AMD CPU speculation... and expert conjecture

[juanrga]

Some time ago I predicted that discrete graphics cards will be killed by 2020 more or less. I shared Nvidia labs quote confirming that they expect the same. AMD's last Papermaster talk was about achieving 25x increase in APUs efficiency by year 2020 (i.e. ~20TFLOPs APUs). Discrete graphics cards were not mentioned during his talk.

AMD is transforming, as everything else, into a SoC company. During last re-organization the CPU division was subsumed inside the APU division.

Now Linus Towalds also agrees on that dGPUs will be killed.

My arguments were centered about the physical inability of an ancient CPU+dGPU to scale up to match a high-end APU due to nonlinear scaling of silicon expected for 2018--2020.

A [ 75W CPU + 225W dGPU ] on 10nm will be about 8--10x slower than a 300W APU on same node, due to the energy wall problem, as every engineer (AMD, Intel, Nvidia...) knows.

I also mentioned economic aspects why discrete GPUs are not viable. Linus arguments are focused in this aspect and are worth mentioning them.

Do you still believe that discrete GPU's have a future?

What do you base that ludicrous belief on? Drugs?

Because everything says that IGP's are getting to be "good enough" for a big enough swath of the market (and that very much includes most gamers - look at the game consoles, for chrissake! You are aware that modern game consoles are IGP's, right?) that the discrete GPU model isn't financially viable in the long run.

So your argument is exactly the wrong way around. It's not that the IGP's can't have a adequate market size, it's the discrete GPU's that have market size problems.

And the IGP's are very much moving in the direction of the GPU being more of an general accelerator (AMD calls the combination "APU"s, obviously). And one of the big advantages of integration (apart from just the traditional advantages of fewer chips etc) is that it makes it much easier to share cache hierarchies and be much more tightly coupled at a software level too. Sharing the virtual address space between GPU and CPU threads means less need for copying, and cache coherency makes a lot of things easier and more likely to work well.

We've seen this before, outside of graphics. Sure, you can use MPI on a cluster, and get great performance for some very specific loads. But ask yourself why everybody ends up wanting SMP in the end anyway. The cluster people were simply wrong when they tried to convince people how hardware cache coherency is too expensive. It's just too complicated to come up with efficient programming in a cluster environment.

The exact same is true in GPU's too. People have spent tons of effort into working around the cluster problems, and lots of the graphical libraries and interfaces (think OpenGL) are basically the equivalent of MPI. But look at the direction the industry is actually going: thanks to integration it actually starts making sense to look at tighter couplings not just on a hardware level, but on a software level. Which is why you see all the vendors starting to bring out their "close to metal" models - when you can do memory allocations that "just work" for both the CPU and the GPU, and can pass pointers around, the whole model changes.

And it changes for the better. It's more efficient.

Discrete GPU's are a historical artifact. They're going away. They are inferior technology, and there isn't a big enough market to support them.

Linus

http://www.realworldtech.com/forum/?threadid=141700&curpostid=141714

I like the last line. Game developers also agree on that discrete GPU's are a historical artifact and they are inferior technology.

 

[gamerk316]

dGPUs aren't going anywhere anytime soon, and I shake my head at anyone who says otherwise. I won't bother to repeat my arguments again though.

 

[colinp]

Juan, you go ahead and game with your APU. I'm going to stick with my CPU and "historical artifact" (I note your use of the present tense) dGPU for now. If you're happy with the experience you get then more power to you. I know I wouldn't be.

 

[-Fran-]

dGPUs aren't going anywhere anytime soon, and I shake my head at anyone who says otherwise. I won't bother to repeat my arguments again though.

Well, 2020 is 5+ years away, so it's not "soon", unless it is to you, gamerk.

Juan, you go ahead and game with your APU. I'm going to stick with my CPU and "historical artifact" (I note your use of the present tense) dGPU for now. If you're happy with the experience you get then more power to you. I know I wouldn't be.

You can't read into the future, so you can't make a claim like that with a straight face. You sound like those gasoline lovers that can't imagine a world without gasoline based vehicles.

--

My point is: wait and see.

