AMD CPU speculation... and expert conjecture

[gamerk316]

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

 

[jdwii]

Funny when you look up the definition of "niche" you get this
"a specialized but profitable corner of the market."

 

[jdwii]

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

Gamer do you think we will ever see 8GB of Vram on the APU alone being sold as mainstream gaming parts? Or do you picture only a small buffer for example at the most 512mb of vram?

 

[de5_Roy]

@esrever: interesting - evolved memory hierarchy. is it something like in-memory processing?

 

[Cazalan]

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

There will be a CPU with 16GB of RAM on it in 2016. For many things that will be plenty but it will cost around $4500.

The question is how quickly they can scale down the costs to something mainstream consumers might buy.

 

[juanrga]

oh, you mean juan? he's parroting some s/a articles from before. and i've verified several times in this thread how he lies. his claims(which are actually blanket statements devoid of any specifics) are of no concern to me. this is why the reply was directed towards you.

I find funny that you accuse others of parroting, when your main activity in this thread consists on copy/pasting new links. The only thing you have verified is that you call liars to people whose arguments you are unable to grasp, but others can understand easily.

 

[juanrga]

@8350rocks, the plans of AMD have not changed. The FF interconnect only provides 10Gbps per socket. This is slow even for future APU--APU communication. AMD exascale uses 40--100 Gbps interconnects for APU--APU.

FD-SOI doesn't scale well beyond 10nm, unlike FINFETs. Even IBM had rejected FD-SOI for its future nodes, but this doesn't matter anymore, with IBM abandoning the foundries business by due to the lack competitiveness.

Kaveri reduced clocks by two reasons: (i) lower TDP and (ii) HDL. FD-SOI was not in the menu because cannot provide the needed properties for GCN. AMD is not the only company that has rejected SOI. The foundry enginner correctly notices how all the big guys have rejected SOI due to being pure hype.


@gamerk316, Nvidia has not changed its plans, because the motivation to do CPU cores has not changed. When the project started out in 2006, it was an x86 core with Fermi GPU shaders. Due to x86 license problems, Nvidia changed to ARM cores and the GPU cores have been updated, but the main goal is unchanged. Nvidia has gave a technical talk about its plans for the 2018--2020. Check the slide that I reproduced a pair of pages ago. You can continue to negate evidence but dGPUs are replaced by HCNs (Heterogeneous Compute Nodes). Now check my signature.


@Cazalan, research follows a well-defined series of steps, starting with ideas in a piece of paper and following with computer simulations and finishig with prototypes on silicon. Exascale research doesn't even need drivers. No sure what are you arguing here.


@jdwii, only your comment was "silly", first because the FPU was integrated in the CPU despite being a big chip by the standards of the epoch

and, second, because at 14nm there is enough space for including a R9 290X on a small die APU. At 10nm there is double of that space.

 

[juanrga]

There will be a CPU with 16GB of RAM on it in 2016. For many things that will be plenty but it will cost around $4500.

In 2015 a 'CPU' with up to 16GB of RAM will be released, but it will not use ordinary DDR3/4 RAM, but MCDRAM with a bandwidth much greater than the GDDR5 memory used in today high-end GPUs. That 'CPU' will be up to 4x faster in real-world HPC workloads than the fastest dGPU made by Nvidia up to now: the $5000 K40.

Moreover the 'CPU' will be easier to program than the GPU.

 

[juanrga]

@Yuka, you are right. From an economic side it was an adquisition from AMD, but I was refering to the technical side, which was a merge, a fusion, between the GPU team and the CPU team. The motivations were:

The AMD/ATI acquisition doesn’t make a whole lot of sense on the discrete graphics side if you view the evolution of PC graphics as something that will continue to keep the CPU and the GPU separate. If you look at things from another angle, one that isn’t too far fetched we might add, the acquisition is extremely important.

Some game developers have been predicting for quite some time that CPUs and GPUs were on this crash course and would eventually be merged into a single device. The idea is that GPUs strive, with each generation, to become more general purpose and more programmable; in essence, with each GPU generation ATI and NVIDIA take one more step to being CPU manufacturers. Obviously the GPU is still geared towards running 3D games rather than Microsoft Word, but the idea is that at some point, the GPU will become general purpose enough that it may start encroaching into the territory of the CPU makers or better yet, it may become general purpose enough that AMD and Intel want to make their own.

