[SiSoft] AMD monster APU?

[The Stilt]

Videocardz confirming a desktop APU with 4c/8t and 28CU Vega with 2 GB HBM2 cache.

https://videocardz.com/74464/amd-preparing-mobile-ryzen-5-apu-with-vega-11-graphics

This proves a lot of people wrong who were claiming HBM2 could never make it to mainstream high volume APUs in 2018. Now we can understand the reason behind Samsung's massive USD 26bn capex for 2018 of which USD 7bn is DRAM. With both Intel and AMD producing HBM2 based products in high volume somebody needed to ramp HBM2 capacity in a big way and thats Samsung.

https://www.electronicsweekly.com/b...ungs-2017-capex-may-kill-competition-2017-11/

Sent from my SM-G935V using Tapatalk

If the thing is real it by no means is "mainstream", the same way KBL-G isn't.
Such chips cost more than an entry-level laptop.

EDIT: And if the thing is real, it is not an APU but a MCM part (just like KBL-G).

 

[jpiniero]

If the thing is real it by no means is "mainstream", the same way KBL-G isn't.

I had wondered if as part of the Kaby-G deal, Intel would have to share the EMIB GPU product with AMD.

Otherwise, how hard do you think it'd be to put a CPU on a GPU interposer? IOW Vega 12, except putting an additional CPU die on it.

 

[Yotsugi]

I had wondered if as part of the Kaby-G deal, Intel would have to share the EMIB GPU product with AMD.

EMIB is exclusive to Intel Custom Foundry.

Otherwise, how hard do you think it'd be to put a CPU on a GPU interposer?

Why would you do that?
PCIe x16 link does not need Si interposer.

 

[raghu78]

If the thing is real it by no means is "mainstream", the same way KBL-G isn't.
Such chips cost more than an entry-level laptop.

EDIT: And if the thing is real, it is not an APU but a MCM part (just like KBL-G).

Of course the AMD APU is mainstream. I don't think AMD would have any problem selling a 4C/8T Zen with 28CU Vega APU and 2GB HBM2 on package for around USD 300. If you call core i5 8600k and 8700k as mainstream then definitely this desktop AMD APU is mainstream. AMD will bring USD 200 mid range GPU performance with 1500X performance in a power and space efficient package for a price of around USD 300. btw this is a desktop APU. OEMs can come with some very cool SFF designs using this APU. I think AMD's Fusion vision is finally coming to fruition. For mobile AMD has Raven Ridge with Vega 11 at 15/35w/65w. btw I think the desktop APU chip could be built at GF 12LP and launch in Q2 2018 closer towards Computex.

EMIB is exclusive to Intel Custom Foundry.

Why would you do that?
PCIe x16 link does not need Si interposer.

I doubt its a MCM design. I am thinking its a 300 - 330 sq mm APU. videocardz is confirming a Zen 4C/8T , 28 Vega NCU APU with 2GB HBM2 on package.

https://videocardz.com/74464/amd-preparing-mobile-ryzen-5-apu-with-vega-11-graphics

"Meanwhile desktop series are also making an appearance. Raven Ridge APU featuring 28 CUs (1792 Stream Processors) has been spotted along alongside Fenghuang Raven platform (engineering board).

What Sisoft information is lacking, but we can already share with you, is the core count of this new APU. It’s a quad-core, 8-thread chip with a codename of “2G11SRD1P4MF6_30_N”. "

 

[Yotsugi]

Of course the AMD APU is mainstream.

It's about as mainstream as entry-mid level gaming laptops: not exactly.

btw this is a desktop APU.

I don't think AM4 can handle this.

 

[Glo.]

None of Semi-Custom chips, that were slated for consoles ever appeared on SiSoft database.

If anything. This APU, at the highest end will cost as high as 399$, which means that Stilt is correct - it costs the same as entry level laptop.

If this is SemiCustom design - then for who? Manufacturing costs of this will be around 80-100$, depending on volume, not accounting for design costs.

Is it possible to modify AM4 socket power delivery to be backwards compatible in 400 chipset series Mobos, with previous generation CPUs, and at the same time - new designs?

 

[raghu78]

It's about as mainstream as entry-mid level gaming laptops: not exactly.

I don't think AM4 can handle this.

