dGPU vs APU spin-off

[Gubbi]

My bet is the next leap in APU performance will materialize in the laptop market. You need fast and low power memory to reach the next level in graphics performance for integrated graphics.

People buy desktops/sides because they want the performance and flexibility (or very low cost). In the laptop market, however, it is perfectly fine to solder RAM in place, and glue the whole thing together to never be opened again.

An APU with a closely integrated memory subsystem is the next logical step for Ultrabooks, we already have tablets (iPad Pro) with twice the bandwidth of high end quad core+iGPU CPUs.

This would kill off the lowest performance tier of GPUs. Discrete GPUs would once again lose volume and be confined to higher performance niches (which also commands higher margins); This would make them vulnerable to further erosion from below.

In five years time I could imagine a PC without DIMM sockets and no PCIe x16 slots. You drop an APU with memory on the package into a socket, - done. The socket only need pins for power and PCIe lanes. There'll be a different socket for power users/workstations, which supports multiple sockets; Double the RAM amount ? You drop in another APU package.

Cheers

 

[Deleted member 13524]

And since you mention the power connectors on "motherboards for high-performance APUs" then I assume that the average/majority of AM4 boards won't have all of those power connectors (or phases/VRMs/whatever). So, if you have a PC with one of those cheaper AM4 boards (and consequently a lower-end APU) and decide to play some AAA games, then you have to upgrade your motherboard and APU

With AM3+ boards, you have to check if the board can support the power demands of some Bulldozer CPUs. With LGA1150, a top-of-the-line Z87 motherboard doesn't support Broadwell CPUs even though it's the same interface. In LGA1151 DRAM is an absolute mess: some support DDR4, others DDR3L, others support both, and others even support 1.65V DDR3. You have to be extra careful with compatible memory when buying a LGA1150 motherboard, especially at the lower-end.
And M.2 is another total mess in motherboards right now. Some boards have PCIe M.2, others only SATA M.2, others can have both.

Point is: people either buy a pre-made system from Dell or HP who are going to make sure everything works, or if they want to build their own PC they have to do proper research to make sure everything works. It's not "if it fits it works" anymore, and hasn't been for a while.

So yeah, people buying an expensive 300W APU have to make sure their motherboard and cooling systems have the necessary requirements to get it to work. Just like pretty much any other component.
Regardless, for someone who bought a cheap AM4 board to go with a 95W Raven Ridge they're probably better served with an add-in graphics card as upgrade.
But those interested in building e.g. a compact solution have much better chances here, and probably at a much lower price too.

My bet is the next leap in APU performance will materialize in the laptop market. You need fast and low power memory to reach the next level in graphics performance for integrated graphics.

People buy desktops/sides because they want the performance and flexibility (or very low cost). In the laptop market, however, it is perfectly fine to solder RAM in place, and glue the whole thing together to never be opened again.

An APU with a closely integrated memory subsystem is the next logical step for Ultrabooks, we already have tablets (iPad Pro) with twice the bandwidth of high end quad core+iGPU CPUs.

This would kill off the lowest performance tier of GPUs. Discrete GPUs would once again lose volume and be confined to higher performance niches (which also commands higher margins); This would make them vulnerable to further erosion from below.

We don't even need to look at leaps in APU performance. Little by little, evolutionary steps in iGPUs have been killing lower size/performance segments of dGPUs:
In 2011 we had AMD's Cedar and nvidia's GT218 at <60mm^2. Sandy Bridge killed that dGPU market segment completely.
In 2014 AMD released 90mm^2 Oland and nvidia 80mm^2 GK208. Skylake has most probably killed that segment too. We're probably not seeing sub-90mm^2 dGPUs again.
Come 2017, Raven Ridge (if successful) will start tearing apart at <110mm^2 dGPUs (which may not even exist anymore BTW). That's a 4-35W mobile APU, and pairing it with a dGPU below 50W just wouldn't make sense (except maybe for laptop VR in case it takes off).

Now the actual leap will be HBM and interposers on socketable MCMs. Put a half Zeppelin (4-core Zen + 32 CU GPU + 2-stack HBM2 + dual-channel DDR4) in consumer space and that thing will start eating away all the market of ~200mm^2 GPUs and below with ease (Polaris 11 and GP107) and that's most probably a sub-200W solution. And this is something AMD could launch as early as late 2017 or early 2018.

 

[swaaye]

Frankly I don't think the future will be much different than today. Pretty graphics developments don't seem to be getting less demanding, and manufacturing and 3D architecture improvements just don't help like they used to because of diminishing returns on various fronts.

Most people couldn't care less about games so the typical APU will be the minimum necessary to satisfy the market there. As usual.

I'm sure there will be fancier APU options, as we see developing now. Which are really a terrible value at the moment because Intel likes it that way. But I can't imagine discrete going away. No way. Not when people are endlessly clamoring for and speculating about prettier pixels that mimic reality.

 

[Anarchist4000]

I'm sure there will be fancier APU options, as we see developing now. Which are really a terrible value at the moment because Intel likes it that way. But I can't imagine discrete going away. No way. Not when people are endlessly clamoring for and speculating about prettier pixels that mimic reality.

