[SiSoft] AMD monster APU?

[raghu78]

RR is a mainstream APU.
This thing is a niche product squarely aimed at "gaming" laptops and such.
It's KBL-G but better (and it's an MCM too).

We will see in 5-6 months time what this chip is. My bet is a desktop single die APU aimed to take share from Nvidia discrete GPUs especially in OEM systems where they can undercut Intel CPU + Nvidia dGPU in a big way.

 

[Yotsugi]

We will see in 5-6 months time what this chip is. My bet is a desktop single die APU aimed to take share from Nvidia discrete GPUs especially in OEM systems where they can undercut Intel CPU + Nvidia dGPU in a big way.

It won't do anything unless the performance is worth it.
It all depends on AMD's ability to make their GPUs work.

 

[Glo.]

The only way this chip lands in desktop is when it is BGA package for STX market, but also has possibility for installing typical AM4 compatible cooler(IHS on the chip). Then whole Fenghuang Raven platform also would make sense.

If it would be this case 399$ price tag for this market is pretty decent value, and it ticks out every chart that APU has to tick: reducing footprint, impact, power, thermals, while giving decent enough performance. It would also open up new market, for custom built NUC's.

 

[Glo.]

GV107 will probably end up 5-10% faster than GP106 and thats a best case. At the low end Nvidia will have to be less aggressive as they need to hit a balance of perf, power, die size, and most crucially cost. Assuming a GV107 for USD 170 and a 4C/8T CPU at USD 160 you are looking at USD 330 for a cpu + dgpu combo. Power draw of this combination will be well above 125w combined for CPU + GPU. If AMD can come deliver Rx 570 perf and R5 1500X CPU perf at 100w with 2GB HBM2 cache for DIY retail at USD 300 they would be able to win a lot of consumers who want to get excellent gaming performance with very competitive efficiency and form factor advantages in SFF designs. When selling to OEMs AMD can sell this full system against Intel CPU + Nvidia GPU combos and undercut the intel cpu + nvidia dgpu combo heavily.

Core i3 8100T - 35W TDP, 119$, GV107 - 149$, 60W TDP. Whole systep power usage under full load - around 100W of power(+/-10%). Price including MoBo lower than MoBo+ APU.

As I have written higher, for me, the only way this combo lands in desktop market is as BGA package, with motherboards themsleves, already.

Nobody have asked a question. Why haven't we seen to this day X300 MoBo's? remember what was supposed to be X300? Enthusiast Small Form Factor. mSTX is Small form factor.

I would definitely see this as explanation what is happening. BGA mSTX solution would open a lot of doors in desktop space.

 

[Yotsugi]

Why haven't we seen to this day X300 MoBo's?

Because X370 gives more I/O and still fits into mITX form-factor.

 

[Glo.]

Because X370 gives more I/O and still fits into mITX form-factor.

Aaaaand mITX starts to be considered too big for being Small Form Factor .

I do agree that it does not make sense at first glance, however, AMD was working on X300 chipset. What happened to this?

Maybe I am biased, however I would definitely love to see BGA packages coming to desktop, and opening possibilities never before possible. Really small, and unusual form factors would be much more common.

Yes, I am SFF enthusiast...

 

[raghu78]

Core i3 8100T - 35W TDP, 119$, GV107 - 149$, 60W TDP. Whole systep power usage under full load - around 100W of power(+/-10%). Price including MoBo lower than MoBo+ APU.

As I have written higher, for me, the only way this combo lands in desktop market is as BGA package, with motherboards themsleves, already.

Nobody have asked a question. Why haven't we seen to this day X300 MoBo's? remember what was supposed to be X300? Enthusiast Small Form Factor. mSTX is Small form factor.

I would definitely see this as explanation what is happening. BGA mSTX solution would open a lot of doors in desktop space.