AM4 is designed for 128w power specification. But thats only theoretical. Motherboard vendors design actual boards to support 150+w power draw with overclocking. In the real world people are overclocking 8C/16T Ryzen to 4+ Ghz and at those speeds it draws 150w easily. I doubt AMD will have power problems given that they will be clocking the CPU and GPU at the sweet spot of efficiency in freq/voltage curve.

None of Semi-Custom chips, that were slated for consoles ever appeared on SiSoft database.

If anything. This APU, at the highest end will cost as high as 399$, which means that Stilt is correct - it costs the same as entry level laptop.

If this is SemiCustom design - then for who? Manufacturing costs of this will be around 80-100$, depending on volume, not accounting for design costs.

Is it possible to modify AM4 socket power delivery to be backwards compatible in 400 chipset series Mobos, with previous generation CPUs, and at the same time - new designs?

I doubt the desktop APU will cost more than USD 350 at max. I think USD 300 is quite realistic. In 2018 14LPP and its optimization 12LP will be mature processes and yields should be very high even for a 300+ sq mm die. The manufacturing cost of the APU even with 2GB HBM2 on package should be < USD 100 at max.

 

[CatMerc]

If the thing is real it by no means is "mainstream", the same way KBL-G isn't.
Such chips cost more than an entry-level laptop.

EDIT: And if the thing is real, it is not an APU but a MCM part (just like KBL-G).

It is an MCM, you're right.

 

[Glo.]

It is an MCM, you're right.

Any clues, or proofs about this?

 

[CatMerc]

Any clues, or proofs about this?

A friend of mine. He wasn't wrong yet, I doubt he is with this.

That said, it's still rumor territory, so I'm not saying 100%. Just that I have high confidence in it being an MCM.

 

[ksec]

Now we can understand the reason behind Samsung's massive USD 26bn capex for 2018 of which USD 7bn is DRAM. With both Intel and AMD producing HBM2 based products in high volume somebody needed to ramp HBM2 capacity in a big way and thats Samsung.

https://www.electronicsweekly.com/b...ungs-2017-capex-may-kill-competition-2017-11/

Sent from my SM-G935V using Tapatalk

HBM volumes are tiny. China has a war chest of 100 billion ready to take the NAND and Memory market. And the Capex increase was simply a message from Samsung that they are not going without a fight.

 

[CluelessOne]

I don't get this. If AMD just release Raven ridge 2700 APU with a TDP 45 W for desktop market on socket AM4 it can be a replacement for ryzen R3. As it is, I think this APU is pointless. Too expensive for lower market, not enough CPU grunt for higher end market. The GPU is too small for higher end, unless it can compete with GTX 1070.

 

[raghu78]

I don't get this. If AMD just release Raven ridge 2700 APU with a TDP 45 W for desktop market on socket AM4 it can be a replacement for ryzen R3. As it is, I think this APU is pointless. Too expensive for lower market, not enough CPU grunt for higher end market. The GPU is too small for higher end, unless it can compete with GTX 1070.

Try getting a RX 570 class GPU (USD 200) and a R5 1500X CPU (USD 160 at microcenter) in a 100-125w TDP using discrete components. Not possible. Thats what AMD is likely aiming for with the desktop APU for USD 300.

 

[Glo.]

Try getting a RX 570 class GPU (USD 200) and a R5 1500X CPU (USD 160 at microcenter) in a 100-125w TDP using discrete components. Not possible. Thats what AMD is likely aiming for with the desktop APU for USD 300.

300$ versions may be cut down ones, with 4C/8T CPU, but for example 1536/1408 GCN cores, and GTX 1050 Ti+10% performance target in 65W TDP(low clocks).

APUs have to be better in some way than CPU+dGPU combo to be viable. If they will cost more while delivering the same performance- they failed.

 

[The Stilt]

There is a possibility of course that this platform is some sort of HW test mule system for console hardware, currently running Windows (current consoles are x86 after all).
4C/8T can easily replace a 8C 16h family Puma CPU due to it's significantly higher IPC and most likely clocks as well.

One thing is certain: if it is a real deal, it ain't a mainstream consumer targeted product (not even remotely).
It would be one hell of a product for ultra high-end gaming laptops and in there it would make sense. Currently AMD has no dGPUs for mobile, which wouldn't absolutely suck compared to nVidia.

