Question [AT] AMD extends GloFo WSA to 2025

[jpiniero]

AMD and GlobalFoundries Wafer Supply Agreement Updated Once More: Now $2.1B Through 2025

www.anandtech.com

AMD has agreed to buy 2.1 billion worth of 12/14 nm wafers between 2022 and 2025. So another year and $500M on top of what they had agreed to in May.

This must confirm that Monet is coming. I personally don't think Monet is worth it but AMD thinks otherwise?

 

[Abwx]

Zen3 backport to 12nm would cos about same as 100 million Monets produced on 7nm. So if it is about money they won't even think Zen3 on 12nm. For cheap 12nm products AMD will use what they had on 14nm - Zen+.

A complete design is something like 400M, so a backport is undoubtly way less than 100M, moreover for a 4C chip with few graphics CUs.

Besides GF s 14/12nm libraries were designed such that AMD could eventually switch to TSMC if the former had trouble finalizing their 7nm, wich indeed happened, but still, AMD released Zen 2 at TSMC within the expected timeframe.

Also given large scale automation in EDA tools it wouldnt take long to backport a design to 12nm+, most time consuming would be to wait for the chip to be fabbed if ever two steppings are necessary.

 

[andermans]

I have a feeling that if there's an cut product coming it's more of an Alder (Raptor?) Lake-M competitor than the space Dali/Picasso is in.

I think Alder Lake-M is more of a match for Ryzen 3? I think Mendocino is lower end than that, especially since it only has 4 Zen2 cores and a tiny GPU. That is squarely in Dali/Pollock territory, and should mostly compete against Intel E-core only stuff (pentium silver etc.)

 

[NostaSeronx]

There is also the gapping hole for Geode, Sempron, etc.

Previous HP+LP
32nm PDSOI -> 28nm Bulk= ~6 million units, Low ASP
40nm Bulk -> 28nm Bulk = ~6 million units, Low ASP

Future HPC+ULP
[14-12] -> [7-6/5-4] -> 3 = ~3 to ~4 million units, High ASP
Based on supply/demand:
22FDX-> 12FDX -> 6FDX = ~8 to ~9 million units, Low ASP

Geode Lineage:
Geode GX 400 MHz, 3.6W TDP <-- 2002 // EOL: 2012
Geode LX 500 MHz, 3.6W TDP <-- 2005 // EOL: 2019
G-T16R 615 MHz, 4.5W TDP <-- 2011 // EOL: 2021
GX-210JA 1 GHz*2, 6W TDP <-- 2013 // EOL: 2023
GX-210KL 1 GHz*2, 4.5W TDP <-- 2016 // EOL: 2021
GX212JJ & GX215JJ 1.2-1.5 to 1.6-2.0 GHz*2, 6W <-- 2018 // EOL: 2021

Geode LX was introduced with the $45 SEP.
Bobcat, Jaguar, Puma, Puma+ had an average $50 price.
Stoney-Excavator had an average <$40 price.

Geode LX = ~100 mm2 (4cm*4cm package, paste rim on bottom of package looks like ~1cm*~1cm)
Bobcat = ~75 mm2
Jaguar/Puma/Puma+ = ~107 mm2 (larger die + 1.3x the cost of 40nm)
Excavator = ~125 mm2 (larger die + depreciated node(production insert when 28nm had lower than 40nm gate cost))

Why there is no G-series with Zen:
Zen = ~150 mm2 (Larger die + 1.95x the cost of 28nm)

Zen isn't in the linear-area IPC growth but in the exponential-area IPC growth region.
Jaguar on 14nm = 1.79+1.79= 3.58 versus 5.5 mm2 w/ CMT-Jag 14nm = ~2.7 mm2

How expensive to design, get processed wafers, and get low-power it makes sense to use 22FDX instead, as well as reviving the most successful niche.
AMD is 6th in wafer purchases and 1st in revenue for GloFo.
Qualcomm is 1st in wafer purchases and 2nd in revenue; 2021E is Qualcomm taking 1st place from AMD.
Low volume of wafers => lower utilization of Malta => Lower revenue to income contribution(It's outright negative).

