Speculation: AMD's APU with HBM

[whm1974]

I think by 20/21 we will be on 16 cores mainstream, also with igpu to worry about, I would hope by then Tripple channel is a go.
What is the feasibility of a crystalwell type L4 cache? On 7nm this has to be feasible for 6 core/16CU APU's does it not?
Alongside quad channel lpddr5 option.

What makes you think that 16 cores will be mainstream consumer by then? Hell six cores are only now mainstream consumer for new mid-range systems.

 

[french toast]

What makes you think that 16 cores will be mainstream consumer by then? Hell six cores are only now mainstream consumer for new mid-range systems.

By mainstream I mean non HEDT, ie AM4/5.
put it this way, if we don't get one consumer die with igpu ala intel with zen2...then we most probably are getting two 6 core CCX (12 cores) for desktop... it's not outside the realms of feasibility that AMD continues the core wars and bumps that into either 4x4 core CCX or more likely 2x 8 core CCX for Ryzen 4/5 on 7nm+.

That is going to swallow a tonne more bandwidth, especially since we expect much wider simd units and imo 3/4x SMT alongside wider cores.
Might wait until 5nm..but it is coming.

 

[whm1974]

I wouldn't be surprised at all if 4c/4t and 4/8t CPUs are still mainstream consumer by 20/21.

 

[french toast]

I wouldn't be surprised at all if 4c/4t and 4/8t CPUs are still mainstream consumer by 20/21.

Ok, that's not what I meant...as clarified I meant desktop die.
We are already past 4/8...6 core is getting mainstream.
Ryzen 5 series will be 16 cores is what I'm saying, and it would be preferable imo to have 3 channels by then.

 

[PeterScott]

I wouldn't be surprised at all if 4c/4t and 4/8t CPUs are still mainstream consumer by 20/21.

Exactly. Most PCs sold last year were probably dual cores. Quad cores are probably overkill for most people.

The increases core counts are mainly a marketing benefit for most of the market.

Half the people buying for the extra performance, might run a few benchmarks to see the extra performance, then go back to consuming media and using the internet, using 3% of their new multi-core CPU.

 

[ZGR]

Exactly. Most PCs sold last year were probably dual cores.

Source?

 

[whm1974]

Exactly. Most PCs sold last year were probably dual cores. Quad cores are probably overkill for most people.

The increases core counts are mainly a marketing benefit for most of the market.

Half the people buying for the extra performance, might run a few benchmarks to see the extra performance, then go back to consuming media and using the internet, using 3% of their new multi-core CPU.

I think it more likely that the mainstream lower end consumers with dual core systems will be replacing them in few years with quad core ones, then 16 core processors becoming mainstream.

 

[AtenRa]

1: Great, you have sales figures for how many GTX1050/1060 laptops are sold? I'd like to see them. I am betting it isn't as high as you think. Most people buy laptops with no dGPU.

By the same logic most Dektops sold are with no dGPUs, so there is no point for AMD/NVIDIA to spend huge amount for R&D to create huge 400-500mm dies (VEGA & GP102).

2: No, creating a multipurpose dies that server multiple markets makes more sense. NVidia owns both the laptop dGPU market, and the discrete GPU card market with the same 1030/1050/1060/1080 chips. It's not creating double the mask for one market. It is using one mask to cover two markets, and that makes much more sense.

You missed the point where you also need a CPU die (AMD or Intel). So you have two dies, one for the CPU and another one for the dGPU.

The problem with the "Super" APU attempting to win this market, is that you actually have to win it with one inflexible, very expensive die. So you come out with your ultra expensive APU, and NVida puts GTX 1160 into laptops (Lenovo already let slip these are coming) for lower cost, and higher performance, and you are totally sunk. You just created a large white elephant.

How exactly Lenovo or any other OEM will have lower BOM creating a laptop when they have to buy a CPU from Intel, a dGPU from NVIDIA, 4GB GDDR5 from Samsung, bigger more complex motherboard, bigger and heavier Cooling solution vs a single APU from AMD with 4GB HBM ???

The dGPU option is flexible. It allows you to reuse your desktop dGPU card dies, as mobile dies, and cover all the price/performance levels.

