Question [AT] AMD extends GloFo WSA to 2025

maddie

Diamond Member
Jul 18, 2010
3,861
2,943
136

jpiniero

Lifer
Oct 1, 2010
11,510
3,153
136
What else would they be making?
In 2025 other than Monet? Possibly the IO die for Milan and Milan-X and also maybe some new low end chipsets. You would think everything else available today that uses 12 nm would be EOL by then.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,193
3,584
136
I believe Monet is for Automotive as well just like V1000 (which is the one used by Tesla Model Y). V2000/3000 cannot be certified for Auto because they are using TSMC N7/6.
I guess they'd rather use GF 12LP+ instead of annoying TSMC by using Samsung 8A (which is a better process and also Automotive certified) besides the fact they already produce IOD2 there.
Looks like they are serious about Auto if they tape out something and also agreed on the wafer contract.
I guess there will be a Monet successor too then.
But Monet has way too less CUs (1TF+ @2GHz) for a modern Automotive chip (these days 2-3 TF is possible to get)

One thing about such products is that you will never hear a word but they exist and they do not show up on any benchmarking website.

Or dirt cheap chrome books...



On second thought this deal is to fab chrome book processors like Monet and beyond. Ryzen C series APUs
 
Last edited:
  • Like
Reactions: Tlh97 and Tarkin77

jpiniero

Lifer
Oct 1, 2010
11,510
3,153
136
Or dirt cheap chrome books...
I am expecting Monet to be more of a Dali/Picasso replacement. But I like the idea of it being used in Automotive too. I'm not sure how big they want to make it so while I'm pretty sure it's 4 Zen 3 cores the CU count is going to be a mystery.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,193
3,584
136
I am expecting Monet to be more of a Dali/Picasso replacement. But I like the idea of it being used in Automotive too. I'm not sure how big they want to make it so while I'm pretty sure it's 4 Zen 3 cores the CU count is going to be a mystery.
Yeah I think it is Dali/Picasso replacement. Way too less CU for Automotive.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,484
981
136
GlobalFoundries is in the process of down ramping 14nm/12nm FinFETs. Fitting room for 45nm FDSOI nodes at Fab8: 45RFE, 45SPCLO, 45SG01, 45SP3 and their 45nm PDSOI counterparts: 45RFSOI, 45CLO. There is also SkyWater's FDSOI RH90 at Malta via MOU. As well as the 22FDX/12FDX integration at Fab8 going on as well. All of which are critical to turning Malta around leading to all fabs being profitable. Rather than just Singapore and Dresden carrying Malta.

FinFETs at GlobalFoundries unless there is a new fab spawning there is no room for them going forward. Thus, it is a matter of when GlobalFoundries will pull the cord for them. Which is probably when 12FDX starts its dev-to-prod ramp at Malta next year.
 
Last edited:

LightningZ71

Golden Member
Mar 10, 2017
1,266
1,284
136
What ever happened to the 12LP+ node that supposedly absorbed their 7LP node improvement efforts?!?!

I can see AMD continuing to use trailing wafer capacity, and maybe producing IO Dies on 22FDX at some point.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,484
981
136
What ever happened to the 12LP+ node that supposedly absorbed their 7LP node improvement efforts?!?!
Why stop at 12LP+ when there is suppose to be a 12LP++ node.

14LPP = 2015
12LP = 2017
12LP+ = 2019
12LP++ = 2021?

The only reason to stop refinement is if the node is being axed. For example, heavily performance orientated refinement;
22FDX = No strain on NFET // 2018
22FDX+ = SiGe below BOX of NFET // 2020
22FDX++ = SiGe below BOX plus sSOI NFET // 2022+
Monet might be using it.
"Worked on delivering CReST patterns for Renoir, Mero, Arden, Oberon, Navi10, Raphael, Mendocino, Pheonix, Navi31 projects."
So far Monet never existed to begin with... since there is only one FT6 product available across AMD Employees.
 
Last edited:

Joe NYC

Senior member
Jun 26, 2021
687
618
96
I see a few uses:
- continuing ramp of Milan, into 2023 - huge IOD
- Zen3, Zen 3D, into 2023 - sizeable IOD
- Moving all of 6xx generation chipsets to GF 12/14nm
- Integration of formerly external components (networking, Wifi, Bluetooth) into the chipset
- Opportunistic use of capacity for Monet or embedded.
 
  • Like
Reactions: Vattila

Joe NYC

Senior member
Jun 26, 2021
687
618
96

There are so many possibilities... could very well be IO-dies for Zen 4 ... or maybe even an Interposer for Zen5 based Designs?
All of the rumors point to AMD moving Zen 4 IOD to TSMC N6. A lot of good reasons for that, including power, die size, ability add advanced features such as RDNA 2 iGPU.

