GlobalFoundries Stops All 7nm Development

GlobalFoundries on Monday announced an important strategy shift. The contract maker of semiconductors decided to cease development of bleeding edge manufacturing technologies and stop all work on its 7LP (7 nm) fabrication processes, which will not be used for any client. Instead, the company will focus on specialized process technologies for clients in emerging high-growth markets. These technologies will initially be based on the company’s 14LPP/12LP platform and will include RF, embedded memory, and low power features. Because of the strategy shift, GF will cut 5% of its staff as well as renegotiate its WSA and IP-related deals with AMD and IBM. In a bid to understand more what is going on, we sat down with Gary Patton, CTO of GlobalFoundries.
https://www.anandtech.com/show/13277/globalfoundries-stops-all-7nm-development
 
There had been mumblings of this through the back-channels for a while now, but this is the first public statements I have seen about it.
 
i understand the reasons they gave but the one about losing customers for 7nm with no viable path to 5 and beyond is going to be the same result just right now instead of later.. you know why would a new customer waste their time tuning their chips for what theyre saying is a dead end process.
 
i understand the reasons they gave but the one about losing customers for 7nm with no viable path to 5 and beyond is going to be the same result just right now instead of later.. you know why would a new customer waste their time tuning their chips for what theyre saying is a dead end process.

Yup, seems very short sighted. Surely all high end mobiles will be 7nm next year?
 
And then there were three.
I’m not sure I would call Intel a foundry at all. And Samsung is hardly what you would call a pure-play foundry. So it depends on your definitions. Pity, since they had some nice things lined up, but economic realities being what they are, it is probably better if they spend their research dollars elsewhere. Competing head-on with the giants must be brutal.
 
Probably i'd be very hard for GF to gain customers beyond AMD for 7nm. There'd be just a few companies willing to pay the price anyway, and one of them is Samsung
 
GF is one fifth the size of TSMC. That means there are a lot less wafers to amortize R&D over.

At the same time, while 7nm looks feasible with fin-fets, 5nm will require gate all around (GAA) to keep leakage in check. GAA is unlikely to improve transistor performance since you have more capacity in the gate and lower drive current. So you only get a density increase, not faster or lower power devices.

GF has state of the art FD-SOI capability, which positions them well for IoT, automotive and upcoming mixed signal SOCs (5G). They probably see an easier way forward, beyond 7nm, with planar FD-SOI than fin-fets.

Cheers
 
Last edited:
On one hand, it seems this might affect either Navi or Picasso's release date (Picasso = Raven Ridge successor).
Anandtech's article about AMD's efforts for 7nm claims there was one design expected to tape out in Q4 2018 that was cancelled. This means it would be something with a release date between Q2 and Q4 2019.
Assuming the first Navi (Navi 10?) is still coming out Q1, that shouldn't be it. Those very old but still very accurate slides from videocardz mention a lower-end Navi 11, and that might be the one affected. One other possibility is "Picasso", the Zen2 APU which also doesn't seem to have any fixed release window so far. The fact that AMD's Hotchips 2018 presentation focused on a seemingly revamped "Raven Ridge 2018" could indicate that Picasso might be coming later than expected.



On the other hand, getting rid of the Wafer Share Agreement with GlobalFoundries might be a godsend for AMD, on the long run.
GF's track record with state-of-the-art processes has been terrible, at least since the fab spinned-off from AMD 9 years ago.
32nm - delayed and with yield issues
28nm - delayed, and underperforming to the point of AMD only being able to use it ~3 years after TSMC's equivalent for GPUs.
20nm - cancelled, killing AMD's Skybridge project in the process.
14XM - cancelled, had to purchase the 14LPP IP from Samsung
14LPP (Samsung) - delayed, reportedly with worse efficiency and performance than Samsung's implementation
10XM - cancelled, supposedly to accelerate 7nm and bring it at a more competitive timing against TSMC and Samsung
7nm - Yeah well.. we decided we won't do this after all.

Seems rather clear to me that GlobalFoundries never had the talent and/or resources to compete with TSMC, Samsung and Intel at the high-end processes.
AMD has been suffering greatly from GF's inability to compete on high-end processes, as we saw with the clocks that Polaris and Vega could achieve before hitting enormous efficiency walls.

On GlobalFoundries' side, selling all those high-end-process chips at a very low price-per-mm^2 to AMD may not have been very profitable, if at all. Even less if what AMD has to pay is aligned with the process' performance.
AMD reportedly was getting their 14nm GPUs at a very low price, but GF's intrinsic low performance eventually meant AMD couldn't get a desirable ASP for their GPUs either (pre-mining craze, of course), since there's much more in a graphics card than the SoC.
In the end, maybe no one was really happy with the Wafer Share Agreement. Align that with Mubadala starting to demand getting some of their money back, and the shelving of high-end processes starts to make sense.
 
Last edited by a moderator:
I’m not sure I would call Intel a foundry at all. And Samsung is hardly what you would call a pure-play foundry. So it depends on your definitions. Pity, since they had some nice things lined up, but economic realities being what they are, it is probably better if they spend their research dollars elsewhere. Competing head-on with the giants must be brutal.