GP107 is a 75w chip. GV107 is not going to give GP106 perf at 1/2 the power. TSMC 12FF is not a full node shrink. Its a 16FF+ optimization for area reduction. btw you will never use exactly the same power as CPU+GPU TDP. It takes power to drive signals across a PCI-E bus. most importantly form factor advantages are huge with single die APUs. Cost also is a major advantage and an AMD APU can significantly undercut the CPU + dGPU combo especially in OEM PCs where AMD is not going to sell at DIY retail prices.

 

[Yotsugi]

What happened to this?

No one used this because sticking x370 worked *just* fine.

 

[PeterScott]

We will see in 5-6 months time what this chip is. My bet is a desktop single die APU aimed to take share from Nvidia discrete GPUs especially in OEM systems where they can undercut Intel CPU + Nvidia dGPU in a big way.

My bet is on MCM just like Kaby-G, in fact I bet it is the same GPU die used for Kaby-G.

Reasoning:
1: Everyone was saying why doesn't AMD do it's own version of Kaby-G, this answers that: They are.

2: The "semi-custom" claim from Intel, implied that the GPU die wasn't exclusive to Intel. Re-using the die is a big money saver, and with 28 CU it would match up well with the die size shown by intel.

3: Why integrate a die this big? Everyone is arguing for smaller GPU dies working together in MCM, well MCM makes even more sense here, as CPU, GPU is a perfect dividing line for MCM packaging.

 

[Glo.]

GP107 is a 75w chip. GV107 is not going to give GP106 perf at 1/2 the power. TSMC 12FF is not a full node shrink. Its a 16FF+ optimization for area reduction. btw you will never use exactly the same power as CPU+GPU TDP. It takes power to drive signals across a PCI-E bus. most importantly form factor advantages are huge with single die APUs. Cost also is a major advantage and an AMD APU can significantly undercut the CPU + dGPU combo especially in OEM PCs where AMD is not going to sell at DIY retail prices.

https://www.techpowerup.com/reviews/Palit/GeForce_GTX_1050_Ti_KalmX/28.html
Does GP107 here look like 75W TDP GPU? . OC'ed GTX 1050 Ti's - yes, they are 75W TDP GPUs.

GV107 will bring GP106 performance at the same power consumption. Nvidia was able to increase the core dramatically of Volta GPUs, compared to GP100, at the same power consumption, and with the similar core clocks. If the change of architecture will bring uplift in performance GV107 will easily outperform GP106, while consuming much less power.

Do not get me wrong. I would love to see this on desktop. But I do agree with Stilt's technical point of view, and that this chip may only be custom design, with MCM, to compete with Intel.

IMO best thing AMD could do is to partner with an OEM, like ASRock, and offer those chips only for SFF market as a BGA package, coming with MoBo, already. mITX sized MoBo, with full 4c/8t/28CU/2GB HBM2 chip soldered into the MoBo with place for standard 160 mm cooler on top of the IHS, for 399$? That is actually very good deal and simplifies a lot of hassle with construction. Anything that saves space, and offers some new possibilities, for new markets is great addition.

Actually Im really hoping that the Kaby Lake-G CPUs will come with BGA package boards, and also will be available for consumers.

 

[Glo.]

My bet is on MCM just like Kaby-G, in fact I bet it is the same GPU die used for Kaby-G.

Reasoning:
1: Everyone was saying why doesn't AMD do it's own version of Kaby-G, this answers that: They are.

2: The "semi-custom" claim from Intel, implied that the GPU die wasn't exclusive to Intel. Re-using the die is a big money saver, and with 28 CU it would match up well with the die size shown by intel.

3: Why integrate a die this big? Everyone is arguing for smaller GPU dies working together in MCM, well MCM makes even more sense here, as CPU, GPU is a perfect dividing line for MCM packaging.

If the die is the same - nothing would stop Intel, or Apple from enabling all of 28 Compute Units in the GPU.

 

[Yotsugi]

in fact I bet it is the same GPU die used for Kaby-G.

No, that one is designed with EMIB in mind.
It's semi-custom design as is.

 

[PeterScott]

No, that one is designed with EMIB in mind.
It's semi-custom design as is.