In turn going to 22FDX means Dresden/22FDX, Malta/22FDX, Singapore/22FDX utilizing all AMD Foundry/Chartered Foundry supply chains. To GlobalFoundries this is more robust on maintaining high utilization and getting higher net income.

14LPP -- 50% Revenue = 1/4th of volume
28BLK -- 50% Revenue = 3/4th of volume
25% volume(14nm) versus 75% volume(28nm) -> 20%(1/2th 2H for no 28nm orders) volume for 2021, 10%(1/4th for reduction of 14nm) volume for 2022, 5-0%(1/8th for reduction and 2H zero% revenue contribution) volume 2023. As GlobalFoundries opts for nodes that have cheaper expansions.

Going forward:
14LPP = 84CPP/78CPP to 64CPP -> too small SATP-SAQP is needed for Fin.
12FDX = 84CPP(12FD)/64CPP(10FD FEOL as 12FD+) -> remains SADP for transistor.
^-- higher SRAM density for same mask layers. It is better to wait for 12FDX+ than to enter into 12FDX early. There is also potential for sSOI in 22FDX++, since the 2014 20FD node at Albany had high mobility/sSOI as well. There has been some Leti/STM slides indicating there is a 96CPP option on 22FDX+.

I think Alder Lake-M is more of a match for Ryzen 3? I think Mendocino is lower end than that, especially since it only has 4 Zen2 cores and a tiny GPU. That is squarely in Dali/Pollock territory, and should mostly compete against Intel E-core only stuff (pentium silver etc.)

Alder Lake-N => 8c+32eu at worst to 8c+96eu at best => 128-bit DDRx
Mendocino => 4c+6cu at worst to 4c+8cu at best => 128-bit DDRx

It is still competitive sort-of. I sort of expect AMD will get Zen4c(8c)+RDNA3(?cu) out for Meteor Lake-N thou.

Example of price points going forward:
R1606G - $100 ==> R3606G(Full enable) - $160, Dual-core w/ big GPU: $120
Ryzen 3 3250U - $60 ==> Ryzen 3 6250U(Full enable) - $120, Dual-core w/ big GPU: $80
Athlon 3000G - $49 ==> Athlon 6100G(Full enable) - $109, Dual-core w/ big GPU: $69
I'm not going to hunt down units per SKU to find out average ASP. Just know that it should be going up between 14nm <-> 6nm.

 

[andermans]

Alder Lake-N => 8c+32eu at worst to 8c+96eu at best => 128-bit DDRx
Mendocino => 4c+6cu at worst to 4c+8cu at best => 128-bit DDRx

Alder Lake-M is 2+8, no? Furthermore, with Zen2 with probably low clock speeds I'd count the 4 cores more like the E-cores? Additionally I was expecting more like 1 WGP/2 CU for the GPU, not 6/8 CU, though I don't think we have a good source right now.

(As an aside it will probably steal a tiny bit of market from the Picasso line, but Picasso->Renoir also doubled cores so who knows.)

 

[NostaSeronx]

Alder Lake-M is 2+8, no? Furthermore, with Zen2 with probably low clock speeds I'd count the 4 cores more like the E-cores? Additionally I was expecting more like 1 WGP/2 CU for the GPU, not 6/8 CU, though I don't think we have a good source right now.

(As an aside it will probably steal a tiny bit of market from the Picasso line, but Picasso->Renoir also doubled cores so who knows.)

Alder Lake-N is 0+8, though.

As shown below they need to chase for performance and in return get a higher average selling price:


A big Mendocino would fit in-between:
Ryzen 3 3200G = $99 @ launch
Ryzen 3 5300G = $160 @ launch (5600G=$260, 5700G=$360)

$49 of 3000G -> >$100, enough to eat the max ~1.94x increase in cost. (or move to 6nm and not reduce the price for increased margins)

---- Not to andermans ----
The above picture, specifically:
There hasn't been any indication that GlobalFoundries increased 22nm prices. While GF 12nm/14nm is only the trend from 28nm TSMC/Samsung to 12nm/16nm TSMC/Samsung given GF's lower price.