 

[raghu78]

300$ versions may be cut down ones, with 4C/8T CPU, but for example 1536/1408 GCN cores, and GTX 1050 Ti+10% performance target.

Why USD 300 for a cut down SKU? I think the fully enabled SKU can sell at USD 300-350. My guess is AMD is building this on 12LP which brings area reductions and performance increase simultaneously at same power and complexity vs 14LPP. This is probably a 320-330 sq mm die. A Q2 launch around Computex is quite possible.

https://www.pcgamesn.com/amd-ryzen-mobile-gaming

"This beefy APU could perform close to AMD’s discrete RX 570 GPU - which sits at 32 CUs, encapsulating 2,048 stream processors. The potential to drop in a processor with rock-solid gaming performance straight into a B350 AM4 board could be a game changer in the budget market, but it all comes down to pricing - especially if HBM2 memory makes the cut."

There is a possibility of course that this platform is some sort of HW test mule system for console hardware, currently running Windows (current consoles are x86 after all).
4C/8T can easily replace a 8C 16h family Puma CPU due to it's significantly higher IPC and most likely clocks as well.

One thing is certain: if it is a real deal, it ain't a mainstream consumer targeted product (not even remotely).
It would be one hell of a product for ultra high-end gaming laptops and in there it would make sense. Currently AMD has no dGPUs for mobile, which wouldn't absolutely suck compared to nVidia.

This is a desktop APU and the Vega GPU would need atleast 50+w on its own even if clocked at 1100 Mhz which is the sweet spot for Vega power efficiency. Add to that the 4C/8T CPU and 100w is the minimum for this APU to provide the best combnation of perf, power, cost and form factor advantages over a CPU + dGPU combo of same performance.

 

[Glo.]

There is a possibility of course that this platform is some sort of HW test mule system for console hardware, currently running Windows (current consoles are x86 after all).
4C/8T can easily replace a 8C 16h family Puma CPU due to it's significantly higher IPC and most likely clocks as well.

One thing is certain: if it is a real deal, it ain't a mainstream consumer targeted product (not even remotely).
It would be one hell of a product for ultra high-end gaming laptops and in there it would make sense. Currently AMD has no dGPUs for mobile, which wouldn't absolutely suck compared to nVidia.

Both consoles: Xbox One X, and PS4 Pro have higher core counts. So this would not have any sense at all, if this is console product. It is PC design.

Why USD 300 for a cut down SKU? I think the fully enabled SKU can sell at USD 300-350. My guess is AMD is building this on 12LP which brings area reductions and performance increase simultaneously at same power and complexity vs 14LPP. This is probably a 320-330 sq mm die. A Q2 launch around Computex is quite possible.

https://www.pcgamesn.com/amd-ryzen-mobile-gaming

"This beefy APU could perform close to AMD’s discrete RX 570 GPU - which sits at 32 CUs, encapsulating 2,048 stream processors. The potential to drop in a processor with rock-solid gaming performance straight into a B350 AM4 board could be a game changer in the budget market, but it all comes down to pricing - especially if HBM2 memory makes the cut."

Because this is premium product that massively simplifies the construction of the computer. 400$ Price tag for something that would cost to manufacture AMD around 80-100$ is not far fetched. AMD has to earn money, not give away hardware.

 

[LTC8K6]

Without it's own HBM2 you will have the shared memory and bandwidth problems, which seems to make it a waste of time? But then if it only has 2gb of ram, it still seems problematic?

Is it going to use the 2gb of HBM2 until it needs more ram and then use system ram?

 

[Glo.]

I don't know how people are calculating this however....


If Nvidia maintains progress on performance of their GPUs, GV107 chip will perform at around +15% of GTX 1060 6GB performance level. Slightly below RX470 performance for this APU is not going to compete with it.

Imagine a scenario, when this APU has 95 or higher wattage TDP. 125W for example. What benefit for it is there to buy this chip, instead of 35W TDP Intel CPU, and GV107 chip which combined will cost less, including low cost MoBo's, will consume less, and perform better than this APU? Remember what I have written, for APUs only way to be viable is if they are in some way better than standard hardware: saving power, cost, are better, or equally performing while maintaining similar amount of thermal output, and power consumption.

In perfect world, perfect APU is: faster, cheaper, more efficient, and producing less heat, and saving space than the CPU and GPU combo, but we haven't seen anything like this, so far, and this APU may be first attempt for this, however - this world is not in a vacuum.

