Article [Anandtech] AMD Amends Wafer Supply Agreement with GlobalFoundries

[NTMBK]

https://www.anandtech.com/show/1391...th-globalfoudries-set-to-buy-wafers-till-2021

Committed to buying 12nm and above dies until 2021. I guess that means lots of I/O dies for Ryzen, and some budget parts on old processes? Or will we see some crazy FD-SOI wafer-sales-special? (I guess the former.)

 

[Kenmitch]

What's up with anandtech.com? Takes forever for the page to load now.

 

[Abwx]

https://www.anandtech.com/show/1391...th-globalfoudries-set-to-buy-wafers-till-2021

Committed to buying 12nm and above dies until 2021. I guess that means lots of I/O dies for Ryzen, and some budget parts on old processes? Or will we see some crazy FD-SOI wafer-sales-special? (I guess the former.)

GF statement :

“AMD remains an important strategic partner as we reshape our portfolio to intensify investment in the technologies that provide the most value to customers. We are proud to be a critical supplier of AMD’s current-generation 14nm and 12nm products, and we look forward to continuing to play a key role in their next-generation products.”

This require more than suppling 14nm based I/Os, at some point they ll have to use smaller nodes than what they currenty have if they want to really play said key role in AMD s next gens, notice the plural...

 

[NTMBK]

GF statement :

“AMD remains an important strategic partner as we reshape our portfolio to intensify investment in the technologies that provide the most value to customers. We are proud to be a critical supplier of AMD’s current-generation 14nm and 12nm products, and we look forward to continuing to play a key role in their next-generation products.”

This require more than suppling 14nm based I/Os, at some point they ll have to use smaller nodes than what they currenty have if they want to really play said key role in AMD s next gens, notice the plural...

I would say providing the northbridge exactly qualifies as playing a "key role".

 

[Mopetar]

This require more than suppling 14nm based I/Os, at some point they ll have to use smaller nodes than what they currenty have if they want to really play said key role in AMD s next gens, notice the plural...

If the IO die is really just all of the IO connections and there's no cache, etc. then AMD doesn't gain much in going to a smaller node since part of what they've relegated to that die doesn't shrink down.

Since older processes are going to be cheaper (less demand and after investment costs are recouped less incentive to keep prices high) AMD isn't going to hurt from sticking with it long term, especially if it's mature and the yields are good.

It makes it easier for AMD to design for as well since they can just reuse what already works and make minor tweaks and additions.

 

[AtenRa]

Ryzen 2 Mobile at 12nm just got launched, they will be in production the entire 2019 and in to 2020.
I could see new ChromBook Zen+ cheap CPUs at GloFo 12nm in 2020-2021.
Also, embedded 12nm Ryzen 2 APUs will continue to be produced for another 5+ years at GloFo.
Another product could be Desktop motherboard Chipsets.
And dont forget 12nm GPUs.

 

[gorobei]

if amd is following the active interposer butterdoughnut plan, then a 12/14nm interposer with io built in will be used for the next gen server parts. that is a large bit of silicon with a ton of tolerance for defects, which will chew through a good chunk of the wsa.

 

[naukkis]

AMD could also produce chipsets at Globalfoundries.

 

[Abwx]

I would say providing the northbridge exactly qualifies as playing a "key role".

As key role as providing the SKU substrate and case..

Seriously.?.

If the I/O device use 20%of the power they can eventualy stretch the thing by switching to 12nm for the 7nm+ Zen, but at the next node from TSMC the I/O would amount to 35% of the power if not iterated to a smaller than 12nm node.

And we re not even talking of mobile wich require both decent perf and perf/watt metrics, in the same way they can buy some time by using GF s 12nm process but the competition wont stay idle in the meantime, once they get their next node AMD will have to follow suite and shrink their mobile APUs to 7nm.

 

[NostaSeronx]

The 7nm node removal includes no fees. So, production at 7nm and below is fully flexible and there is no penalty. However, the 12nm node and above is not fully flexible.

My speculation on the pricing;
Existing 14nm/12nm FinFET pricing is fixed with static decreases per year.
While 22FDX/12FDX pricing is dynamic to lure AMD into being a double digit customer.

Further speculation;
If AMD produces on the FDX side, the penalties for 14/12 -nm FinFET flexibility is written off.

Malta is operated at a loss with FinFETs. So, persuading AMD to move to Chengdu/Dresden on new products is more worthwhile.

 

[amd6502]

Further speculation;
If AMD produces on the FDX side, the penalties for 14/12 -nm FinFET flexibility is written off.

Malta is operated at a loss with FinFETs. So, persuading AMD to move to Chengdu/Dresden on new products is more worthwhile.

Is there anything in the amendment's wording about 22FDX(or 12/14FDX).

I would like to see them go 22FDX or better for small SoC's (under 2B transistors), especially for new sub 10W A-series. You mentioned in another post 12FDX may not be ready for GPU portion due to 3D SRAM not yet being ready.

Till 3D SRAM comes along at GloFo, it is unlikely that the I/O chip will have ROPs or any display capability.

That would kind of rule out any APU or GPU on this leading edge FDSOI (or I'm not correctly understanding the above).

I'm just guessing/hoping this wouldn't be an issue on 22FDX either, otherwise it could explain why there is no 22FDX Stoney Ridge successor yet.