28nm to 22nm = ~0.97x (TSMC/UMC)
28nm to N-1 FinFET = ~1.62x (TSMC/Samsung)
N-1 FinFET to N+1 FinFET = ~1.94x (TSMC) || 6nm has similar decreasing as below so min price is: ~$6909.5 (reduction via mask: GDS to 193i = >88 masks, GDS to EUV = >70)
N+1 FinFET to N+2 FinFET = ~1.42x (TSMC)

 

[jpiniero]

I think Alder Lake-M is more of a match for Ryzen 3? I think Mendocino is lower end than that, especially since it only has 4 Zen2 cores and a tiny GPU. That is squarely in Dali/Pollock territory, and should mostly compete against Intel E-core only stuff (pentium silver etc.)

The full M die is branded i7 so it implies it's going to be more pricey. Thinking about it more I don't think M is going to be that widespread given that Windows Tablets are dead and most of the 2-in-1s are thicc bois that can easily take a 15W processor so it's not clear a cut down die is really needed for them. The salvage from the 2+8 and the 6+8 is definately a competitor for the low end.

 

[andermans]

Alder Lake-N is 0+8, though.

As shown below they need to chase for performance and in return get a higher average selling price:
View attachment 55305

A big Mendocino would fit in-between:
Ryzen 3 3200G = $99 @ launch
Ryzen 3 5300G = $160 @ launch (5600G=$260, 5700G=$360)

$49 of 3000G -> >$100, enough to eat the max ~1.94x increase in cost. (or move to 6nm and not reduce the price for increased margins)

Yes, the cost will increase but I doubt by that much. Remember that Pollock/Dali cost AMD <= $10 to make. If I take a yield calculator and an assumption of a ~100 mm2 chip that gives me about $17 for a 7nm chip. So the cost can definitely increase but it is more about product placement than about fab costs in the segment you're looking at. The place where the production prices really matter are for the $200-$300 entry level notebook market, where CPUs like that will be sold much cheaper to OEMs.

As an aside I didn't hear anything about Mendocino having DDR support at all? ExecuFix only mentioned LPDDR so I wonder if it will come to desktop/DIY at all. If it doesn't desktop pricing is fairly moot.

 

[xilli_fiberbit]

A another car maker will use a custom AMD APU like Tesla, to replace an intel atom-based infotainement :

https://twitter.com/x/status/1477249165818306562

So could be too produced at GF in 12nm, we will see if AMD talks about it at CES

 

[Abwx]

Yes, the cost will increase but I doubt by that much. Remember that Pollock/Dali cost AMD <= $10 to make. If I take a yield calculator and an assumption of a ~100 mm2 chip that gives me about $17 for a 7nm chip. So the cost can definitely increase but it is more about product placement than about fab costs in the segment you're looking at. The place where the production prices really matter are for the $200-$300 entry level notebook market, where CPUs like that will be sold much cheaper to OEMs.

As an aside I didn't hear anything about Mendocino having DDR support at all? ExecuFix only mentioned LPDDR so I wonder if it will come to desktop/DIY at all. If it doesn't desktop pricing is fairly moot.

A 4C/8T Zen 2 or 3 with a few CUs wouldnt be more than 130mm2 on GF s 12nm+ and would cost barely 7$/die, besides efficency would be quite good enough at 0.6x the 14nm power vs 0.5x for TSMC s 7nm process.

 

[andermans]

A 4C/8T Zen 2 or 3 with a few CUs wouldnt be more than 130mm2 on GF s 12nm+ and would cost barely 7$/die, besides efficency would be quite good enough at 0.6x the 14nm power vs 0.5x for TSMC s 7nm process.

Note that Dali is 150 already. Adding more & bigger cores or CUs wouldn't help. The problem is these are very low margin products. Backporting zen 2/3 (note zen3 has never appeared with a CCX < 8 cores) and RDNA2 would be pretty expensive in design and tooling costs.