What do you think has to be special for EMIB? Do you think the HBM modules used were new customs designs for EMIB? Do you think Intel designed and new custom CPU for EMIB?

EMIB is just an interface medium, like the interposer that AMD uses for Vega cards. It is more cost effective, but it is functionally similar. What can be integrated via EMIB can be integrated via Interposer.

 

[PeterScott]

If the die is the same - nothing would stop Intel, or Apple from enabling all of 28 Compute Units in the GPU.

You mean the same nothing that doesn't stop board partners from turning all their Vega 56 chips into Vega 64 chips? And just about every other GPU chip sold in multiple configurations since the dawn of time.

 

[raghu78]

My bet is on MCM just like Kaby-G, in fact I bet it is the same GPU die used for Kaby-G.

Reasoning:
1: Everyone was saying why doesn't AMD do it's own version of Kaby-G, this answers that: They are.

2: The "semi-custom" claim from Intel, implied that the GPU die wasn't exclusive to Intel. Re-using the die is a big money saver, and with 28 CU it would match up well with the die size shown by intel.

3: Why integrate a die this big? Everyone is arguing for smaller GPU dies working together in MCM, well MCM makes even more sense here, as CPU, GPU is a perfect dividing line for MCM packaging.

1. This is AMD's first version of their long awaited single die APU with HBM. We will see when it comes if its single die or MCM. Till then we agree to disagree.

2. the semi custom claim from Intel came due to the work done by AMD to get the GPU to work on EMIB which is a completely Intel proprietary technology.

https://www.anandtech.com/show/1200...with-amd-radeon-graphics-with-hbm2-using-emib

The next bit is the interesting one: ‘custom-to-Intel … discrete graphics chip’ from AMD RTG. This means that none of AMD’s current product stack has silicon dedicated to EMIB, but AMD is going to leverage its semi-custom design to provide graphics chiplets for Intel to add to its silicon.

‘In close collaboration, we designed a new semi-custom graphics chip, which means this is also a great example of how we can compete and work together, ultimately delivering innovation that is good for consumers… Similarly, the power sharing framework is a new connection tailor-made by Intel among the processor, discrete graphics chip and dedicated graphics memory. We’ve added unique software drivers and interfaces to this semi-custom discrete GPU that coordinate information among all three elements of the platform.’

3. We will see how big the AMD desktop APU die is. My guess is 12LP 7.5T libraries and a die size around 300 sq mm which is very reasonable.

 

[raghu78]

What do you think has to be special for EMIB? Do you think the HBM modules used were new customs designs for EMIB? Do you think Intel designed and new custom CPU for EMIB?

EMIB is just an interface medium, like the interposer that AMD uses for Vega cards. It is more cost effective, but it is functionally similar. What can be integrated via EMIB can be integrated via Interposer.

dude do some reading of the anandtech article. EMIB is Intel proprietary. The packaging and electrical specifications are given by Intel. AMD does not have a EMIB based GPU die at all. Period. a EMIB die will differ from a die which is built on Si Interposer using TSVs. These are just basic fundamental logic checks you should do before making a statement.

 

[PeterScott]

dude do some reading of the anandtech article. EMIB is Intel proprietary. The packaging and electrical specifications are given by Intel. AMD does not have a EMIB based GPU die at all. Period. a EMIB die will differ from a die which is built on Si Interposer using TSVs. These are just basic fundamental logic checks you should do before making a statement.

Dude, that's like arguing that since we have been using copper wires, we need to design new CPUs to work with silver wires.

EMIB is a dumb interconnect layers. It is not active electronics. It is functionally equivalent to interposer.

You didn't answer the question about the HBM and CPU used in Kaby G. Do you really think those were all new part designed for EMIB.

They weren't. Because EMIB is just an interconnect. It doesn't require all new parts to support it.

 

[raghu78]

Dude, that's like arguing that since we have been using copper wires, we need to design new CPUs to work with silver wires.