It’s tough to say if and when this convergence between the CPU and GPU would happen, but if it did and you were in ATI’s position, you’d probably want to be allied with a CPU maker in order to have some hope of staying alive. The 3D revolution killed off basically all giants in the graphics industry and spawned new ones, two of which we’re talking about today. What ATI is hoping to gain from this acquisition is protection from being killed off if the CPU and GPU do go through a merger of sorts.

This is the same motivation why Nvidia began to design its own APUs in the same year AMD brought ATI and rejected Nvidia.

 

[juanrga]

Umm you do realize we still have off die cache's ... we call it system memory. After that is slow long term storage, HDD's. We have since created a memory hierarchy that is as follows.

L1 cache
L2 cache
L3 cache
System Memory
SSD storage
HDD storage

The fastest memory possible is the L1 cache which is only one step away from physical CPU registers, it's also the most expensive in space utilization. After that is L2 and L3 cache. We then move off the CPU die and can utilize different transistor technology for modular memory which is slower but far bigger. Once we're done with memory we move onto the non-volatile storage mechanisms which have the greatest storage density.

Just to illustrate how obscene your comment is I will use a common power setup.

i7-4790K
Total L1 cache: 64KB per core: 256KB total per chip
Total L2 cache: 256KB per core: 1MB total per chip
Total L3 cache: 8MB

System Memory: 16GB DDR3 (8x2)
SSD: 256GB
HHD: 2TB
GTX 780 (3GB GDDR5) GTX 880 (4~8GB GDDR5)

Notice the insane leap of sizes from 8MB of L3 cache to 16GB of system memory and 3~8GB of dGPU memory. So even assuming you could get 256MB of on die memory using today's technology, your still 64 times lower then what the user had in system memory and 12~16 times lower then that same power use's graphics memory.

Congrats you can now load a small linux kernel with nothing else happening. Once you put 1GB of on-die super memory, your now talking a system with 64GB of memory and 12~16GB of fast graphics memory. This would all be with the same technology. You can never catch up.

Using your exaggerated example of DNA memory. 1TB of magical on-die memory would have 64TB of system memory at 12~16TB of fast graphics memory. So while "ONE TERABYTE" sounds big in today's terms, it's awfully small when contrasted to what else will be available.

Since you are a Nvidia owner, let us contrast your ruminations to actual data from the Nvidia project that will kill its own GPUs. Nvidia use the buzzword HCN: Heterogeneous Compute Node.

The design has a L0/L1/L2 hierarchy cache with three topology. The LLC is 256MB. The DRAM is 256GB. The integrated SSD is 1TB.

This is not the more agressive design. Since you mention a Haswell CPU, let me remark that Intel mentions 512GB of DRAM for its own project.

As a final note, since you mention linux, let me inform you that we can run several lighweigth linux distros on 256MB or even less. You can run the kernel and the applications on that small memory footprint.

 

[palladin9479]

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

There will be a CPU with 16GB of RAM on it in 2016. For many things that will be plenty but it will cost around $4500.

The question is how quickly they can scale down the costs to something mainstream consumers might buy.

There most certainly won't be a CPU with 16GB of dram on die, the manufacturing technology isn't there yet. It should be possible to put 16GB of 3D stacked dram on the same socket, would just be a bunch of IC's surrounding the die inside the same package. You can make the package as large as you want it to be so there isn't as much of a space limitation. The real issue is cost, 16GB per chip is extremely small in the server industry where you get 64~128GB per chip is common. It would only be useful for a small set of circumstances where you'd want a multi-socket system that had a small memory footprint but needed to access that memory at ultra fast speeds / incredibly low latency. It wouldn't be useful for commodity computing which is what we're discussing.

It's funny that some people here are basically saying "640KB ought to be enough for anybody" in regards to integration of memory / graphics.

 

[juanrga]

A future dGPU could come self contained in it's own box and sit on the shelf like any eSATA or USB3 drive does today. Giving rise to better board layouts and cooling options.

 

[8350rocks]

Get the point yet? Now, ask yourself, if ~2.2 bil transistors can power 4k in 2020, what can a 6 bil transistor dGPU do with the same tech?

You'll want to rethink the math there. GPUs have only gotten faster by throwing MORE transistors on them. Sure they can do some tweaks with optimizations but largely its by adding transistors.