I doubt it will go away right away, but it seems likely the mid-range segment will start to transition and low continue if not entirely erode. APUs tend to largely replace dGPUs with comparable performance. For a system not requiring high performance graphics, integrated and small form factors make a ton of sense. So it seems likely most, if not all, OEM systems not used for gaming would make the move. It's the availability of small form factor that really seems significant. For an OEM there are a lot of shipping and material costs with a full ATX case as opposed to SFF.

50º using the cheapest solution ($60 I think) while limiting the fan speed.
It doesn't look like such an APU would require a very high-end cooling solution.

Keep in mind those numbers are "delta over ambient" so you'll need to add 20 degrees to each. Regardless, the worst cooler on there is now running 70C which isn't uncommon for a GPU.

 

[DavidGraham]

but the market trend is clear. dGPUs have started to grow in the middle and high end as well, driven by the growth in the gaming sector.

I doubt it will go away right away, but it seems likely the mid-range segment will start to transition and low continue if not entirely erode.

PC gaming hardware market broke $30 billion worldwide for the first time ever in 2016. The Asia-Pacific region saw the most growth overall, while Western European and North American growth focused on high-end hardware. JPR attributes the increased growth in the Asia-Pacific region to a lack of traction by gaming consoles and "an entrenced PC gaming culture."

High-end gaming hardware brought in the bulk of the profits, pulling in over $13 billion in 2016, while mid-range hardware accounted for $10.6 billion. While entry-level hardware is not a minor market at $6.7 billion, it's clear that demand for the best hardware is high, and JPR says that "the western appetite for PC gaming systems costing thousands of dollars is strong."

JPR notes that more and more of our daily computing needs are being taken care of by our phones, and says that as those needs are met, "the PC is ultimately becoming a power user's tool." More and more PC purchases are thus made with gaming in mind.

JPR also offered some thoughts about the draw of PC gaming to explain its faster-than-expected growth. Analyst Ted Pollak offered up reasons like the quality of the gaming experience offered by high-definition and ultra-high-definition monitors, as well as the often-superior controls thanks to mice and keyboards. Additionally, PC gaming's wide variety of hardware options offer more room for expression and customization than ever before.

http://techreport.com/news/31313/report-pc-gaming-hardware-market-expands-to-an-all-time-high

 

[snarfbot]

As long as there exists a benefit for using a discrete GPU they will be made and used. Right now we're not even close to saturating pcie bandwidth. Somewhere along the line when that can't keep up anymore, or external interconnects are consuming more die area than a comparable GPU core + memory would, then I can see it being possible. But I think optical interconnects are going to wind up being integrated first.

 

[Anarchist4000]

High-end gaming hardware brought in the bulk of the profits, pulling in over $13 billion in 2016, while mid-range hardware accounted for $10.6 billion. While entry-level hardware is not a minor market at $6.7 billion, it's clear that demand for the best hardware is high, and JPR says that "the western appetite for PC gaming systems costing thousands of dollars is strong."

That has very little bearing on the original argument. That number includes CPUs, GPUs, integrated/APU, peripherals, etc. No reason a high performance APU for example couldn't drive the market higher. They will continue their trend of supplanting discrete cards with comparable performance.

We are also very excited about the prospects for the AMD Ryzen CPU platform and think it will be adopted at all three hardware tiers. Of course Intel CPUs currently offer superior power and value for gamers of every budget level, and their integrated graphics now rival game consoles.”

This would be an interesting bit of speculation right here. There is very little reason for Ryzen to have a positive impact on revenues as a product already exists on the market that can be substituted for it. Competition doesn't generally drive prices higher. So unless the author is excited about a revenue decrease, he's looking at that TAM AMD keeps mentioning. Shifting that price/performance curve enough to grow the overall market by pulling in new consumers. As the mid to high tier APUs haven't released, they wouldn't have an effect on these numbers.

 

[DavidGraham]

That has very little bearing on the original argument.

It has, people buy high end hardware the same way they buy consoles, the demand for an ever increasing processing power is very much alive and increasing rapidly.

No reason a high performance APU for example couldn't drive the market higher

It would of course, but then we are back to the same argument, high end APUs will NOT compete with high end CPUs + high end GPUs. And since the market is hungry for even more power due to the never ending quest for better graphics and game simulation, high end is here to stay.

 

[Davros]

Heres a point of view I dont think anyone's expressed
Apu's may be increasing in gpu power but its at the expense of cpu power, if that gpu wasnt there they could add a lot of cpu cores
(/me waits for someone to point out somebody's already said that )

 

[Anarchist4000]

It has, people buy high end hardware the same way they buy consoles, the demand for an ever increasing processing power is very much alive and increasing rapidly.

It would of course, but then we are back to the same argument, high end APUs will NOT compete with high end CPUs + high end GPUs. And since the market is hungry for even more power due to the never ending quest for better graphics and game simulation, high end is here to stay.

In which case they'd want APUs, as they would perform better and cost less. Why even bother with a smaller process node if keeping everything spread out is such a good thing? Electronics have moved towards faster, smaller integrated devices ever since the computer was invented. I would expect that trend to continue.