In laptops - yes it has a lot of sense, especially as a competition for Intel+Radeon combo, and here - AMD has clear winner, however, only if it is in comparable TDP ranges: 65 and 100/95W TDP ranges, because those are TDP packages in which Kaby Lake-G will come.

Yes, the GPU can be packed in low TDP, the same way as Intel+AMD are able to pack 1536 GCN core+ 4 Core Kaby Lake in 65W TDP. 12 nm process may be good, if the rumors are correct, and we will see around 4.5 GHz core clocks in 95W Ryzen CPUs.

Without it's own HBM2 you will have the shared memory and bandwidth problems, which seems to make it a waste of time? But then if it only has 2gb of ram, it still seems problematic?

Is it going to use the 2gb of HBM2 until it needs more ram and then use system ram?

High Bandwidth Cache Controller - inherent part of every Vega architecture GPU.

 

[raghu78]

Both consoles: Xbox One X, and PS4 Pro have higher core counts. So this would not have any sense at all, if this is console product. It is PC design.

Because this is premium product that massively simplifies the construction of the computer. 400$ Price tag for something that would cost to manufacture AMD around 80-100$ is not far fetched. AMD has to earn money, not give away hardware.

Firstly AMD is trying to increase its share in discrete GPUs against Nvidia and the APU is a tremendous tool in that fight. AMD will undercut a CPU + dGPU combo handsomely to win over consumers. btw a CPU + dGPU will provide more performance. The key here is AMD will try to hit the best combination of perf, power efficiency, cost and form factor advantages to make it a really compelling product. This chip will be a perfect fit for OEM SFF designs.

I don't know how people are calculating this however....

If Nvidia maintains progress on performance of their GPUs, GV107 chip will perform at around +15% of GTX 1060 6GB performance level. Slightly below RX470 performance for this APU is not going to compete with it.

Imagine a scenario, when this APU has 95 or higher wattage TDP. 125W for example. What benefit for it is there to buy this chip, instead of 35W TDP Intel CPU, and GV107 chip which combined will cost less, including low cost MoBo's, will consume less, and perform better than this APU? Remember what I have written, for APUs only way to be viable is if they are in some way better than standard hardware: saving power, cost, are better, or equally performing while maintaining similar amount of thermal output, and power consumption.

In perfect world, perfect APU is: faster, cheaper, more efficient, and producing less heat, and saving space than the CPU and GPU combo, but we haven't seen anything like this, so far, and this APU may be first attempt for this, however - this world is not in a vacuum.

In laptops - yes it has a lot of sense, especially as a competition for Intel+Radeon combo, and here - AMD has clear winner, however, only if it is in comparable TDP ranges: 65 and 100/95W TDP ranges, because those are TDP packages in which Kaby Lake-G will come.

Yes, the GPU can be packed in low TDP, the same way as Intel+AMD are able to pack 1536 GCN core+ 4 Core Kaby Lake in 65W TDP. 12 nm process may be good, if the rumors are correct, and we will see around 4.5 GHz core clocks in 95W Ryzen CPUs.

High Bandwidth Cache Controller - inherent part of every Vega architecture GPU.

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.

Most importantly this is a single die APU and not a MCM like Kaby G and there are significant perf, power, cost advantages with single die on a mature 14LPP or 12LP.

Without it's own HBM2 you will have the shared memory and bandwidth problems, which seems to make it a waste of time? But then if it only has 2gb of ram, it still seems problematic?

Is it going to use the 2gb of HBM2 until it needs more ram and then use system ram?

2GB of High Bandwidth cache. Thats where the High Bandwidth cache controller of Vega comes into action and allows for smooth gameplay without needing vast amount of VRAM.

https://www.pcworld.com/article/319...or-you-and-gamers-may-be-waiting-a-while.html

Another demo capped Rise of the Tomb Raider at 2GB of memory to show that the card’s radical high-bandwidth cache controller can significantly increase minimum and average frame rates in VRAM-constrained scenarios, as shown in the image above.

https://wccftech.com/amds-vega-doubles-usable-graphics-memory/

 

[Yotsugi]

I don't know how people are calculating this however....