The premise has a good background at least, but time will tell if they're actually right about things. Until a product is done, they're just theory and not facts, but it doesn't mean they can't forecast something interesting: in this case, APUs being more than "CPU"+"dGPU".

Cheers!

 

[gamerk316]

Well, 2020 is 5+ years away, so it's not "soon", unless it is to you, gamerk.

I always take the long view on things. And dGPU's are still going to be in most all gaming rigs in 5 years. You simply are not going to be able to make a dGPU class APU, simply due to power/heat constraints. Its really that simple.

 

[etayorius]

Some time ago I predicted that discrete graphics cards will be killed by 2020 more or less. I shared Nvidia labs quote confirming that they expect the same. AMD's last Papermaster talk was about achieving 25x increase in APUs efficiency by year 2020 (i.e. ~20TFLOPs APUs). Discrete graphics cards were not mentioned during his talk.

AMD is transforming, as everything else, into a SoC company. During last re-organization the CPU division was subsumed inside the APU division.

Now Linus Towalds also agrees on that dGPUs will be killed.

My arguments were centered about the physical inability of an ancient CPU+dGPU to scale up to match a high-end APU due to nonlinear scaling of silicon expected for 2018--2020.

A [ 75W CPU + 225W dGPU ] on 10nm will be about 8--10x slower than a 300W APU on same node, due to the energy wall problem, as every engineer (AMD, Intel, Nvidia...) knows.

I also mentioned economic aspects why discrete GPUs are not viable. Linus arguments are focused in this aspect and are worth mentioning them.

Do you still believe that discrete GPU's have a future?

What do you base that ludicrous belief on? Drugs?

Because everything says that IGP's are getting to be "good enough" for a big enough swath of the market (and that very much includes most gamers - look at the game consoles, for chrissake! You are aware that modern game consoles are IGP's, right?) that the discrete GPU model isn't financially viable in the long run.

So your argument is exactly the wrong way around. It's not that the IGP's can't have a adequate market size, it's the discrete GPU's that have market size problems.

And the IGP's are very much moving in the direction of the GPU being more of an general accelerator (AMD calls the combination "APU"s, obviously). And one of the big advantages of integration (apart from just the traditional advantages of fewer chips etc) is that it makes it much easier to share cache hierarchies and be much more tightly coupled at a software level too. Sharing the virtual address space between GPU and CPU threads means less need for copying, and cache coherency makes a lot of things easier and more likely to work well.

We've seen this before, outside of graphics. Sure, you can use MPI on a cluster, and get great performance for some very specific loads. But ask yourself why everybody ends up wanting SMP in the end anyway. The cluster people were simply wrong when they tried to convince people how hardware cache coherency is too expensive. It's just too complicated to come up with efficient programming in a cluster environment.

The exact same is true in GPU's too. People have spent tons of effort into working around the cluster problems, and lots of the graphical libraries and interfaces (think OpenGL) are basically the equivalent of MPI. But look at the direction the industry is actually going: thanks to integration it actually starts making sense to look at tighter couplings not just on a hardware level, but on a software level. Which is why you see all the vendors starting to bring out their "close to metal" models - when you can do memory allocations that "just work" for both the CPU and the GPU, and can pass pointers around, the whole model changes.

And it changes for the better. It's more efficient.

Discrete GPU's are a historical artifact. They're going away. They are inferior technology, and there isn't a big enough market to support them.

Linus

http://www.realworldtech.com/forum/?threadid=141700&curpostid=141714

I like the last line. Game developers also agree on that discrete GPU's are a historical artifact and they are inferior technology.

Going to side on this one with Juanrga, APUs in the long term sound quite awesome and this is in the next 5-6 years so it`s no where near the term "soon".

dGPUs will probably still be around, mostly for very old PCs who have cannot be upgraded to the APUs design.

 

[jdwii]

Isn't getting harder for shrinks as well? Making gamers view-point more valid

With stacked memory i guess it could be possible for a high end APU with GDDR5 memory with high bandwidth and then you can have lower latency since its on die. Just seems like even in 5 years a 295X won't be able to fit into a APU(with 4 CPU cores) even with the assumption that we will shrink are fabrication 2 time(28nm-20nm-14nm). Right now the best APU we have is around a low-mid range GPU that is barely as good as a 80$ GPU. Then we are trying to do Ray Tracing in graphics and 4K will be more mainstream by then.

