.

A 22nm SOI 4/8 + 1070-level GPU HBM2 solution would be pretty nice to see. Basically a 6700K and a 1070 on a chip. I bet the cooling and motherboard power furnishing requirements would be interesting, though. Some sort of vapor chamber design, possibly with out-of-the-case passive heat sinks would likely be mandatory. It would probably have to be a two chips on one die solution or the size of the chip would be enormous.