GDDR5 RAM vs On-Package Cache RAM to improve IGP performance?

[Arachnotronic]

The rumors I heard are slightly different - GT3e is only going to hit some of the lower power mobile parts. Not necessarily limited to those with i7 branding, but that could be the case. I don't know where the line is drawn between mainstream but this definitely won't be paired with the highest performance CPUs.

Intel is pushing a perf/W solution that could be very competitive, with both the on-package memory and a huge number of shaders. This is afforded them by their process technologies. This advantage is indispensable for mobile gaming platforms. It makes it hard for AMD to compete with APUs, nevermind anyone competing with GPUs. If there's limited competition and it's serving that gives Intel more room to raise margins..

Of course we're going to have to see how Kaveri does here but if it takes them many months later to release it it may not matter..



Rumor is that Crystalwell isn't stacked, it's side by side. But AMD may not have the option of making such a large package instead.

Every mobile SoC has been using stacking on top of the package for years, this is surely different from the kind of stacking you refer to but I hope GF at least has this capability or they're in serious trouble.

BTW, everyone's referring to the (supposedly 128MB) of (probably) eDRAM Crystalwell (GT3e) offers as "cache." I doubt it's cache. It's probably just a straight memory buffer that has to be managed manually.

GT3e is for the highest end mobile i7 chips.

 

[Piroko]

I don't think a 128MB buffer needs to be aggressively optimized by the game developer for to make a tangible difference. Just having the driver put what render target objects it can fit on it can make a big difference. Even with a less than sophisticated allocation policy. This isn't really new ground either, there were discrete GPUs in the past with only small amounts of dedicated RAM that were meant to directly address the rest from system memory.

You're either talking about a really long time ago or I missed something. Performance usually tanks so far that a game goes from 60+fps to unplayable as soon as the local GPU memory is exhausted. Also, current game engines at least trade a bit ram usage for more performance with all the buffering going on at the moment.

I don't see how power consumption is going to be comparable. RAM with 3 times the throughput per pin doesn't use the same amount of power, even if you only use it 1/3rd as frequently.
Saving DIMM slots as a cost saving measure doesn't really factor into this if you're comparing against motherboards that could have (and often do) have soldered down DDR3.

http://www.anandtech.com/show/2906/14 There aren't many comparisons between equally equipped DDR3 and GDDR5 cards, but as a tendency the GDDR5 equivalents are not more power hungry while being 30% faster (as a card, memory speed in this case is more than doubled). I'm not concerned with power usage when even 6 Ghz chips get away without a heatsink.

The price of the Dimm slots won't make much of a difference anyways, but if the MCM were a cheap solution I'd think Intel would push it more than a select few bins.
In comparison, Gddr5 was 18-21$ per GB in 2011 (the only reliable source I could find with a quick search) and it's reasonable to expect prices below 10$ per GB for the 5Ghz chips by now compared to about 4.6$ per GB of 1600 Mhz DDR3 chips (source). It is higher cost, but very manageable and probably decreasing further.

 

[jpiniero]

GT3e is for the highest end mobile i7 chips.

That's the current rumor. I think Intel intended it to be on all of the GT3 models, but OEMs balked at the additional cost. Don't be surprised if that fruit company ends up being the only one offering GT3e.

 

[ShintaiDK]

GT3e is HD5200. Its the top line.

 

[Fjodor2001]

Regarding the price and speed of the GDDR5 RAM primarily to be used for AMD's Kaveri, I found this:

http://www.brightsideofnews.com/news/2013/3/5/amd-kaveri-unveiled-pc-architecture-gets-gddr5.aspx

saying:

"While the information we glanced at is clearly preliminary and could be changed till the release of the chip, the focus should be less expensive GDDR5 chips with moderate clock speeds – not the 6GHz hotness you find in high-end graphics cards like GeForce GTX Titan. Specifically the document lists 800 MHz QDR and 850MHz QDR (3200MHz and 3400MHz) clocks which would result with 51.2 GB/s and 54.4 GB/s of system memory bandwidth."