 

[-Fran-]

Well, 2020 is 5+ years away, so it's not "soon", unless it is to you, gamerk.

I always take the long view on things. And dGPU's are still going to be in most all gaming rigs in 5 years. You simply are not going to be able to make a dGPU class APU, simply due to power/heat constraints. Its really that simple.

Makes sense, and you're right in a broad sense.

To deepen that, do you know if PCIe 3 increased the Watt ceiling for devices? I remember vaguely about it, but I'm not 100% sure they did. 300W is the current ceiling for PCIe 2, AFAIK. What about PCIe 4? Will there be another spec to accommodate higher wattage?

Also, from what I've seen with nVidia, they've been playing the "perf/watt" card since Kepler and it looks like they'll continue that way with Denver (the dinosaur? HAHA; man I'm old!) in the ARM space and Maxwell already has big claims that that department (750ti, for example) for dGPUs. AMD, one way or the other, has been playing that card with their uArchs since the 6000 series.

Since most HW makers are playing the Perf/Watt battle (nVidia, Intel, AMD and VIA, me thinks) the only way to lower your overall consumption is to integrate stuff into tighten designs; meaning APUs and/or SoCs.

I wouldn't be surprised to actually see less performing dGPUs in the upcoming years because of this "trend". The high consuming parts would go into the ultra-expensive high end and/or professional space me thinks. I'd be sad, but not surprised.

Cheers!

 

[juanrga]

dGPUs aren't going anywhere anytime soon, and I shake my head at anyone who says otherwise. I won't bother to repeat my arguments again though.

We could agree/disagree depending on what you mean by "anytime soon".

Juan, you go ahead and game with your APU. I'm going to stick with my CPU and "historical artifact" (I note your use of the present tense) dGPU for now. If you're happy with the experience you get then more power to you. I know I wouldn't be.

"historical artifact" are the words from Linus... but I agree with him: computer history is towards integration.

Why would I want to use a slow CPU+dGPU in the future when APUs will be about 10 times faster?

 

[colinp]

Good question, Juan. Why don't we revisit it when an APU is even 25% as fast as a dCPU+dGPU?

 

[juanrga]

Well, 2020 is 5+ years away, so it's not "soon", unless it is to you, gamerk.

I always take the long view on things. And dGPU's are still going to be in most all gaming rigs in 5 years. You simply are not going to be able to make a dGPU class APU, simply due to power/heat constraints. Its really that simple.

You start with a fallacy because precisely power constraints are the reason why CPU+dGPU cannot scale up and will be replaced by APUs, but ignore the physics... I have a question for you: Who will fabricate those dGPUs for your future gaming rigs?

Your bellowed Nvidia? I guess no:

In this time frame, GPUs will no longer be an external accelerator to a CPU; instead, CPUs and GPUs will be integrated on the same die with a unified memory architecture. Such a system eliminates some of accelerator architectures’ historical challenges, including requiring the programmer to manage multiple memory spaces, suffering from bandwidth limitations from an interface such as PCI Express for transfers between CPUs and GPUs, and the system-level energy

Intel? No. The next year Intel starts replacing its GDDR5-based discrete cards by much faster CPUs with MCDRAM. Larrabe will reappear in a SoC

AMD? Just say me how many times "discrete GPUs" are mentioned in this two page article

http://www.theregister.co.uk/2014/06/20/amd_25x20_power_efficiency_pledge/

The same number, zero, than in this formal announcement about AMD plans for the 2020

http://www.amd.com/en-us/press-releases/Pages/amd-accelerates-energy-2014jun19.aspx

Pay attention to the part where says:

AMD expects to outpace the power efficiency gains expected from process technology transitions through 2020 for typical use based on successfully executing three central pillars of the company’s energy efficient design strategy:

[...]

Through Heterogeneous System Architecture (HSA), AMD combines CPU and GPU compute cores and special purpose accelerators such as digital signal processors and video encoders on the same chip in the form of APUs. This innovation from AMD saves energy by eliminating connections between discrete chips, reduces computing cycles by treating the CPU and GPU as peers, and enables the seamless shift of computing workloads to the optimal processing component.