 

[Abwx]

Note that Dali is 150 already. Adding more & bigger cores or CUs wouldn't help. The problem is these are very low margin products. Backporting zen 2/3 (note zen3 has never appeared with a CCX < 8 cores) and RDNA2 would be pretty expensive in design and tooling costs.

Dali use 14nm or eventually 12nm but without the density advantage since they use the same layout for Zen and Zen+, that s why the 14nm 1000 series have same area as the 12nm 2000 series.

If they backport a 7nm design to 12nm+ they will inherently use as much density as possible, FTR a proper layout of Dali at 12nm+ would be something like 110mm2 at most, so even if a 4C Zen3 + 4 RDNA2 CUs is 150mm2 that would still be the same cost as Dali but at higher margins given the perfs and perf/watt improvements.

 

[naukkis]

If they backport a 7nm design to 12nm+ they will inherently use as much density as possible, FTR a proper layout of Dali at 12nm+ would be something like 110mm2 at most, so even if a 4C Zen3 + 4 RDNA2 CUs is 150mm2 that would still be the same cost as Dali but at higher margins given the perfs and perf/watt improvements.

Look how it goes with Rocketlake - bakcporting advanced node to older node resulted absolutely massive sized core. And AMD itself have said that features like 256 bit SIMD weren't feasible until 7nm process - backported 7nm design would loose big part of frequency potential so resulting performance would probably not be any better than with proper 14nm Zen design.

 

[jpiniero]

Look how it goes with Rocketlake - bakcporting advanced node to older node resulted absolutely massive sized core. And AMD itself have said that features like 256 bit SIMD weren't feasible until 7nm process - backported 7nm design would loose big part of frequency potential so resulting performance would probably not be any better than with proper 14nm Zen design.

And yet they need to do something if they are still interested.

 

[Abwx]

Look how it goes with Rocketlake - bakcporting advanced node to older node resulted absolutely massive sized core. And AMD itself have said that features like 256 bit SIMD weren't feasible until 7nm process - backported 7nm design would loose big part of frequency potential so resulting performance would probably not be any better than with proper 14nm Zen design.

RKL is not comparable since its 512b SIMD is even bigger.

256b wasnt feasible at 14nm because of power comsumption, 8C design with big caches and undoubtly time to market being too long.

Comparatively to 14nm 12nm+ has 50% higher density and 40% lower power comsumption.

Besides frequency is not a problem for low cost parts, and AMD said that is was harder to extract high frequencies from 7nm than from 12nm due to inherently lower voltage, Zen 2/3 clock higher than ZEN because they gave up some density gains within the cores.

 

[naukkis]

And yet they need to do something if they are still interested.

Interested in what? If they want to do 14/12nm designs they use their 14nm IP or design new - backport or new cpu design for 12nm would cost maybe $300m. Even Intel tried as much as possible to not do something so stupid as backporting cpu core to older process - AMD sure doesn't even consider it as possibility.

AMD cpu designers are busy doing future cpu revisions on newer processes, older processes have what they have designed for them.

 

[jpiniero]

Interested in what?

The ultra low end market where Dali and Picasso are now.

 

[naukkis]

The ultra low end market where Dali and Picasso are now.

AMD's only problem is that they are capacity limited. So they produce Zen+ based products on GF and Zen2/3-based product on TSMC - and for their cheap 7nm products they probably use Zen2-cores as they are smaller than Zen3 and as they aren't tuned to maximum performance whether they used Zen2 or 3 doesn't make any difference.

 

[jpiniero]

AMD's only problem is that they are capacity limited. So they produce Zen+ based products on GF and Zen2/3-based product on TSMC - and for their cheap 7nm products they probably use Zen2-cores as they are smaller than Zen3 and as they aren't tuned to maximum performance whether they used Zen2 or 3 doesn't make any difference.

Right, which is why they need to use 12 nm GloFo for their cheap products. They don't have a cheap 7 nm product right now, even 7 nm Ryzen 3 is pretty rare. Plus you have the TSMC price hikes.