 

[Khato]

1. What could be the reasons Intel vs AMD have selected different solutions?

I still have my doubts that AMD actually will have a DDR3/GDDR5 memory controller (there are pretty big differences in interface protocols between the two) but if they do... Yeah, it'd basically be their only realistic option for increasing available memory bandwidth. It's not really feasible to increase the width of the bus, and they can't wait for DDR4. As for why they wouldn't take Intel's approach, well, recall that AMD took a far different approach to GPU integration than Intel (at least the presented block diagrams look more like an evolution of NB integrated graphics rather than actual integration into the CPU.) Other potential reason could be that such an approach wouldn't play as nice with their discrete GPu architecture/drivers. aka, Intel's GPU is designed exclusively for the bandwidth constrained integrated role whereas I'd be somewhat surprised if AMD isn't designing for discrete and then just stuffing it into the integrated role.

2. What are the pros & cons of each solution? E.g. costs, possibility of the CPU also getting benefits of faster RAM, amount of "fast RAM" available to the IGP, memory bandwidth differences (perhaps On-Package RAM is faster, but smaller amount available?), etc.

Pretty much the only pro for AMD's approach is that it should fix the GPU bandwidth issues in 3D workloads. It's possible that it could also help their GPGPU performance, but that's far more dependent upon the specifics of the workload. Should have a negligible effect upon CPU performance. The cons have already been pretty well covered - far higher price for a given capacity and higher power consumption.

Now as expected, the first pro for Intel's approach is exactly the same as it is for AMD - remove the bandwidth cap from the performance equation. It can do such despite its small size for the simple fact that 3D workload bandwidth requirements don't correspond to the amount of memory space they take up - Intel should be able to satisfy something like half of their bandwidth requirements with their small on-package memory. The other important benefit of this approach is that it actually reduces power consumption - talking to an on-package memory device at lower speeds with a wider bus takes far less energy than getting the data from main memory. The con, of course, is cost. Who knows whether it's more or less expensive than changing main memory to GDDR5... I'm sure Intel likes the fact that the extra 'expense' on the platform is more money in their pockets, whereas AMD is just making more money for the GDDR5 manufacturers.

 

[ShintaiDK]

GDDR5 can also only be soldered and needs to be very close due to limited tracelengths.

 

[Phynaz]

Regarding the price and speed of the GDDR5 RAM primarily to be used for AMD's Kaveri, I found this:

http://www.brightsideofnews.com/news/2013/3/5/amd-kaveri-unveiled-pc-architecture-gets-gddr5.aspx

BSN - enough said. Just because you read it on the internet doesn't make it true.

 

[Idontcare]

I still have my doubts that AMD actually will have a DDR3/GDDR5 memory controller (there are pretty big differences in interface protocols between the two) but if they do...

Maybe I am missing something in the context of the discussion, but AMD already does have that. It just hasn't been put into an APU yet, or rather an APU that is available on the market yet.

But DDR3/GDDR5 integrated memory controller has already been implemented in 28nm CMOS by AMD at TSMC.

 

[SPBHM]

shouldn't be a big problem, their GPUs like the one used for 5570 can work with DDR2, DDR3, GDDR5...

 

[wlee15]

GDDR5 can also only be soldered and needs to be very close due to limited tracelengths.

Not necessarily as there's a JEDEC standard for a GDDR5m SODIMM.

http://www.jedec.org/standards-documents/docs/so-018a

 

[Homeles]

Do you see how fast the GPU is developing towards Gflops rates, compaired to the CPU?

Yes, because we know how GPUs work. GPU FLOP performance scales almost perfectly with transistor count. CPUs do not.

It's very, very safe to count on a doubling of performance every time a GPU is moved to a new process node.

Right now, alot of that "potential" is just wasted, because CPU&GPU dont work together well.
(The "idea" is to fix that, and make use of wasted potential, so you get MUCH more performance)

While true, GPUs aren't a magical panacea for making everything run faster.

 

[Exophase]

GT3e is for the highest end mobile i7 chips.

He did say highest performance parts, without the "mobile" qualifier.

I hadn't heard anything about it only being in a 47W TDP part but if that's the case I really have to question Intel's strategy. If you can burn that much on the CPU you could have probably afforded a decent discrete GPU instead. I was under the impression that the premium was to get Crystalwell in ultrabooks and even tablets where that TDP is far too high.

You're either talking about a really long time ago or I missed something. Performance usually tanks so far that a game goes from 60+fps to unplayable as soon as the local GPU memory is exhausted. Also, current game engines at least trade a bit ram usage for more performance with all the buffering going on at the moment.