As I mentioned plenty of times before all this info, there is no way that you can obtain future APU level of performance relying only on silicon improvements. Physics is very clear.

 

[gamerk316]

Plans change. You simply are not going to be able to physically fit a high performing GPU on the same die as a high performing CPU within any power/temp threshold, and no amount of die shrinks is going to change that very simple fact.

AMD can scream HSA as much as they want, but they get most of their revenue out of their GPU division. I really don't see them shutting that down.

Same with NVIDIA, which despite its best attempts, has failed every single time they've attempted to make a SoC (Tegra). They specialize in big, powerful, stand alone GPUs. Their designs simply do not scale down well. Nevermind that, lacking an X86 license, they would be dependent on ARM taking over, which it won't (despite your claims otherwise).

But hey, last year, MSFT was a "devices and services" company. Or how Intel was promoting Netbust 10GHz. Sometimes, plans fail when confronted with reality.

 

[gamerk316]

As I mentioned plenty of times before all this info, there is no way that you can obtain future APU level of performance relying only on silicon improvements. Physics is very clear.

I'm amused how you are projecting APU performance 5 years out to be 10x what dGPUs are today, when the top GPU on an APU is, what, 50% the performance of the best dGPU out there?

I got news: HSA doesn't really bring any performance increases by itself, and brings a LOT of performance headaches. You have to worry about memory bandwidth, which you didn't before. You have to worry about the CPU and GPU operating on the same data at the same time, which you didn't before. And so on. HSA is nice since you avoid having to copy a large amount of data over the PCIe bus, but that itself isn't a significant performance bottleneck since PCIe can handle the bandwidth we need for the foreseeable future.

 

[juanrga]

Going to side on this one with Juanrga, APUs in the long term sound quite awesome and this is in the next 5-6 years so it`s no where near the term "soon".

dGPUs will probably still be around, mostly for very old PCs who have cannot be upgraded to the APUs design.

Excellent point! Precisely Intel will be pursuing a similar strategy for HPC/servers. The next year Intel will start replacing its own discrete cards with a new 'CPU' that will be much faster and will only offer discrete cards for legacy customers that already have a Xeon CPU based system and want to upgrade cheap instead changing the whole platform.

 

[Cazalan]

So in 2020 instead of buying a dGPU to increase gaming performance I'll be buying a second CPU/APU interfaced via an optical interconnect. Wow big difference!

 

[juanrga]

Isn't getting harder for shrinks as well? Making gamers view-point more valid

With stacked memory i guess it could be possible for a high end APU with GDDR5 memory with high bandwidth and then you can have lower latency since its on die. Just seems like even in 5 years a 295X won't be able to fit into a APU(with 4 CPU cores) even with the assumption that we will shrink are fabrication 2 time(28nm-20nm-14nm). Right now the best APU we have is around a low-mid range GPU that is barely as good as a 80$ GPU. Then we are trying to do Ray Tracing in graphics and 4K will be more mainstream by then.

Nobody mentioned GDDR5 memory because it is too slow for the performance discussed here. AMD will be using HBM and Nvidia/Intel/ARM will use HMC.

A 295X2 'only' has 11TFLOPS. AMD is talking about 20TFLOPS APUs for the year 2020. Nvidia is much more aggressive and schedules a 40TFLOP APU/SoC (with 8 CPU cores) for the year 2020.

 

[gamerk316]

Nobody mentioned GDDR5 memory because it is too slow for the performance discussed here. AMD will be using HBM and Nvidia/Intel/ARM will use HMC.

Call me when they hit mass production with decent yields and profit margin, and get integrated into chipsets.

A 295X2 'only' has 11TFLOPS. AMD is talking about 20TFLOPS APUs for the year 2020. Nvidia is much more aggressive and schedules a 40TFLOP APU/SoC (with 8 CPU cores) for the year 2020.

Not happening. Simple math:

A10-7850K: .86 TFLops (CPU + GPU)
R9 290x: 5.6 TFLOPS

So you'd need APU's to catch up nearly an order of magnitude over a 5 year timespan, with dGPUs standing still, to get performance out of an APU that matches TODAYS dGPUs. And the A10 7850k is over-inflated since we're counting CPU performance as part of compute performance, where the 290x is pure GPU performance.