NVidia actually blankets the laptop dGPU market with the dies from GT1030/GTX1050/GTX1060/GTX1080. All 4 have mobile variants, and you can flexibly pair it with your CPU of choice (almost all still choosing Intel CPU for design wins).

A Super APU really can't compete with the flexibility and economy of scale of simply re-using desktop GPU as mobile GPUs to cover the whole range of options.

APUs will stick to low performance(relatively), low cost GPU sections, to server that much bigger TAM that doesn't buy dGPU in their laptops (and desktops).

The higher performance GPU niche will continue to be met by reusing desktop GPUs.

You forget again that you also need a CPU either from AMD or Intel, dGPU alone will not work in a Laptop or Desktop. Also in Desktop, the dGPU is very expensive with way higher BOM to create due to extra PCB, extra cooling etc etc vs a single APU with HBM.

Not to mention that you can always use a single die for multiple SKUs, a 6-core + 20/24 CU APU die could also make a Quad core + 16 CU SKU much like todays Ryzen, you may have 3-5 different SKUs from a single 300mm die easily.

Consider also this, such a high performance APU will command higher ASP. Example for desktop,

6-Core CPU = 200 USD
GTX1050Ti 4GB = 150 USD

Total = 350 USD

APU + HBM with same performance as above hardware could be sold at the same or lower price and create higher profits for the manufacturer than the combination of two different dies for CPU + dGPU + PCB + dual coolers etc.

 

[beginner99]

By the same logic most Dektops sold are with no dGPUs, so there is no point for AMD/NVIDIA to spend huge amount for R&D to create huge 400-500mm dies (VEGA & GP102).

Yeah most sold without dGPU but the dGPU desktop market is not a niche, it's a fairly large market. The market for the super APU is a niche market in comparison. So I question if any profit could be made at all. I doubt it. But let's assume profit can still be made. You still need to assign resoruces to this project. Resources which then are missing say on the Zen3 or Navi project which both are far more important for AMDs bottom line.

Simply said there already would be such a product available if a lot of money could be made and yes that product could as well come from intel.

 

[whm1974]

Yeah considering that the Ryzen APUs and CPUs, and dGPUs products are keeping AMD afloat, I would be putting my resources into instead of something niche with limited market appeal myself.

 

[PeterScott]

By the same logic most Dektops sold are with no dGPUs, so there is no point for AMD/NVIDIA to spend huge amount for R&D to create huge 400-500mm dies (VEGA & GP102).

Except, I can find actual numbers on discrete GPU sales and revenues:
https://www.anandtech.com/show/10864/discrete-desktop-gpu-market-trends-q3-2016/4

Selling about 9 Million Discrete GPUs/quarter on average, and since then mining and gaming hardware sales have been booming, so I expect it is 10 Million+ GPU cards/quarter.

So there is some idea that this is a big enough niche to develop GPU cards for.

You missed the point where you also need a CPU die (AMD or Intel). So you have two dies, one for the CPU and another one for the dGPU.

You missed the point, that you already need to have CPU dies anyway, and that if you create your expensive Super APU, you would still need those normal CPUs anyway, because your Super APU is an expensive niche part, so you still need more normal CPUs for everyone else, and unless you are giving up on the dGPU markets, you still need those as well.

You haven't saved anywhere, you are just creating one hugely expensive extra die for a niche product.

How exactly Lenovo or any other OEM will have lower BOM creating a laptop when they have to buy a CPU from Intel, a dGPU from NVIDIA, 4GB GDDR5 from Samsung, bigger more complex motherboard, bigger and heavier Cooling solution vs a single APU from AMD with 4GB HBM ???

Simple. AMD will be selling it's Super APU for more that the total cost of separate CPU/dGPU/RAM.

Consider also this, such a high performance APU will command higher ASP. Example for desktop,


6-Core CPU = 200 USD
GTX1050Ti 4GB = 150 USD

Total = 350 USD

APU + HBM with same performance as above hardware could be sold at the same or lower price and create higher profits for the manufacturer than the combination of two different dies for CPU + dGPU + PCB + dual coolers etc.