 

[Abwx]

backport or new cpu design for 12nm would cost maybe $300m.

In 2013 a full design was 300-400M according to an AMD exec, so a backport is certainly not as costly as 100M, eventually 50M.

That said they have another solution since they recently tested Samsung processes using a Radeon GPU, guess that it was also some indirect indication to a somewhat greedy TSMC that they could use Samsung for some products, and is also the prove that a port to another process is not that costly.

 

[NostaSeronx]

Yes, the cost will increase but I doubt by that much. Remember that Pollock/Dali cost AMD <= $10 to make. If I take a yield calculator and an assumption of a ~100 mm2 chip that gives me about $17 for a 7nm chip. So the cost can definitely increase but it is more about product placement than about fab costs in the segment you're looking at. The place where the production prices really matter are for the $200-$300 entry level notebook market, where CPUs like that will be sold much cheaper to OEMs.

As an aside I didn't hear anything about Mendocino having DDR support at all? ExecuFix only mentioned LPDDR so I wonder if it will come to desktop/DIY at all. If it doesn't desktop pricing is fairly moot.

That $10 factor is ignoring the whole equation.
GF-14nm => 9.60591133 for 150 mm2 for all dies.
Good dies => 10.103626943 for 150 mm2 for good w/ defects dies and 12.58 for 150 mm2
<--> Extensive margin plus packaging costs ~4.7x but rounding to ~5x
9.60591133 * 5 => 48.02955665
10.103626943 * 5 => 50.518134715
12.580645161 * 5 => 62.903225805
These are minimums, not maximums hence why R1606G = $100 and Ryzen 3 3250U = $60.

Even if there is only LPDDR5, there is still cases where desktops can appear;
A9-series 15/25W Stoney(125 mm2 die) desktop board for example:


RPi 400 7W Hudson(120 mm2 die) desktop board for example:


Ryzen 3 3250U/3150U/3050U 15W Dali(149 mm2 die) desktop motherboard:


Mendocinio:
All dies => $23.480149254 * 5 => $117.400746269
Good dies => 24.70947644 * 5 => $123.547382199 && 28.690030395×5=$143.450151976
These are minimum bands for 7nm, but can easily be the maximum band for 6nm.

--------

Right, which is why they need to use 12 nm GloFo for their cheap products. They don't have a cheap 7 nm product right now, even 7 nm Ryzen 3 is pretty rare. Plus you have the TSMC price hikes.

However, GlobalFoundries appears to be down ramping FinFETs at Malta. To fit in FDSOI and RFSOI, which is why they paid GlobalWafers to revive the SunEdision/MEMC SOI(A SmartCut licensee) Fab: https://gf.com/press-release/global...-partnering-expand-semiconductor-wafer-supply :: "The 300mm pilot line is on track to be completed in Q4 this year. => In particular, the 300mm wafers made at GWC’s MEMC site in Missouri will be used at GF’s most advanced manufacturing facility, Fab 8 in Malta, New York"

Going forward, the only nodes available post-shortage at GlobalFoundries Fab8 is 45nm PD/FD, 22nm FD, 12nm FD. Which has a higher volume of customers being >100 on SOI vs the <10 on FinFET.

Fab 1 - Fab 7 - Fab 8 => 73% of revenue is above 14nm FinFET, only 27% is below 22nm FDSOI. This is compared to TSMC which has more Fabs than GF; 65% of revenue is FinFET, and 45% of revenue is Planar.

Without exclusivity commitments GlobalFoundries would be like UMC: 0% on FinFET, 100% on Planar/DDC.

 

[Joe NYC]

A another car maker will use a custom AMD APU like Tesla, to replace an intel atom-based infotainement :

https://twitter.com/x/status/1477249165818306562

So could be too produced at GF in 12nm, we will see if AMD talks about it at CES

Maybe the 2nd carmaker is what's behind the additional orders from GF.

As Bits and Chips tweeted before, GF has the AMD orders partially categorized as Automotive...