I don't know exactly what the performance profile was, but the performance isn't going to from 60+ FPS to unplayable for a current IGP with shared memory, even if it were to go from using the local memory well to not at all. I was more thinking more along the lines that it wasn't a new hurdle software-wise, but if it was so botched as to be useless in the past then it might not make a difference.

http://www.anandtech.com/show/2906/14

There aren't many comparisons between equally equipped DDR3 and GDDR5 cards, but as a tendency the GDDR5 equivalents are not more power hungry while being 30% faster (as a card, memory speed in this case is more than doubled). I'm not concerned with power usage when even 6 Ghz chips get away without a heatsink.

Then why do GDDR5 versions have higher TDP ratings? Just as another blurb from another AT article "Typically we see GDDR5 cards sport a higher TDP thanks to the memory’s higher power consumption, and this would be further driven up by the fact that the GTX 650 is clocked higher than the GT 640."

http://www.anandtech.com/show/6289/nvidia-launches-geforce-gtx-650-gk107-with-gddr5

Isn't it pretty typical for the same HSF to cover the GPU and the RAM? Even if the GDDR5 really is bare that doesn't mean there's no power consumption issue, afterall they tend to distribute that power over a lot of chips...

The price of the Dimm slots won't make much of a difference anyways, but if the MCM were a cheap solution I'd think Intel would push it more than a select few bins.
In comparison, Gddr5 was 18-21$ per GB in 2011 (the only reliable source I could find with a quick search) and it's reasonable to expect prices below 10$ per GB for the 5Ghz chips by now compared to about 4.6$ per GB of 1600 Mhz DDR3 chips (source). It is higher cost, but very manageable and probably decreasing further.

We don't really know what Intel's plans are. We don't know what they'll release later this year. The thing isn't even out.

A > 2x cost for RAM is pretty huge. Maybe if you only care about 4GB. I wouldn't even consider a laptop with only 4GB of RAM. And it's not like there aren't plenty of Trinity laptops with 8GB so that market does exist.

MCM with on-package RAM isn't the only alternative to what AMD is doing. They could have had separate DDR3 and GDDR5 buses. That'd have increased the cost and complexity of the APU but decreased the cost and excess power consumption of the RAM. I don't really know if it would have been worth it or not.

 

[Ajay]

Pretty much the only pro for AMD's approach is that it should fix the GPU bandwidth issues in 3D workloads. It's possible that it could also help their GPGPU performance, but that's far more dependent upon the specifics of the workload. Should have a negligible effect upon CPU performance. The cons have already been pretty well covered - far higher price for a given capacity and higher power consumption.

No, it will help all compute workloads as well as 3D. AMD has been putting a fair amount of work into OpenCL and OpenCL libraries. All sorts of things can be accelerated via OpenCL like H.264 encode/decode, super fast STL libraries like the CUDA based Thrust library and so on. Low speed GDDR5 memory can easily double the bandwidth available to the APU, plus some costs will be avoided such as DIMM slots. For commodity desktops and laptops this scenario could be a Win for AMD. This would certainly boost AMD's advantage in GFX performance. But, I'll require AMD to shift more resources to developer support to get OpenCL libraries in key consumer and business applications.

As a side note, Since the big two console manufacturers will be using AMD APUs, OpenCL use could grow and that would force Nvidia to devote more energy towards OpenCL support to have compatible performance in console ports (currently NVs OpenCL support is lame, especially on Kepler since NV has stated that they will not optimize OpenCL drivers for Kepler and AMD GPUs support the newest version of OpenCL and crush Nvidia in OpenCL performance).

 

[Dark Shroud]

Adobe & ArcSoft have ported part of their code to OpenCL and other smaller companies are taking note at the savings of using OpenCL. Handbrake will also be using OpenCL.

The same OpenCL code runs on AMD & Nvidia GPUs and Intel has it running on their CPU cores.

 

[podspi]

A > 2x cost for RAM is pretty huge. Maybe if you only care about 4GB. I wouldn't even consider a laptop with only 4GB of RAM. And it's not like there aren't plenty of Trinity laptops with 8GB so that market does exist.

MCM with on-package RAM isn't the only alternative to what AMD is doing. They could have had separate DDR3 and GDDR5 buses. That'd have increased the cost and complexity of the APU but decreased the cost and excess power consumption of the RAM. I don't really know if it would have been worth it or not.

Only 4gb of ram is fine as long as it comes with an SSD. Heck, most people I know only have 4gb of RAM with a spinning platter, 5200rpm of course...