Point being, you'd need over 100% performance increases per year, every year, for the next 5 years, to match todays dGPUs. Its not happening. The End.

 

[cemerian]

The only times it made it difference were in extreme multiple GPU setups where the normal x16 bandwidth would of been divided up amongst several GPU's. This was a test of PCIe 2.0 8x vs PCIe 3 8x. A PCIe 2.0 16x channel has about the same bandwidth as a PCIe 3 8x channel.

Also game development is NOT being nerfed nor hindered due to PCIe, that is horribly incorrect. The very nature of how GPU's process data means that the graphics processor or the geometry source (CPU) will be the limitation. dGPU's tend to have 3~6GB of high speed memory, with PCIe 2.0 the per-lane transfer speed is 500MB's. That would allow you to transfer the entire contents of nearly any mid to upper midrange dGPU in under one second. With a 16x slot your looking at 8GB's which is the entire capacity of everything but the Titan Z in under one second. Games aren't anywhere close to needing more then 8GB's worth of new graphics data per second. They upload everything they need to the dGPU, then just go about their business rasterizing frames while slowly swapping out unneeded data for needed data.

Heck a 32-bit NT executable can't address more then 2GB of application memory, with only about 1.7~1.8GB being usable for the application. So there's no way a 32-bit application can even dream about using more then ~3GB of video memory, hell most don't use more then 512MB ~ 1GB. Which leads me to the real user of graphics memory, the GPU itself as a working area. Many graphic enhancement techniques employed at the driver level are executed inside the dGPU itself, the textures are loaded into graphics memory then rendered to larger levels inside the memory with the shaders and aliasing being done. There is next to no additional burden placed on the interconnect bandwidth, unless your rendering at a much higher internal resolution then your dGPU's memory can support and thus need to constantly swap in and out graphics resources. Of course that creates other problems which are significantly more severe then graphics bandwidth.

Anyway the technical details of how all this works leads to the result that graphics I/O interconnects are usually well ahead of what the current need is in the consumer space, usually by an entire generation.

*gasp* PCIe3 does slightly better in high-bottleneck situations at insane resolutions? I'm SHOCKED I tell you.

Again, you take an edge case and consider it a rule. There is no significant difference between the two at this point, and nowhere near enough to warrant a $200 platform upgrade.

Besides failing to interpret correctly my post, both of you show a lack of familiarity with the edge of game development.

Against the usual myth, I have provided gaming examples where PCIe3 provides a performance advantage over PCIe2. I didn't say that it was the general rule; at contrary, I wrote "here we can see some benchmarks showing that PCIe3 can provide 5%, 10%... gains over PCIe2 in some situations/configurations".

Game developers know that PCIe slots are crippling game development. I wrote before: "In fact, the big advantage of a console such as the PS4 is that avoids the PCIe bottleneck and can use the GPU for novel applications including sound processing or physics". PS4 architect confirms how the outdated PC architecture is crippling development of next gen games:

If you look at a PC, said Cerny, "if it had 8 gigabytes of memory on it, the CPU or GPU could only share about 1 percent of that memory on any given frame. That's simply a limit imposed by the speed of the PCIe. So, yes, there is substantial benefit to having a unified architecture on PS4, and it’s a very straightforward benefit that you get even on your first day of coding with the system. The growth in the system in later years will come more from having the enhanced PC GPU. And I guess that conversation gets into everything we did to enhance it."

Cerny envisions "a dozen programs running simultaneously on that GPU" -- using it to "perform physics computations, to perform collision calculations, to do ray tracing for audio."

But that vision created a major challenge: "Once we have this vision of asynchronous compute in the middle of the console lifecycle, the question then becomes, 'How do we create hardware to support it?'"

One barrier to this in a traditional PC hardware environment, he said, is communication between the CPU, GPU, and RAM. The PS4 architecture is designed to address that problem.