You seem to be forgetting (#1 is key):
1: That it cost more to build one larger die than two smaller size dies. So your silicon is more expensive, not less. Not only that, but your product is lower volume niche, while CPU + dGPU are amortized across a wider range of products, so double whammy on your pricing.
2: That HBM memory costs more than GDDR memory.
3:That Silicon interposer technology is VERY expensive.

You still need a PCB anyway, and few more traces will only add pennies compared to the dollars of silicon interposer.

More expensive silicon, more expensive memory, and more expensive connectivity.

Coolers is nearly a wash. You need one bigger CPU block, vs two smaller ones, but in laptop with that much CPU/GPU power, you likely need dual fans either way. But sure you might save a dollar or two on cooling, but nothing like you cost increase on silicon/HBM/interposer.

 

[Vattila]

An organic substrate can't do HBM.

Yet.

Low-cost HBM on organic substrate may be coming eventually, though.

https://www.extremetech.com/gaming/...tions-to-blow-the-doors-off-the-memory-market

"Right now, Intel has a potent GPU competitor with its Crystal Well, which puts 64-128MB of EDRAM on-package with the CPU. AMD doesn’t really have an answer to Crystal Well at present, and the company’s on-die graphics are already bandwidth-limited. One potential solution is to adopt HBM for APUs and offer a chip with a unified memory pool for CPU and GPU in a single package — but that can only happen if HBM prices drop enough to justify its inclusion."

EDIT: Here is a nice article on upcoming packaging technologies:

"Cheaper packaging options ahead — low cost alternatives to interposer could have big impact on chip design."

https://semiengineering.com/cheaper-packaging-options-ahead/

For further reading, see SemiconductorEngineering's "Packaging, Test & Materials" section:

https://semiengineering.com/category-main-page-packaging-test-electronic-systems/

 

[CatMerc]

Yet.

Low-cost HBM on organic substrate may be coming eventually, though.

https://www.extremetech.com/gaming/...tions-to-blow-the-doors-off-the-memory-market

"Right now, Intel has a potent GPU competitor with its Crystal Well, which puts 64-128MB of EDRAM on-package with the CPU. AMD doesn’t really have an answer to Crystal Well at present, and the company’s on-die graphics are already bandwidth-limited. One potential solution is to adopt HBM for APUs and offer a chip with a unified memory pool for CPU and GPU in a single package — but that can only happen if HBM prices drop enough to justify its inclusion."

EDIT: Here is a nice article on upcoming packaging technologies:

"Cheaper packaging options ahead — low cost alternatives to interposer could have big impact on chip design."

https://semiengineering.com/cheaper-packaging-options-ahead/

Fair enough.

 

[PeterScott]

"Cheaper packaging options ahead — low cost alternatives to interposer could have big impact on chip design."

https://semiengineering.com/cheaper-packaging-options-ahead/

Cheaper doesn't mean cheap. Packaging costs are only one of a multitude of reasons already covered why the Super APU won't be here in the foreseeable future. We can revisit in 5 years and see how the landscape looks then.

Until then I am confident that mainstream APUs won't use HBM, and anything beyond mainstream (for GPU performance) will use a dGPU (the dGPU will often use HBM though).

I think most important thing AMD needs on the APU front is an 8 core version (with same or even lesser GPU). But I wouldn't call that an Super APU. Actually it's more than their 8Core CPU needs at least some basic GPU functionality. This is nice fallback to have in any desktop CPU.

 

[LightningZ71]

I can imagine that an 8 core RavenRidge derived product would have some significant issues with memory starvation. There are times now that you can demonstrate the issues with data starvation on the 2400g when you really push both its CPU cores and it's graphics subsystem at the same time. If you double the cores and ACTUALLY MAKE USE of them, how is dual channel DDR4 supposed to keep up with that demand? Even if they went with a 7nm product that kept the 11CU VEGA solution from RR and combined it with two Pinnacle Ridge CCXs and enabled all 8 cores and all 16MB of L3 cache, it would still have significant memory bus contention issues. Perhaps when they get to DDR5, and Zen2, they might be able to have something that's workable.