The problem now is the transistor cost curve is going sideways and slightly up. When GPUs went from 40nm to 28nm the transistors got cheaper and we got faster graphics cards for nearly the same price. Now they're not. With some clever packaging like 2.5D they can continue making larger parts but we'll be paying more for them.

Like I've mentioned before you can buy a 20 Billion transistor part from Xilinx but it will cost you 2 arms and 2 legs, making the GTX Titan Z look like child's play. The GPU makers can learn a lot from Xilinx.

My point was essentially that a bargain bin GPU in 2020 will run 4K, but a top end GPU will be able to accomplish so much more.

For example, for 1080p you can get by with a HD 7770 for 1080p with ~1.5 B transistors...while back in the day you needed the HD 5850 with ~2.15B transistors...

 

[jdwii]

"@jdwii, only your comment was "silly", first because the FPU was integrated in the CPU despite being a big chip by the standards of the epoch"
No juan yet again yours is silly since a FPU is much smaller compared to a GPU such as 295X. So again you are using silly claims.

 

[jdwii]

oh, you mean juan? he's parroting some s/a articles from before. and i've verified several times in this thread how he lies. his claims(which are actually blanket statements devoid of any specifics) are of no concern to me. this is why the reply was directed towards you.

I find funny that you accuse others of parroting, when your main activity in this thread consists on copy/pasting new links. The only thing you have verified is that you call liars to people whose arguments you are unable to grasp, but others can understand easily.

That goes for everyone at tomshardware juan i guess its you then sense you always claim things and find ways out of them when you are proven wrong such as the silly FPU argument or the silly 2500K is even with a A10 7850K

 

[Fidgetmaster]

Yes juan lost all credibility and sanity with the allegation of the A10 7850k being on par with 2500k.....hahahah

 

[juanrga]

My predictions about the A10 7850k were verified up to the single digit percent. And my predictions about future APUs will be verified as well.

 

[Fidgetmaster]

No your predictions are just really off/bad and very narcissistic haha, You believe you have just stellar Narcissistic accuracy down to single percentiles hahaha

 

[gamerk316]

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

Gamer do you think we will ever see 8GB of Vram on the APU alone being sold as mainstream gaming parts? Or do you picture only a small buffer for example at the most 512mb of vram?

"Ever" is such a loaded term, especially since CPUs are going to look very different in 20 or so years. Short term though? Maybe 128MB or so of cache on die just for the GPU portion. Any more makes the part too expensive to mass produce. And I suspect I'm being VERY generous here.

 

[cdrkf]

Yes juan lost all credibility and sanity with the allegation of the A10 7850k being on par with 2500k.....hahahah

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

Gamer do you think we will ever see 8GB of Vram on the APU alone being sold as mainstream gaming parts? Or do you picture only a small buffer for example at the most 512mb of vram?

"Ever" is such a loaded term, especially since CPUs are going to look very different in 20 or so years. Short term though? Maybe 128MB or so of cache on die just for the GPU portion. Any more makes the part too expensive to mass produce. And I suspect I'm being VERY generous here.

I think IRIS PRO already uses 128mb, personally I think *if* stacked memory actually turns into a 'thing' we might see a bit more than that in the not too distant future (maybe)....

 

[de5_Roy]

I find funny that you accuse others of parroting, when your main activity in this thread consists on copy/pasting new links. The only thing you have verified is that you call liars to people whose arguments you are unable to grasp, but others can understand easily.

see, this is where you're lying. again. your whole reply is a strawman argument(edit2: distorting my stance and p.o.v. with intents to troll/bait is lying, btw.) with sprinkled ad hominem here and there.
i am gonna clarify things here: i accused you of parroting in the argument, only after noticing it. not plural. i do post links from other sites. that activity is very much different from what you do. you posted links/spammed promo slides in arguments/discussions without pointing out relevant specifics, most of the time without relevance or specifics at all and without displaying any familiarity or understanding of the content.
most of the time i post links is outside any argument/discussion, mostly for reference. almost everyone does that.

i've verified that you frequently lie in attempts to win arguments instead of engaging contructive discussions and i've pointed out several occasions where you've done so. you did however.....fail to debunk my accusations, so they still stand.
i don't see what attacking my lack of grasp is supposed to accomplish, though. i question the same things everyone else do. LOL.