"A typical PC GPU has two buses," said Cerny. "There’s a bus the GPU uses to access VRAM, and there is a second bus that goes over the PCI Express that the GPU uses to access system memory. But whichever bus is used, the internal caches of the GPU become a significant barrier to CPU/GPU communication -- any time the GPU wants to read information the CPU wrote, or the GPU wants to write information so that the CPU can see it, time-consuming flushes of the GPU internal caches are required."

The three "major modifications" Sony did to the architecture to support this vision are as follows, in Cerny's words:

* "First, we added another bus to the GPU that allows it to read directly from system memory or write directly to system memory, bypassing its own L1 and L2 caches. As a result, if the data that's being passed back and forth between CPU and GPU is small, you don't have issues with synchronization between them anymore. And by small, I just mean small in next-gen terms. We can pass almost 20 gigabytes a second down that bus. That's not very small in today’s terms -- it’s larger than the PCIe on most PCs!

We already saw some early PS4 games filling 3GB for video memory[#] and using the GPU in ways that are not possible to use in a PC with a discrete card attached to a PCIe slot.


[#] Partitioned as follows,
* Non-Streaming Textures: 1321MB
* Render Targets: 800MB
* Streaming Pool (1.6GB of streaming data): 572MB
* Meshes: 315MB
* CUE Heap (49x): 32MB
* ES-GS Buffer: 16MB
* GS-VS Buffer: 16MB

to the above add a CPU/GPU shared Memory pool of 128MB

ROFL

 

[juanrga]

Good question, Juan. Why don't we revisit it when an APU is even 25% as fast as a dCPU+dGPU?

The laws of physics will not change. There is little to revisit. If plans are aggressively accelerated the replacement could happen before, if delays happen the plans could be postponed for 2025, but that is all. It will happen yes or yes the question is when.

Plans change. You simply are not going to be able to physically fit a high performing GPU on the same die as a high performing CPU within any power/temp threshold, and no amount of die shrinks is going to change that very simple fact.

AMD can scream HSA as much as they want, but they get most of their revenue out of their GPU division. I really don't see them shutting that down.

Same with NVIDIA, which despite its best attempts, has failed every single time they've attempted to make a SoC (Tegra). They specialize in big, powerful, stand alone GPUs. Their designs simply do not scale down well. Nevermind that, lacking an X86 license, they would be dependent on ARM taking over, which it won't (despite your claims otherwise).

But hey, last year, MSFT was a "devices and services" company. Or how Intel was promoting Netbust 10GHz. Sometimes, plans fail when confronted with reality.

You star with the same fallacy once again. As mentioned before, there is no way you can fit the same high performance in the same power/temp threshold using CPU+GPU, and no amount of die shrinks is going to change that very simple fact.

Nvidia most revenue comes from selling discrete cards, but the laws of physics announce their death. I have given you a quote where they admit that discrete cards will disappear by the year 2020. Nvidia is not designing any discrete card for the year 2020, they are only designing APU/SoC.

I asked you to answer who will do the discrete cards of your imagined 2020 gaming rigs. Please answer the question, don't evade it.

Nvidia has 'failed' to make a SoC because (i) they used standard ARM cores whereas competition used custom ARM cores, (ii) they used an ancient graphics architecture (only the new Tegra K1 will bring by first time a decent GPU), (iii) they lacked needed integration such as modems, and (iv) they are competing against giants such as Apple, Qualcomm,...

Why do you believe Intel is losing billions per month trying to enter that market whereas AMD avoids it as the plague?

Everyone knows that ARM will take on Intel. Why do you believe that Intel is accelerating its Server plans, increasing cores, and adding FPGAs?

As I mentioned plenty of times before all this info, there is no way that you can obtain future APU level of performance relying only on silicon improvements. Physics is very clear.

I'm amused how you are projecting APU performance 5 years out to be 10x what dGPUs are today, when the top GPU on an APU is, what, 50% the performance of the best dGPU out there?

I got news: HSA doesn't really bring any performance increases by itself, and brings a LOT of performance headaches. You have to worry about memory bandwidth, which you didn't before. You have to worry about the CPU and GPU operating on the same data at the same time, which you didn't before. And so on. HSA is nice since you avoid having to copy a large amount of data over the PCIe bus, but that itself isn't a significant performance bottleneck since PCIe can handle the bandwidth we need for the foreseeable future.