EMIB is a dumb interconnect layers. It is not active electronics. It is functionally equivalent to interposer.

You didn't answer the question about the HBM and CPU used in Kaby G. Do you really think those were all new part designed for EMIB.

They weren't. Because EMIB is just an interconnect. It doesn't require all new parts to support it.

Come on man. Do some reading. EMIB is different from Si Interposer as it does not need TSV. Its definitely different at the physical level and at the electrical level to Si interposer. the HBM2 will be integrated into package by Intel at their fabs. You should try and ask Intel about who is their HBM2 provider and did it require custom work ? The CPU and GPU are definitely custom design for EMIB.

https://www.intel.com/content/www/us/en/foundry/emib.html

 

[Glo.]

Dude, that's like arguing that since we have been using copper wires, we need to design new CPUs to work with silver wires.

EMIB is a dumb interconnect layers. It is not active electronics. It is functionally equivalent to interposer.

You didn't answer the question about the HBM and CPU used in Kaby G. Do you really think those were all new part designed for EMIB.

They weren't. Because EMIB is just an interconnect. It doesn't require all new parts to support it.

Im still having a hard time, how the hell you have believe the die between two products is the same, when both of them use DIFFERENT GPU ARCHITECTURES.

AMD chip is using Vega, and Intel got Polaris with HBM2.

 

[turtile]

It could also be Atari's custom chip.

 

[PeterScott]

Come on man. Do some reading. EMIB is different from Si Interposer as it does not need TSV. Its definitely different at the physical level and at the electrical level to Si interposer. the HBM2 will be integrated into package by Intel at their fabs. You should try and ask Intel about who is their HBM2 provider and did it require custom work ? The CPU and GPU are definitely custom design for EMIB.

https://www.intel.com/content/www/us/en/foundry/emib.html

You need to read more, including your own link.

There is nothing there indicating any need to change chip designs for EMIB. The whole point of EMIB is that you mix and match anything in an EMIB package and it uses standard flip chip bumps and microbumps to connect chips, stated in your own link.

Silicon interposer uses the same standard bumps and microbumps to connect chips.
http://electroiq.com/blog/2011/06/silicon-interposers-building-blocks-for-3d-ics/

Same connections for chips, and the interconnects are effectively just dumb wires in either case.

If it works with EMIB, it works with Silicon Interposer, and vice-versa. They are functionally the same.

EMIB is really just a cost reduction, compared to Silicon interposer.

You do not need to redesign chips to work on either, they are both designed to interface common chips.

 

[itsmydamnation]

Im still having a hard time, how the hell you have believe the die between two products is the same, when both of them use DIFFERENT GPU ARCHITECTURES.

AMD chip is using Vega, and Intel got Polaris with HBM2.

Care to show proof that intel product it is polaris based. I haven't seen any.

 

[moinmoin]

Care to show proof that intel product it is polaris based. I haven't seen any.

Leaks so far referred to graphics as gfx804 which is Polaris.

 

[Glo.]

With the theory, that this is MCM package there is slight problem.

There is no 4C/8T CPU design from AMD, alone.

Only Raven Ridge 4C/8T+ Vega 11 CU VGA, and 8C/16T Zeppelin die.

So the only two possibilities is that the GPU is connected via MCM to another GPU, that has connection to HBM2, but that is ridiculous concept. The GPU in Raven Ridge die is disabled, and only CPU is connected to 28CU GPU, which has HBM2 PHY.
Third possibility is that AMD is using Zeppelin die, with disabled cores. But that is another ridiculous idea. All of the engineering samples so far we have seen were full versions of the chips, they were not cut down. Only later we have seen the cut down parts. But it was way down the road.

This is early engineering sample.

 

[Yotsugi]

With the theory, that this is MCM package there is slight problem.

There's no problem.
It's 4/8 RR + Vega32/28 on the same package.
Vega32/28 is connected using PCIe link, just as usual GPU.
It's KBL-G, but with a better GPU and without Intel tax.