HETEROGENEOUS HPC, ARCHITECTURAL OPTIMIZATION, AND NVLINK : http://on-demand.gputechconf.com/gtc/2015/presentation/S5649-Steve-Oberlin.pdf

Linking this from the Volta thread. Nvidia's own research shows the benefits. Plenty of research and common sense out there to back it up. The ideal computer system will ALWAYS be a SoC as burning power and time through a slow bus is never a good thing. Network communications also never improve as distance increases. Simple physics. That's a very well understood practice and the entire industry bears it out. The real benefit of the APU design is being able to execute scalar graphics work efficiently without wasting power. Efficient use the hardware instead of having discrete parts that are only marginally effective. Not ruling out discrete APUs yet either. That does however challenge the definition of a discrete card. We're still probably a year or so away from seeing those models properly implemented to appreciate the benefits.

Heres a point of view I dont think anyone's expressed
Apu's may be increasing in gpu power but its at the expense of cpu power, if that gpu wasnt there they could add a lot of cpu cores
(/me waits for someone to point out somebody's already said that )

Or use the GPU to accelerate parts of the CPU workload. Many CPU bottlenecks are associated with SSE/AVX functionality. Instructions that tend towards being GPU oriented and highly parallel. Many games end up GPU bound, especially with LLAPIs. Adding more CPU cores isn't necessarily the best solution. With indirect execution that requirement will go down further. Just look at the CPU cores used by consoles.

 

[Davros]

Many games end up GPU bound,

which is an argument for dgpu's

 

[Anarchist4000]

which is an argument for dgpu's

Except the APU design makes the GPU faster and more efficient.

 

[Davros]

Not faster than a discreet gpu

 

[DavidGraham]

smaller integrated devices ever since the computer was invented.

They get smaller because they fulfilled most of their functions, and newly added functions don't require huge processing power, applying your logic to GPUs and Graphics doesn't work, we are nowhere near satisfying our Graphics and visual fidelity needs. Hence bigger GPUs are always needed.

In which case they'd want APUs, as they would perform better and cost less

A high end APU will never perform better than a high end dCPUs +dGPUs, it's a simple logic really. Simple semiconductor physics.

The ideal computer system will ALWAYS be a SoC as

Ideally slow, SoCs can't compete with discrete in performance, the insistence on integration doesn't work in Graphics and new computing workloads, these needs as much powerful chips as they can get, and SoCs can't provide that.

Except the APU design makes the GPU faster and more efficient.

Not faster than a discreet gpu

Exactly, Davros.

 

[Gubbi]

I find the Vega architecture schematics interesting.

APUs and dGPUs have different requirements for their memory system. An APU needs to utilize the cheapest possible ram and support multiple capacities because of the competitive environment of the PC market, a dGPU needs bandwidth, and will sacrifice capacity flexibility and accept higher cost to get there.

Vega bridges this to some extend with its HBM2 cache. You get a fair chunk of very high bandwidth RAM without giving up capacity flexibility. This could be a big deal, not only in the PC market, but also in the server space. I'd certainly love to have a 16+core/32+context CPU with >200GB/s to plough through some of our heavier queries.

Cheers

 

[Blazkowicz]

Not faster than a discreet gpu

Intel GPU are very discreet. They try to make you feel like you don't have a GPU at all


(not really . I'd love to have an i5-5675C, I'm sure that can run Crysis). The major problem with AMD APU is the fear or reality of being CPU starved. When you're CPU starved it's game over, relatively speaking.
It's about possible to stay on 1024x768 CRT forever to do some pretty high end gaming on APU. But people seem to think 1080p is low res now. Perhaps because they can, high res like 1440p and up is an ever cheaper luxury. For some reason nobody makes high quality, fast refresh LCD at low res (e.g. 20" 1366x768 at 144Hz)

I say, bring it on. 768p 144Hz would be useful! But maybe it would fail hard and tank the company that made it.

 

[Alexko]

I don't know why anyone would want a display like that. I mean, it might be good for smooth gaming on a budget, but for anything other than gaming it would be a pain.

 

[Davros]

768p 144Hz

If you had a normal res 144hz monitor and ran your game at 768p It would be just as smooth, turn off scaling if you dont want upscaling
If your going to complain about upscaling wouldnt a big low res screen be just as bad because of the low ppi

 

[Anarchist4000]

A high end APU will never perform better than a high end dCPUs +dGPUs, it's a simple logic really. Simple semiconductor physics.

Why should the APU be any less capable? Because you choose to impose artificial limits to the design? Your entire argument is pure BS. Discrete is faster simply because you choose to think it's faster? Some sort of alternative facts? I've never seen a design where moving components further away from each other improves performance. I've never known reducing latency between components and increasing bandwidth to be detrimental to performance in any of my designs. Tightly integrated designs ALWAYS result in lower power usage and higher performance in electrical systems. There would be no reason at all to create smaller process nodes or integrate components otherwise. I'd challenge you to find a single example where spreading out processors increased performance.

 

[Davros]

Discrete is faster simply because you choose to think it's faster? Some sort of alternative facts?

alternative facts really
name the apu faster than a gtx1080