It's even worse on the mobile front. SO-DIMMs that clock higher than 2400/2666 (at least on paper) are rare to come by, and are almost never seen in OEM machines on the value side of the spectrum (and, lets face it, an APU only laptop will always be on that end).

This is why many are making the point that AMD should focus more on an MMC/Emib based solution with a higher end CPU chip that has something like a VEGA 3 class tiny iGPU for low power operations, an integrated dGPU chip, and a stack or two of HBM. They can keep the low end APUs for that end of the market, but, they need something for the rest of the laptop market, especially the higher margin mid-range and premium mobile market. In that price range, that sort of product can do well if priced and implemented well.

 

[PeterScott]

I can imagine that an 8 core RavenRidge derived product would have some significant issues with memory starvation. There are times now that you can demonstrate the issues with data starvation on the 2400g when you really push both its CPU cores and it's graphics subsystem at the same time. If you double the cores and ACTUALLY MAKE USE of them, how is dual channel DDR4 supposed to keep up with that demand? Even if they went with a 7nm product that kept the 11CU VEGA solution from RR and combined it with two Pinnacle Ridge CCXs and enabled all 8 cores and all 16MB of L3 cache, it would still have significant memory bus contention issues. Perhaps when they get to DDR5, and Zen2, they might be able to have something that's workable.

Moot point. Since you don't max the CPU and GPU in real world situations.

The only time both are highly active is in gaming, and the 8 CPU cores will be underutilized when gaming, because the iGPU is just not fast enough to get anywhere near maxing out the CPU while gaming. Even if there is some badwidth contention, such is life.

The 8 cores aren't really there for gaming, they are there for Render/Encoding, when you really can make full use of them, while the CPU sits essentially idle.

This is totally workable option and one AMD needs. If Ryzen 2600-2700 had an iGPU, It would be a lot more points convincing me to get it over an Intel solution, because I want to postpone getting a GPU for a while, which I can do with an Intel CPU, but not a Ryzen.

This is why many are making the point that AMD should focus more on an MMC/Emib based solution with a higher end CPU chip that has something like a VEGA 3 class tiny iGPU for low power operations, an integrated dGPU chip, and a stack or two of HBM. They can keep the low end APUs for that end of the market, but, they need something for the rest of the laptop market, especially the higher margin mid-range and premium mobile market. In that price range, that sort of product can do well if priced and implemented well.

You are back to talking about Kaby Lake -G solution. Which isn't really an APU, but just housing a CPU and GPU on a small circuit board together. Sure you can do that, but do not that the advantages of Kaby-G, are negligible compared to the same solution on separate packages. Kaby-G isn't exactly scoring many design wins.

 

[whm1974]

You are back to talking about Kaby Lake -G solution. Which isn't really an APU, but just housing a CPU and GPU on a small circuit board together. Sure you can do that, but do not that the advantages of Kaby-G, are negligible compared to the same solution on separate packages. Kaby-G isn't exactly scoring many design wins.

I agree as the Kaby-G is more of solution looking for a problem more then anything else. Other then in laptop or AIO systems, I really don't see where this CPU/dGPU package fits in.

SFF desktop? Well we have some compact MiniITX cases available that will take a video card that are not that all that big.

 

[PeterScott]

I agree as the Kaby-G is more of solution looking for a problem more then anything else. Other then in laptop or AIO systems, I really don't see where this CPU/dGPU package fits in.

SFF desktop? Well we have some compact MiniITX cases available that will take a video card that are not that all that big.

And even in Laptops and pre-built SFFs, the advantages are negligible compared to just having a separate CPU and GPU on the MB. You need custom MB of about the same size in either case.

For SFF, I would be a lot more excited if something like MXM cards for GPUs took off and really became a standard. Then you could build some really space efficient SFF designs with a choice of GPU options.

I am continually disappointed that we are still stuck with 1980's form factor of giant bulky card cages for expansion. It's too bad we can't get industry players to improve SFF modularity, instead of just putting stupid RGB LEDs on everything.

 

[whm1974]

And even in Laptops and pre-built SFFs, the advantages are negligible compared to just having a separate CPU and GPU on the MB. You need custom MB of about the same size in either case.