Then there is ability to use advanced TSMC packaging technologies from stacking potentially bridges. And then, the full packaging itself could be done at TSMC facilities.

As far as interposers - I think they are dead from AMD POV. EFB bridges can achieve most of the benefit at fraction of cost, and there will be other bridging technologies in the future.

But as far as current IODs in Zen 3, for comparison, it will take probably at least a year before Zen 4 overtakes Milan / Milan X. So, if you start from mid 2022 launch of Zen 4, at mid 2023, Milan will still be the majority of server chips. By the time Zen 4 and Sapphire Rapids launch, Milan could be > 20% of server chips, as high as 25% market share.

And that is going to be a lot of gigantic Zen 3 IODs that AMD will be buying, their production will not end before 2024-2025/
 
  • Like
Reactions: Vattila

moinmoin

Diamond Member
Jun 1, 2017
3,442
4,820
136
Interesting that AMD is voluntarily extending the WSA. 14nm/12nm being on the way out is a good point there since AMD promised 10 years of availability for its embedded products, so would need to ensure that until at least early 2028 (for V1000) if not longer (the latest R1000 launched early 2020, so until early 2030). Keeping the WSA alive longer may be cheaper than stock-keeping for that long.
 
  • Like
Reactions: Tlh97 and Joe NYC

Joe NYC

Senior member
Jun 26, 2021
687
618
96
Interesting that AMD is voluntarily extending the WSA. 14nm/12nm being on the way out is a good point there since AMD promised 10 years of availability for its embedded products, so would need to ensure that until at least early 2028 (for V1000) if not longer (the latest R1000 launched early 2020, so until early 2030). Keeping the WSA alive longer may be cheaper than stock-keeping for that long.
Since AMD is upping the amount by $500 million and additional year, AMD is looking at strong demand and possible new uses of GF.
 

moinmoin

Diamond Member
Jun 1, 2017
3,442
4,820
136
Since AMD is upping the amount by $500 million and additional year, AMD is looking at strong demand and possible new uses of GF.
The article itself already stated that the WSA amounts to $500 million a year, so adding a year and $500 million changes nothing per year. Overall constant demand is what I'd call that. It will be interesting how AMD will keep it constant, to me it doesn't sound like a major new product would come, but something clearly needs to replace the old outgoing products to keep the output constant overall. $500 million per year may also well be just the minimum investment needed to keep 14nm/12nm running, and AMD will see later how to best make use of the spare capacity.
 
  • Like
Reactions: Tlh97 and Joe NYC

Joe NYC

Senior member
Jun 26, 2021
687
618
96
The article itself already stated that the WSA amounts to $500 million a year, so adding a year and $500 million changes nothing per year. Overall constant demand is what I'd call that. It will be interesting how AMD will keep it constant. $500 million may also well be just the minimum investment needed to keep 14nm/12nm running.
My expectation of the original agreement was that it would tail off in the last year to next to nothing as opposed to it being a cliff from 500m to zero. So, it is possible that now the tail is still going towards zero, now in 2025, but first years will get additional $100 - $150 million per year of wafer purchases.
 
  • Like
Reactions: Tlh97 and moinmoin

NostaSeronx

Diamond Member
Sep 18, 2011
3,484
981
136
Interesting that AMD is voluntarily extending the WSA. 14nm/12nm being on the way out is a good point there since AMD promised 10 years of availability for its embedded products, so would need to ensure that until at least early 2028 (for V1000) if not longer (the latest R1000 launched early 2020, so until early 2030). Keeping the WSA alive longer may be cheaper than stock-keeping for that long.
The WSA might extend all the way to 2035+ for AMD.

This WSA reduced the price AMD had to pay for 14nm/12nm; V1000 = 2028 EOL, R1000 = 2029 EOL. There is also EPYC Embedded as well that is covered. The reduced price of 14nm/12nm in turn increased the minimum annual wafer purchase limit.