 

[NostaSeronx]

Maybe the 2nd carmaker is what's behind the additional orders from GF.

As Bits and Chips tweeted before, GF has the AMD orders partially categorized as Automotive...

AMD is listed under given examples of market to customer to application growth:
45RFSOI/45RFe/22FDX for Radio/Modem Chipsets
45CLO/45LPCLO for Silicon Photonics
22FDX for Battery Management/mmWave Radar
22FDX for Connectivity/Touch/Re-drivers/Power

GlobalFoundries is targeting single-source only opportunities going forward. Providing the only source of performance-class FDSOI w/ feature richness on the market, helps a lot.


Picasso = Not Automotive-certified, can be used as a non-critical pipe-cleaner product.
Rembrandt = Automotive-certified
Van Gogh/Mendocino = Automotive-certified

The only processors so far are Picasso chips. Primarily because only up to 2022 is when there will be consistent FinFET supply, and sometime in 2023 it will be gone.

If it can be done at another fab better, then GlobalFoundries isn't going to touch it. The huge elephant in room everyone is ignoring, GlobalFoundries FinFETs suck across the board. While cheap to customers, no profits to foundry.

Even SMIC is better than GloFo: (GlobalFoundries < SMIC, with UMC bum-rushing ahead in marketshare for 2021)
14SFE -> 14SF Plus -> { 8/10SFE -> 8/10SF Plus -> 7 }

Only used for decoying:
Option 1 -> Pay least, get least || Bulk
Option 2 -> Pay more, get most || FDSOI
Option 3 -> Pay most, get least || FinFET

----
Found GlobalFoundries selling prices at followers(after 4 years) edge from December 2016:

FDSOI capacity at Fab 1, via expectations. It is best to assume that had been delayed to 2023. Also, the number above that is the 300mm FDSOI wafer cost; 2016: 550+mask&processed steps vs 2020:420+mask&processed steps
Premium EPI = $200, SOI-to-GF: ~$400
$350 premium in 2016, $200 premium in 2020+

Then using the STM/EU roadmap:

22FDX++ at EU = Q1 2023(max)
12FDX at EU = Q1 2024(max)

We are overdue a 12LP++ node, if the American roadmap is going to introduce 12FDX. It can be used as a hypothetical 12LP++ and even 12LP+++ node if Tensile Si on Insulator is a target;


12LP+ = +20% over 14LPP and -40% over 14LPP
12FDX(2019) = +30% over 14LPP and -53% over 14LPP with a >1.2x increase in good dies.
12FDX(2023-2024) = +60% over 14LPP and -75% over 14LPP, if 64CPP is a thing as I expect it is; >1.4x increase in good dies.

 

[moinmoin]

In 2013 a full design was 300-400M according to an AMD exec, so a backport is certainly not as costly as 100M, eventually 50M.

We are in 2022 though, and the upfront cost depends on the node, and for the latest nodes that cost is what nearly doubled each gen. Now we don't know how high that cost is for 14LPP/12LP or whatever at GloFo, but I doubt AMD would pull a RKL and just do a straight backport for a part intended to be sold unchanged for years. 50M is lowballing a lot.

 

[Mopetar]

Maybe the 2nd carmaker is what's behind the additional orders from GF.

As Bits and Chips tweeted before, GF has the AMD orders partially categorized as Automotive...

Could be. Really depends on the manufacturer. The biggest companies make around 10 million automobiles per year. However, once you move out of the top 10, it's down to around a quarter of that amount so there's a lot of variability.

I also don't know what percentage of vehicles a company sells would require an AMD CPU either. Frankly I'm not even sure why they would need an x86 CPU when a cheap ARM processor would be more than sufficient in my mind.

 

[moinmoin]

Frankly I'm not even sure why they would need an x86 CPU when a cheap ARM processor would be more than sufficient in my mind.

The x86 CPU is accompanied by a Radeon GPU. Tesla tooted the "it's like PS5" horn, I guess delivering both the CPU and GPU "as seen in popular gaming consoles" is now a competitive advantage AMD offers.