 

[logainofhades]

A future dGPU could come self contained in it's own box and sit on the shelf like any eSATA or USB3 drive does today. Giving rise to better board layouts and cooling options.

Been talk for years, and a few products, that were external boxes with a dGPU inside. I would actually like to have such a thing, with proper interface connection, for laptops. It is one of those products that makes perfect sense, if the details on getting it to work properly, and laptop makers providing a usable interface to connect to them.

 

[de5_Roy]

i've been ignoring the money aspects in favor of the technological and performance improvement aspects. in reality, chip designers won't make anything if they can't make money off of it. this is why intel doesn't sell hex cores under $250, ddr4 is still unavailable, stacked memory repeatedly gets pushed back, 1366x768 is still norm.
for example: we could have an lga 2011 platform with cheap 128GB (4x 32GB of ddr3 1866) system memory and 1TB ssd and eliminate hdd entirely. we can't, because memory manufacturers decided to prioritize mobile over pc. we had lpddr2 for a long time and suddenly lpddr3 and lpddr4 came out within a rather short time span.
as a result ddr3 kit prices went up and you can't build a small box with kaveri a10 7800 non-K with 64GB ddr3 2133 ram and 512TB SSD (ssd prices stagnated iirc). imagine booting windows/linux, playing games, edit documents, convert videos from and in system ram and then save the system snapshot on the SSD in periodic intervals or at shutdown (provided that the o.s. enables doing this).

 

[cdrkf]

i've been ignoring the money aspects in favor of the technological and performance improvement aspects. in reality, chip designers won't make anything if they can't make money off of it. this is why intel doesn't sell hex cores under $250, ddr4 is still unavailable, stacked memory repeatedly gets pushed back, 1366x768 is still norm.
for example: we could have an lga 2011 platform with cheap 128GB (4x 32GB of ddr3 1866) system memory and 1TB ssd and eliminate hdd entirely. we can't, because memory manufacturers decided to prioritize mobile over pc. we had lpddr2 for a long time and suddenly lpddr3 and lpddr4 came out within a rather short time span.
as a result ddr3 kit prices went up and you can't build a small box with kaveri a10 7800 non-K with 64GB ddr3 2133 ram and 512TB SSD (ssd prices stagnated iirc). imagine booting windows/linux, playing games, edit documents, convert videos from and in system ram and then save the system snapshot on the SSD in periodic intervals or at shutdown (provided that the o.s. enables doing this).

What I find frustrating is how hard it is to get a stock laptop with a reasonable screen and an SSD?!

I mean you can add them with the customisable options available through Dell or Lenovo, but they want something stupid like £300 for a 128gb ssd 'upgrade' when I can buy the same drive for £60 on line. As for the fact that my phone has a higher resolution screen (in total pixels not just density) than my laptop... it's crazy.

 

[gamerk316]

Yes juan lost all credibility and sanity with the allegation of the A10 7850k being on par with 2500k.....hahahah

RAM isn't moving onto the die for reasons already explained. Its costs MORE to RAM on the CPU die. It draws MORE power. You don't have the transistor budget. And apps aren't RAM bandwidth limited, so there's no reason for it.

Gamer do you think we will ever see 8GB of Vram on the APU alone being sold as mainstream gaming parts? Or do you picture only a small buffer for example at the most 512mb of vram?

"Ever" is such a loaded term, especially since CPUs are going to look very different in 20 or so years. Short term though? Maybe 128MB or so of cache on die just for the GPU portion. Any more makes the part too expensive to mass produce. And I suspect I'm being VERY generous here.

I think IRIS PRO already uses 128mb, personally I think *if* stacked memory actually turns into a 'thing' we might see a bit more than that in the not too distant future (maybe)....

Except the eDRAM isn't on the CPU die itself; its external.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/3

Also worth noting:

I asked about the potential to integrate eDRAM on-die, but was told that it’s far too early for that discussion. Given the size of the 128MB eDRAM on 22nm (~84mm^2), I can understand why. Intel did float an interesting idea by me though. In the future it could integrate 16 - 32MB of eDRAM on-die for specific use cases (e.g. storing the frame buffer).

So we're right back to limited die space, hence why its external of the main CPU die. Its essentially acting as an external L4 cache.