As far as I know there is a replacement clause for dropped parts:
28nm Puma = 2024-2025 EOL, but was EOL'd in 2021
28nm Excavator = 2025-2026 EOL, but was EOL'd in 2021

Following this:
{28/22} Internally "N-2 AMD/GF", {14/12} Internally "N-1 AMD/GF"
AMD has two years from last shipment to replace "N-2" products: Steppe/Crowned/LX, Merlin/Brown/Prairie, at GlobalFoundries.
Prior proof of two-year replacement: Krishna/Wichita (2012-GF28) -> Bhavani/Beema/Mullins (2014-GF28)

----
14/12 is being axed at Malta:
2017-2018 = 45,000 wspm
2019-2020 = 35,000 wspm <-- 14LPP new design halt == 45PDSOI introduction
2021-2022 = 15,000 wspm <-- 12LP new design halt == 45FDSOI introduction
2023-2024 = 10,000 wspm <-- 12LP+ new design halt == 22FDX/12FDX introduction

12LP is actually two nodes: 12HPC and 12HD. 12LP+ covers both thus the new nodes are 12HPC+(below Fin booster) and 12HD+(0.5V SRAM)
HPC is the 10.5T/9T library with >=13 metal layers
HD is the 7.5T library with <=11 metal layers

The current axeing of 14nm/12nm and the prior May 2021 WSA actually removed the exclusivity agreement for 14nm/12nm products. Leading to those products being shifted to TSMC to cover for same EOL.

GF 14nm/12nm (N-1) -> TSMC 16nm/12nm (N-1) for;
Ryzen Embedded, EPYC Embedded, V-series, R-series, cIOD/sIOD, etc.
 
Last edited:

gdansk

Senior member
Feb 8, 2011
974
678
136
AMD said it isn't related to Xilinx acquisition. Presumably after Ian's tweet.
But Monet seems too weak to be worth $500 million plus of wafers alone. Possibly some embedded products getting more (long term) orders.
 

Mopetar

Diamond Member
Jan 31, 2011
6,742
3,809
136
Another rumor that I heard some time ago was that they were going to use GF for HBM manufacturing, but I have no idea how legitimate that was. I'm rather skeptical that all $2.1 billion of that is for Monet since that's basically the Athlon replacement.

If we assume $3,000 per wafer, that amounts to 700,000 wafers in total. Picasso had a die area of about 210 mm^2 and was build on GF's 12nm node so we can assume a similar size. The process should be quite mature by now and they can probably bin anything that isn't completely disabled for some reason. Using node enhancements to reduce the area should give them ~300 dies per wafer that they can use.

So the question is if you think they're going to push 210 million Athlon APUs into the market over the next four to five years. That seems like much too large a number to be reasonable. I know that there are a lot of people who think AMD is going to start making IO dies on TSMC, but I think the TSMC wafers are in too short supply and gain too little benefit over using a low power 12 nm node since IO (especially the physical interconnects to busses) doesn't shrink well.
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,484
981
136
But Monet seems too weak to be worth $500 million plus of wafers alone. Possibly some embedded products getting more (long term) orders.
New Product w/ 2 year HVM does not make sense for 14LPP/12LP/12LP+ since 45SOI at GF is already swamping Fab8 w/ higher profits.

6FF Mendocino ~Q2-Q3 2022 or later => 12FF Monet ~Q2-Q3 2023 doesn't make sense.

AMD given the WSA is receiving in-depth information from GlobalFoundries. Pushing a HVM product on a node with a known supply shrink doesn't make sense.

Given Chengdu Fab11.1(180nm/130nm) and 11.2(22FDX) is being replaced by Singapore Fab7.H. Plus, low $Billions for Dresden Fab1.3 and Malta Fab8.2 would make the more viable option to be a 22FDX-HVM product line.

There is a higher demand for low-cost/low-power/low CPU core count:
Banana Pi R1/R2/R2 Pro versus PC Engines APU2/APU3/APU4

AMD since 2018 is no longer rushing nodes at GlobalFoundries. The reason for the removal of low-cost investment from 2015 is primarily based on unfeasible options of low-cost being on 14LP/7LP/5LP/3LP, etc. However, this is no longer really the case since GlobalFoundries pivoted to Pervasive/Low-cost differentiated markets instead of bleeding edge/scaling leadership.

AMD's strategy at GlobalFoundries must reflect GlobalFoundries' strategy as a whole. It in turn also must reflect the dual roadmap introduction in 2016.
Performance at Any Cost = FinFET track
Cost-effective Performance = FDSOI track
With the subsequent pivot in 2018 favoring FDSOI nodes going forward.
N-2 = 28BLK -> 22FDX
N-1 = 14/12 -> 12FDX
As shown below:
fdsoi.png

"The pervasive semiconductor market is driving breakthrough innovation across broad applications such as longer battery life for mobile devices, always-on access to connected devices, high data throughput for work from home, streaming, gaming and AR/VR, powerful sensing for safe and comfortable autonomous driving and embedded memory for secure cryptographic credentials. Unlike processor-centric compute devices, pervasive semiconductor performance is driven more by circuit design, specialty materials and specialized manufacturing processes.