If Nvidia maintains progress on performance of their GPUs, GV107 chip will perform at around +15% of GTX 1060 6GB performance level. Slightly below RX470 performance for this APU is not going to compete with it.

Imagine a scenario, when this APU has 95 or higher wattage TDP. 125W for example. What benefit for it is there to buy this chip, instead of 35W TDP Intel CPU, and GV107 chip which combined will cost less, including low cost MoBo's, will consume less, and perform better than this APU? Remember what I have written, for APUs only way to be viable is if they are in some way better than standard hardware: saving power, cost, are better, or equally performing while maintaining similar amount of thermal output, and power consumption.

In perfect world, perfect APU is: faster, cheaper, more efficient, and producing less heat, and saving space than the CPU and GPU combo, but we haven't seen anything like this, so far, and this APU may be first attempt for this, however - this world is not in a vacuum.

In laptops - yes it has a lot of sense, especially as a competition for Intel+Radeon combo, and here - AMD has clear winner, however, only if it is in comparable TDP ranges: 65 and 100/95W TDP ranges, because those are TDP packages in which Kaby Lake-G will come.

Yes, the GPU can be packed in low TDP, the same way as Intel+AMD are able to pack 1536 GCN core+ 4 Core Kaby Lake in 65W TDP. 12 nm process may be good, if the rumors are correct, and we will see around 4.5 GHz core clocks in 95W Ryzen CPUs.

High Bandwidth Cache Controller - inherent part of every Vega architecture GPU.

Firstly AMD is trying to increase its share in discrete GPUs against Nvidia and the APU is a tremendous tool in that fight. AMD will undercut a CPU + dGPU combo handsomely to win over consumers. btw a CPU + dGPU will provide more performance. The key here is AMD will try to hit the best combination of perf, power efficiency, cost and form factor advantages to make it a really compelling product. This chip will be a perfect fit for OEM SFF designs.



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.

Most importantly this is a single die APU and not a MCM like Kaby G and there are significant perf, power, cost advantages with single die on a mature 14LPP or 12LP.

The thing is not a single die; it also wont are retail channels.
It's a laptop solution.

 

[raghu78]

The thing is not a single die; it also wont are retail channels. It's a laptop solution.

Ryzen Mobile has already launched and we know they are limited to 4C/8T ,11 Vega NCU . Stop the misinformation. Anyway videocardz is saying desktop APU. btw you are forgetting 12LP brings a 15% reduction in area by using 7.5T libraries. 12LPP 7.5T provides > 10% perf than 14LPP 9T at same power and complexity.

https://www.globalfoundries.com/new...-technology-for-high-performance-applications

The new 12LP technology provides as much as a 15 percent improvement in circuit density and more than a 10 percent improvement in performance over 16/14nm FinFET solutions on the market today.

https://www.anandtech.com/show/11854/globalfoundries-adds-12lp-process-tech-amd-first-customer

The new 12LP relies on the groundwork set by the 14LPP, but uses 7.5T libraries, which is one of the ways that enables GlobalFoundries to shrink die sizes by increasing transistor density. Since the library contains different elements, IC developers have to “recompile” their designs to take advantage of the process. Meanwhile, since the 12LP and the 14LPP are very similar, for GlobalFoundries’ existing customers migration path to the 12LP is pretty straightforward. As for others, they are more likely to use the upcoming 7 nm anyway given the fact that it is not far away.

https://videocardz.com/74464/amd-preparing-mobile-ryzen-5-apu-with-vega-11-graphics

 

[Yotsugi]

Anyway videocardz is saying desktop APU.

Oh, WhyCry says a lot of things.

btw you are forgetting 12LP brings a 15% reduction in area by using 7.5T libraries. 12LPP 7.5T provides > 10% perf than 14LPP 9T at same power and complexity.

And?

 

[raghu78]

Oh, WhyCry says a lot of things.

And?

Desktop APU is launching well into Q2 2018 and after Pinnacle Ridge which is confirmed to be built at 12LP. AMD has confirmed 2018 client and graphics products are built at 12LP. Read the anandtech GF 12LP article and GF press release. Connect the dots. 12LP 7.5T is a good choice for a mainstream APU aimed at hitting sweet spot of perf, power and price.

 

[Yotsugi]

12LP 7.5T is a good choice for a mainstream APU aimed at hitting sweet spot of perf, power and price.

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).