In terms of being manufacturers at the feature sizes of the 14/10nm range and still being on the process node treadmill, are there any others?
With the number of remaining nodes on standard materials dwindling and the cost of them rising, I'm curious how much oxygen is left at these heights of capital-intensive global-volume manufacturing.
There's what, perhaps 1-2 nodes at this rate needing double or more the cost and volume on silicon? Curious whether that leaves enough room to breathe for 1 or 2 players at that point.


On one hand, it seems this might affect either Navi or Picasso's release date (Picasso = Raven Ridge successor).
AMD did state its roadmaps had not changed, though resorting to corporate speak touting how they had not changed the most fluid sort of public guidance they have can hide varying levels of disruption. Losing a chunk of wafer starts that AMD had ready access for could constrain its flexibility in filling the gap Intel's process woes had opened up.

On the other hand, getting rid of the Wafer Share Agreement with GlobalFoundries might be a godsend for AMD, on the long run.
GF's track record with state-of-the-art processes has been terrible, at least since the fab spinned-off from AMD 9 years ago.
32nm - delayed and with yield issues
28nm - delayed, and underperforming to the point of AMD only being able to use it ~3 years after TSMC's equivalent for GPUs.
To add some recollection of the prior nodes AMD had deployment issues at 90nm SOI that I saw it stated IBM had to help resolve. That said, 90nm and its refinements kept AMD's highest performing chips at that node, given AMD's problems with 45nm. The successor chips on 45nm did not reach the same clocks, and AMD's SRAM density significantly lagged at that node--for which I've seen variability due to SOI and/or 300mm wafers contributing. Possibly some of that went into why Bulldozer's 45nm debut was cancelled.

Given the long lead times, 32nm and AMD's choices were baked-in by the time AMD decided on spinning off the fabs. AMD's over-extension that led to asset-light being necessary so soon likely did not give GF a good buffer at the start, either.
28nm was to various degrees still bound to some of the device choices made for the 32nm node, so AMD may still deserve some blame.

At 20nm, I believe the decision to stick with gate-first happened against the wishes of AMD and GF's partners in its foundry alliance so that might be where AMD can be let off the hook.

* late edit:
Correction: The 300mm wafer and SOI double-whammy was 65nm. 45nm seemed to have some carry-over malaise when it came to cache density, which I did see some hint as difficulty getting the SRAM cells to meet error rates needed for the server market on SOI unless AMD sacrificed density.
 
Last edited:
In terms of being manufacturers at the feature sizes of the 14/10nm range and still being on the process node treadmill, are there any others?
No, there isn't.
That's not the full picture though. There are sophisticated fabs for memory, and it is not obvious how FDSOI will evolve over time. Also, there are initiatives within mainland china that may or may not push to the frontline. But when it comes to cutting edge processes for logic, it is now basically TSMC, Samsung and Intel, with only TSMC and Samsung having any real foundry contract business to speak of, and TSMC of course dominating that. Intel certainly has the funds to stay at the forefront, but it is not quite as clear that it will make economic sense for them to do so much further than their projected 7nm process. Samsung will push their lithographic technology driven both by their enormously successful memory business, and their own SoC needs. TSMC advances have been fuelled by Apple, and TSMC have executed well enough that competing with them on technology would seem increasingly difficult as they have won a broad customer base for their advanced processes and packaging.

In spite of the attention the leading edge lithography nodes get, most ICs in terms of volume and revenue both is manufactured on more established nodes.
If you happen to have an interest in the particular areas that have most benefitted from advances in lithographic process tech however, the future beyond the next five years looks rather grim. For someone my age, it feels like an era drawing to a close.
 
Why would they be dead? There's a ton of money in producing all sorts of chips for customers around the world that isn't necessarily bleeding edge CPUs and GPUs. You'd be surprised the amount of chips are produced on "old" processes like 28nm.
Maybe so but they're about to lose what I understand to be by large margin their biggest customer: AMD who are going to 7nm on both CPU & GPUs.

If they canned the apparently nearly ready 7nm because it was going to be too expensive to finish off and they need to make a profit, do they have customers to make up for that huge loss of revenue in 14/12nm?

Cutting only 5% of staff and starting probably expensive development on several new techs (RF etc) that might come to fruition some undefined time in future doesn't seem like a very viable way to replace that loss of AMD revenue.

Maybe I'm wrong & they'll be fine but I feel like it'd be more viable to complete bringing 7nm to market while announcing slower/no further development after that.
 
We don't know the 7nm yields of TSMC or Samsung. We know Intel struggles with it (they call it 10nm). So you have a massive investment in tooling, and an uncertain chance of succeeding.

Then there is the amount of customers. As the price goes up, the number of customers goes down. 10/14nm saw half the volumne of 28nm. Doing 7nm with quadruple patterning requires in excess of 80 mask sets for a full metal stack SOC. Even if your yield is adequate, few customer can afford the mask sets, we're talking Apple, Qualcomm, Nvidia and AMD.

By ditching 7nm, GF is choosing a path of significantly lower financial risk. Their 22FDX process has performance of a 14/16nm process but the cost of 28nm. They have chosen to move forward with their 12FDX, which will have good performance, good density and excellent power characteristics, - instead of betting on 7nm.

Cheers
 
Last edited:
Back
Top