Innovation in pervasive ICs is measured in terms of precision, accuracy, bandwidth, efficiency and sensitivity. When combined with greater breadth and diversity of customers and end markets, these factors tend to result in more stable demand and pricing for pervasive semiconductors than processor-centric compute semiconductors. Additionally, capital expenditure requirements are generally significantly lower for manufacturers of pervasive semiconductors."
FDX.png

New products by AMD at GlobalFoundries obviously will reflect the change.

12nm APU = N-1 Cost/Perf, N+1 Power competes with R1000(14nm-150mm2) and R2000(14nm-210mm2) and R3000(6nm-162 mm2)
22nm APU = N-2 Cost, N-1/N+1 Perf/Power doesn't compete with anything, there is no G1000 or G2000 series.

There is also the case 22nm doesn't compete with prior G-series since those are already end of production at GloFo.
 
Last edited:

NTMBK

Diamond Member
Nov 14, 2011
9,767
3,763
136
New Product w/ 2 year HVM does not make sense for 14LPP/12LP/12LP+ since 45SOI at GF is already swamping Fab8 w/ higher profits.

6FF Mendocino ~Q2-Q3 2022 or later => 12FF Monet ~Q2-Q3 2023 doesn't make sense.

AMD given the WSA is receiving in-depth information from GlobalFoundries. Pushing a HVM product on a node with a known supply shrink doesn't make sense.

Given Chengdu Fab11.1(180nm/130nm) and 11.2(22FDX) is being replaced by Singapore Fab7.H. Plus, low $Billions for Dresden Fab1.3 and Malta Fab8.2 would make the more viable option to be a 22FDX-HVM product line.

There is a higher demand for low-cost/low-power/low CPU core count:
Banana Pi R1/R2/R2 Pro versus PC Engines APU2/APU3/APU4

AMD since 2018 is no longer rushing nodes at GlobalFoundries. The reason for the removal of low-cost investment from 2015 is primarily based on unfeasible options of low-cost being on 14LP/7LP/5LP/3LP, etc. However, this is no longer really the case since GlobalFoundries pivoted to Pervasive/Low-cost differentiated markets instead of bleeding edge/scaling leadership.

AMD's strategy at GlobalFoundries must reflect GlobalFoundries' strategy as a whole. It in turn also must reflect the dual roadmap introduction in 2016.
Performance at Any Cost = FinFET track
Cost-effective Performance = FDSOI track
With the subsequent pivot in 2018 favoring FDSOI nodes going forward.
N-2 = 28BLK -> 22FDX
N-1 = 14/12 -> 12FDX
As shown below:
View attachment 54882

"The pervasive semiconductor market is driving breakthrough innovation across broad applications such as longer battery life for mobile devices, always-on access to connected devices, high data throughput for work from home, streaming, gaming and AR/VR, powerful sensing for safe and comfortable autonomous driving and embedded memory for secure cryptographic credentials. Unlike processor-centric compute devices, pervasive semiconductor performance is driven more by circuit design, specialty materials and specialized manufacturing processes.

Innovation in pervasive ICs is measured in terms of precision, accuracy, bandwidth, efficiency and sensitivity. When combined with greater breadth and diversity of customers and end markets, these factors tend to result in more stable demand and pricing for pervasive semiconductors than processor-centric compute semiconductors. Additionally, capital expenditure requirements are generally significantly lower for manufacturers of pervasive semiconductors."
View attachment 54883

New products by AMD at GlobalFoundries obviously will reflect the change.
The node supplied by GloFo is not set in stone. If AMD will guarantee hundreds of millions of dollars to continue producing 14/12nm parts, you can be damn sure those lines will keep running. That's just how business works. Deals are made, and plans shift to make the most of opportunities.

EDIT: More to the point- GloFo and AMD are clearly not aligned on strategy. "Pervasive" means "low margin", and AMD aren't interested in that. They shut down SeaMicro, they didn't produce a Jaguar successor, they are charging high end performance and the high profits it can produce. They aren't going to change strategy just to make GloFo happy. They are going in different directions, and this is just an opportunity for both parties to make good business out of the divorce. AMD can secure supply for their lower performance chips (IODs, active interposers, low end laptop chips, last-gen console SoCs, whatever) in the middle of a global supply shortage, while GloFo can make a truckload of cash off a process they have already developed, on depreciated equipment. AMD aren't going to buy into FD-SOI just to make GloFo happy, no more than GloFo are going to restart 7nm FinFET development just to make AMD happy.
 
Last edited:

ASK THE COMMUNITY