You start with a fallacy. It is not me alone who is doing the prediction, it is the entire industry, including your bellowed Nvidia.

I also recommend you to read the link to the 25x20, because it is evident that you didn't. There you can see why APU will outperform GPUs.

The whole of HSA pivots around two points: performance and simplification of programing.

The superior performance of HSA has been shown, everyone knows that. And Nvidia and Intel are pursuing similar approaches: Nvidia with its unified-memory ARM+CUDA and Intel with its neo-heterogeneity on MCDRAM.

 

[gamerk316]

I asked you to answer who will do the discrete cards of your imagined 2020 gaming rigs. Please answer the question, don't evade it.

Both AMD and NVIDIA. NVIDIA can say whatever they want, but until they can design a competitive SoC, and as long as it makes money, they will produce dGPUs. As for AMD, its their highest grossing part of the company.

Nvidia has 'failed' to make a SoC because (i) they used standard ARM cores whereas competition used custom ARM cores, (ii) they used an ancient graphics architecture (only the new Tegra K1 will bring by first time a decent GPU), (iii) they lacked needed integration such as modems, and (iv) they are competing against giants such as Apple, Qualcomm,...

All of which add on costs, which decreases profit, making the part less attractive. It's easy to say "add this missing stuff", but its much harder to do so while also making a profit.

Why do you believe Intel is losing billions per month trying to enter that market whereas AMD avoids it as the plague?

Because its VERY high profit margin. If you get in, there's a lot of money to be made. I advocated for AMD focusing on servers YEARS ago, and argued BD would be a good platform to do so (while also arguing about its flaws as a HEDT chip).

Everyone knows that ARM will take on Intel. Why do you believe that Intel is accelerating its Server plans, increasing cores, and adding FPGAs?

See above.

 

[gamerk316]

The superior performance of HSA has been shown, everyone knows that.

No, it hasn't. HSA, by itself, has no performance benefit.

The fallacy you are making is that APUs require an HSA design, where they don't.

 

[juanrga]

So in 2020 instead of buying a dGPU to increase gaming performance I'll be buying a second CPU/APU interfaced via an optical interconnect. Wow big difference!

The difference will be "big". In the first place, instead developing two processors, the company resources will be spent on only one; thus, that second CPU/APU/SoC will be cheaper. In the second place, that second CPU/APU/SoC will not work as accelerator/coprocessor but as peer to the first CPU/APU/SoC which will simplify programming. In the third place, hat second CPU/APU/SoC will have access to main memory.

Nobody mentioned GDDR5 memory because it is too slow for the performance discussed here. AMD will be using HBM and Nvidia/Intel/ARM will use HMC.

Call me when they hit mass production with decent yields and profit margin, and get integrated into chipsets.

A 295X2 'only' has 11TFLOPS. AMD is talking about 20TFLOPS APUs for the year 2020. Nvidia is much more aggressive and schedules a 40TFLOP APU/SoC (with 8 CPU cores) for the year 2020.

Not happening. Simple math:

A10-7850K: .86 TFLops (CPU + GPU)
R9 290x: 5.6 TFLOPS

So you'd need APU's to catch up nearly an order of magnitude over a 5 year timespan, with dGPUs standing still, to get performance out of an APU that matches TODAYS dGPUs. And the A10 7850k is over-inflated since we're counting CPU performance as part of compute performance, where the 290x is pure GPU performance.

Point being, you'd need over 100% performance increases per year, every year, for the next 5 years, to match todays dGPUs. Its not happening. The End.

Not sure which is your first point about yields. HBM/HMC are already in the roadmaps. AMD has already showed prototypes (behind closed doors) of HBM, and Intel will start shipping a processor with stacked DRAM the next year. This processor will kill many Nvidia GPUs.

The problem with your math is that you start with a fallacy. You are comparing a GDDR5 290W GPU to a DDR3 95W APU.

 

[blackkstar]

Oh, the assumptions. AMD says that the traditional GPGPU configuration requires copying memory from system RAM to VRAM. Then AMD releases slides showing GPU can access system memory via pointers. And then everyone somehow manages to take that as it only being a thing APUs can do.