For SFF, I would be a lot more excited if something like MXM cards for GPUs took off and really became a standard. Then you could build some really space efficient SFF designs with a choice of GPU options.

I am continually disappointed that we are still stuck with 1980's form factor of giant bulky card cages for expansion. It's too bad we can't get industry players to improve SFF modularity, instead of just putting stupid RGB LEDs on everything.

Well I have seen some really compact MiniITX and MicroATX cases that are not all that big and still allow for some expansion. PITA to build a system with however.

 

[PeterScott]

Yes, but those small cases have to resort to kludges using some kind of slot relocation adapter, and usually end up with the cooling CPU facing out once side of the case, and GPU out the other all uneven and messy. Example is the Silverstone Raven:
https://www.silverstonetek.com/raven/products/index.php?model=RVZ01-E&area=en&top=C

It would be a whole lot cleaner/smaller/easier if something like a small MB with MXM slots (like you used to see in some Laptops) were a standard.

 

[whm1974]

Yes, but those small cases have to resort to kludges using some kind of slot relocation adapter, and usually end up with the cooling CPU facing out once side of the case, and GPU out the other all uneven and messy. Example is the Silverstone Raven:
https://www.silverstonetek.com/raven/products/index.php?model=RVZ01-E&area=en&top=C

It would be a whole lot cleaner/smaller/easier if something like a small MB with MXM slots (like you used to see in some Laptops) were a standard.

Some Steam Machines did something like the last I thought was interesting. It had a MXM graphics card, used SO-DIMM memory, and had M.2 and 2.5" storage. I think Zotac made them but I'm not sure and remember the models.

 

[PeterScott]

I Googled and found the Zotac you mentiond? Nen Steam machine, which reportedly has an MXM card GPU but I can't find anything about MXM listed on the Zotac site. It is very small for something with GTX 1060:
https://www.zotac.com/product/mini_pcs/nen-steam-machine

More recent is the Asrock DeskMini GTX/RX. They are clear that it has MXM and it's a feature they market:
https://www.asrock.com/nettop/intel/DeskMini GTXRX/index.asp

Something like that STX + MXM motherboard becoming a standard would be awesome, along with MXM GPU's taking off.

The problem is these tend to be OEM only products.

 

[whm1974]

I Googled and found the Zotac you mentiond? Nen Steam machine, which reportedly has an MXM card GPU but I can't find anything about MXM listed on the Zotac site. It is very small for something with GTX 1060:
https://www.zotac.com/product/mini_pcs/nen-steam-machine

More recent is the Asrock DeskMini GTX/RX. They are clear that it has MXM and it's a feature they market:
https://www.asrock.com/nettop/intel/DeskMini GTXRX/index.asp

Something like that STX + MXM motherboard becoming a standard would be awesome, along with MXM GPU's taking off.

The problem is these tend to be OEM only products.

Yeah that sounds right. Is Zotac still selling that model? The Asrock DeskMini GTX/RX looks like something I would buy if it supported Coffee Lake or Ryzen CPUs.

 

[GodisanAtheist]

Jumping into the conversation a bit late, but is there anything regarding HBM that prevents it from being used as system RAM as well?

AMD can put together an x86 SOC that has a solid CPU + GPU with 2/4 GB of HBM for the GPU and 6-12GB of HBM as system RAM.

I know AMD has been angling for more of the profit margin pie since they started(and stopped I suppose) selling AMD branded RAM for their systems.

Cutting the RAM vendors out of the loop with some potential performance uplift for the whole system + more profit for AMD sounds like a nice venture all around.

 

[whm1974]

Jumping into the conversation a bit late, but is there anything regarding HBM that prevents it from being used as system RAM as well?

AMD can put together an x86 SOC that has a solid CPU + GPU with 2/4 GB of HBM for the GPU and 6-12GB of HBM as system RAM.

I know AMD has been angling for more of the profit margin pie since they started(and stopped I suppose) selling AMD branded RAM for their systems.

Cutting the RAM vendors out of the loop with some potential performance uplift for the whole system + more profit for AMD sounds like a nice venture all around.

High cost and very high latency. And such an x86 SOC will be priced much higher then what most people would be willing to pay.