 

[Denithor]

This situation unfortunately & ironically reminds me of the P4 days when Intel pushed DDR2 out so their bandwidth-starved chips would stand a chance against AMD with an IMC (huge advantage). In this case however we are going to be facing the reverse, I think - AMD will have to rely on the expensive brute-force solution while Intel will have the more elegant solution.

And I think it comes down to a matter of size. AMD simply doesn't have the luxury of space to integrate that quantity of on-die cache/memory buffer that Intel can thanks to their much smaller process node. Yet another advantage of being ahead in the technology race...

 

[Idontcare]

Is Intel manufacturing the ram chip? I had assumed it would come from Samsung or Micron. Can't say I know much about it though so I wouldn't be surprised if I am wrong.

 

[Roland00Address]

AMD also doesn't have the option of stacking ram on the CPU's PCB as GloFo is not aggressively pursuing that option, unlike Intel and TSMC (Nvidia).

For AMD, they have to make GDDR5 work, they don't have a lot of options because when it comes to their foundry choices they also don't have a lot of options.

Bingo. Intel chose the option they did for they control their fabs and they can make it happen. AMD no longer has its fabs so it must go for the more "commodity" option which is buying gddr from a source such as Samsung.

 

[Ajay]

Only 4gb of ram is fine as long as it comes with an SSD. Heck, most people I know only have 4gb of RAM with a spinning platter, 5200rpm of course...

Oddly, 4GB is still plenty for the vast majority of consumers. It's the smaller number of power users and enthusiasts like us that gag at the thought of < 8GB.

I imagine Best Buy Associates will push people towards 8 GB systems just because 4GB seems so passe.

 

[Ajay]

Adobe & ArcSoft have ported part of their code to OpenCL and other smaller companies are taking note at the savings of using OpenCL. Handbrake will also be using OpenCL.

The same OpenCL code runs on AMD & Nvidia GPUs and Intel has it running on their CPU cores.

That cool. CUDA will likely dominate in high end professional markets for sometime, but I think OpenCL will be the tool of choice for the desktop. If Intel is still on track for introducing a new iGPU architecture for Broadwell, I imagine that they will likely provide support for OpenCL 1.2. Good news all around.

 

[Phynaz]

As a side note, Since the big two console manufacturers will be using AMD APUs, OpenCL use could grow

Are the consoles running OpenCL?

 

[Khato]

Maybe I am missing something in the context of the discussion, but AMD already does have that. It just hasn't been put into an APU yet, or rather an APU that is available on the market yet.

But DDR3/GDDR5 integrated memory controller has already been implemented in 28nm CMOS by AMD at TSMC.

Sorry 'bout that, I didn't really phrase that very well. Technically it's perfectly plausible - it's already been done on multiple discrete cards after all. I was just expressing my uncertainty regarding whether that single 'report' from BSN of Kaveri supporting both DDR3/GDDR5 is accurate in part due to increased complexity of such designs. The differences between DDR3 and GDDR5 are more than enough that they'd have to give their memory controller a pretty hefty overhaul - it's not as simple as just reusing the discrete GPU design. They very well could feel that they need to do so though in order to remain competitive.

Is Intel manufacturing the ram chip? I had assumed it would come from Samsung or Micron. Can't say I know much about it though so I wouldn't be surprised if I am wrong.

Don't recall seeing anything more than speculation on that one. It's been called embedded DRAM frequently, which implies a non-standard device. SemiAccurate seems to think that it's an ultra-wide bus device running at lower speeds which would definitely make sense in terms of power. Really I'd never given it much thought and had always assumed it'd be something Intel makes itself.

 

[Ajay]

Are the consoles running OpenCL?

I've read that it's part of the dev kit for PS4 (it was available on the PS3, IIRC), but support for OpenCL is already there for AMD's APUs, so I expect that some developers will choose to use it. Why have an APU and a GFX card unless you have some plan to use the iGPU on the APU?

 

[Fjodor2001]

BSN - enough said. Just because you read it on the internet doesn't make it true.

Please cool it with the attitude dude. People on this forum are perfectly capable of making their own judgments regarding the validity of content on the Internet. You don't need to post generic remarks like that every time someone links to something. Since we're discussing technology for which the details have not been completely disclosed, we're bound to deal with uncertainties. We have to pick up the information that is available and make our own judgment whether it's likely to be true or not. That's the best we can do.

Regarding this specific article, what is it you think is not true? That slower GDDR5 RAM is cheaper than faster GDDR5 RAM? Or the AMD NDA document they referred to focusing primarily on lower speed GDDR5 RAM for Kaveri?