Here is a slide discussing hUMA, a necessary part of HSA:

http://i.imgur.com/QHP1KQf.jpg

Can you spot the APU?

Does someone have actual proof that APU is the only thing HSA and HUMA can work on or is it just baseless conjecture because a dGPU and dCPU have too much latency over PCIe 3.0, which can (you're all assuming) never be solved and that latency is completely relevant in GPGPU.

You know, latency. There is a huge concern for doing tasks the AMD HSA guide says is not for latency intensive workloads being overly affected by latency and I don't really understand it.

If the added latency plus the decreased time to perform a task is faster than one with lower latency and longer time to perform the task, it's a win.

Would you rather have 100ms of latency and a second to finish a task or 1ms of latency and 4 seconds to finish a task?

 

[gamerk316]

The difference will be "big". In the first place, instead developing two processors, the company resources will be spent on only one; thus, that second CPU/APU/SoC will be cheaper. In the second place, that second CPU/APU/SoC will not work as accelerator/coprocessor but as peer to the first CPU/APU/SoC which will simplify programming. In the third place, hat second CPU/APU/SoC will have access to main memory.

First, Development Cost != Profit. If APUs make $x, and dGPUs make $5x, guess what? I'm not cutting my dGPU development, regardless of the cost savings. These are the type of idiotic management decisions that bankrupt companies.

Second, the code doesn't care about where and how the actual GPU is located and communicated with. To tell the dGPU/iGPU/APU to do some DX/OGL function, I still call the same bloody function. The only people who care are the people who write OS Kernels. To 99.9999999% of the planet, there is no difference between a dGPU and HSA APU as far as coding is concerned.

Third, everything having access to the same memory space is nice, but has its own downsides.

Example: Under HSA, you can NEVER have the CPU and GPU manipulating the same data at the same time. One of the two will have to stop processing until the other finishes in order to maintain the memory state. It's the age old thread problem: Whenever you have multiple threads working on the same data, you lose performance whenever you have to stop one to allow the other to finish. That's why a "proper" threading design only threads totally independent sub-tasks, that can run independently of the main program (and why gains due to threading are always limited for complicated programs).

Point being, to avoid that performance loss, you have to design your program in a way so that the CPU and GPU are always working on different datasets, and overlap as little as possible. Wait, thats already what we're doing. Hence: No benefit.

The problem with your math is that you start with a fallacy. You are comparing a GDDR5 290W GPU to a DDR3 95W APU.

The funny thing about "maximum theoretical performance" is you eliminate any delays due to data not being available for processing. Its based sorely on the computation resources, not memory bandwidth. So your argument here is a fallacy.

 

[juanrga]

I asked you to answer who will do the discrete cards of your imagined 2020 gaming rigs. Please answer the question, don't evade it.

Both AMD and NVIDIA. NVIDIA can say whatever they want, but until they can design a competitive SoC, and as long as it makes money, they will produce dGPUs. As for AMD, its their highest grossing part of the company.

Nvidia has 'failed' to make a SoC because (i) they used standard ARM cores whereas competition used custom ARM cores, (ii) they used an ancient graphics architecture (only the new Tegra K1 will bring by first time a decent GPU), (iii) they lacked needed integration such as modems, and (iv) they are competing against giants such as Apple, Qualcomm,...

All of which add on costs, which decreases profit, making the part less attractive. It's easy to say "add this missing stuff", but its much harder to do so while also making a profit.

Why do you believe Intel is losing billions per month trying to enter that market whereas AMD avoids it as the plague?

Because its VERY high profit margin. If you get in, there's a lot of money to be made. I advocated for AMD focusing on servers YEARS ago, and argued BD would be a good platform to do so (while also arguing about its flaws as a HEDT chip).

Everyone knows that ARM will take on Intel. Why do you believe that Intel is accelerating its Server plans, increasing cores, and adding FPGAs?

See above.

NVIDIA says that by 2020 dGPUs will be replaced. Nvidia is not making any dGPU design for that timeframe. You cannot ignore what the company says and pretend that they will do a GPU in your imagination.

I think you didn't understand my question about why Intel is loosing bilions and AMD avoiding that market.