Market timing and delays are really critical, especially for higher margin or high volume products, and even consumer mainstream. So with risks of delays and other unknowns of FDSOI it's not a mystery that they went finFET with the IO hubs.

 

[amd6502]

Given that 7nm design costs are (and will continue to be) astronomical does anyone think there will be a new Zen uarch on 12nm-22nm (either finfet or fdsoi)? Same question for a new non-Zen uarch (XV+, K12/arm, risc-v, jag++, a hybrid virtual ccx, or something completely different).

 

[DrMrLordX]

Given that 7nm design costs are (and will continue to be) astronomical does anyone think there will be a new Zen uarch on 12nm-22nm (either finfet or fdsoi)? Same question for a new non-Zen uarch (XV+, K12/arm, risc-v, jag++, a hybrid virtual ccx, or something completely different).

Probably not. New designs take time, money, and people; even if 12nm design costs are lower than 7nm, you still take away some of your engineers' time to do this thing. Anything "cheap" that AMD produces on 12nm will increasingly face threats from the ARMy that appears to be deploying everything on 7nm. I don't see low-cost, low-power 12nm Zen++ doing very well against Qualcomm's products. They aren't going to mess with K12 anymore (I don't think), they show no signs of interest in RISC-V, and XV/Con cores have been consigned to the rustheap.

The only way I see them producing more 12nm Zen-based chips is if they continue to iterate on Picasso for customers that want it at higher clocks/lower power.

 

[NostaSeronx]

Is there anything in the amendment's wording about 22FDX(or 12/14FDX).

I would like to see them go 22FDX or better for small SoC's (under 2B transistors), especially for new sub 10W A-series. You mentioned in another post 12FDX may not be ready for GPU portion due to 3D SRAM not yet being ready.

That would kind of rule out any APU or GPU on this leading edge FDSOI (or I'm not correctly understanding the above).

- They use the word "node" which means anything and everything most of the time.
12nm and above, includes 12LP/14LPP of the leading-edge roadmap and 12FDX/22FDX/28BLK of the advanced roadmap.

- The GPU chiplet side would want 3D SRAM for a shared VCN/ROP I/O+Uncore chiplet. This way the I/O chiplet can be used for CPUs(VCN/ROPs disabled) or GPUs(VCN/ROPs enabled) and hybrid-CPU+GPU.

I'm just guessing/hoping this wouldn't be an issue on 22FDX either, otherwise it could explain why there is no 22FDX Stoney Ridge successor yet.

The lack of 22FDX is mostly do to capacity ramp.

2017 -> 1x capacity
1H2018 -> 2x capacity
2H2018 -> 4x capacity
1H2019 -> 5x capacity
2H2019 -> 6x capacity
1H2020 -> 8x capacity
2H2020 -> 10x capacity

https://imgur.com/vhVxrj5

2H2019 looks like the best point to start production if one wants cheap processed 22FDX wafers. I also think that GloFo is still toning down the variability, and increasing Digital/Analog/RF figures of merit.

On another note, with FinFETs pushed out of R&D. That leaves room for High Mobility sSOI to be ramped early for 22FDX++ or something.

 

[darkswordsman17]

AMD's semi-custom unit would be the best use to leverage 14/12nm. They already know it, they already made CPUs, GPUs, and APUs (I think the PS4/One SoCs were at TSMC though, weren't they?) on it, so they won't even need to do much real design work.

 

[amd6502]

AMD's semi-custom unit would be the best use to leverage 14/12nm. They already know it, they already made CPUs, GPUs, and APUs (I think the PS4/One SoCs were at TSMC though, weren't they?) on it, so they won't even need to do much real design work.

Yes, jaguar console APU got 16nm shrink with TSMC.

NG console may be very similar to Zen+/Picasso with big GPU upgrade. I wonder if console sales count towards the WSA.

On another note, with FinFETs pushed out of R&D. That leaves room for High Mobility sSOI to be ramped early for 22FDX++ or something.

I don't think that would happen. But I don't see why both wouldn't be developed. Another benefit is you could use the cheaper nodes as testing grounds for concepts; so big core mainstream finfet branch would benefit from FDX branch.

 

[krumme]

Aliens from the Galaxy Micron, sector 14 ordered 100.000, 22 FDX wafers 5 years ago. That's why GF is in full swing.

 

[Headfoot]

this announcement is good if only for the simple reason that it gives actual facts on what is in the "WSA." I swear, the WSA has the WORST ratio of "Cited authoritatively by forumites" to "Actual knowledge of the thing they're talking about" of any topic on this forum. "x86 license agreement" being probably just as bad.

 

[amd6502]

this announcement is good if only for the simple reason that it gives actual facts on what is in the "WSA." I swear, the WSA has the WORST ratio of "Cited authoritatively by forumites" to "Actual knowledge of the thing they're talking about" of any topic on this forum. "x86 license agreement" being probably just as bad.

Put on your Nosta hat and try again.

 

[NTMBK]

Put on your Nosta hat and try again.

something something FDSOI

 

[itsmydamnation]

something something FDSOI

something something zen bad, cmt best.

 

[dark zero]

CMT is deader than dead, heck, even the Cat Core has more future than BD!