GlobalFoundries Halts 7nm Development

[Stefan Payne]

2 hours ago, leadeater said:

Not everything needs ultra density or ultra performance either, 

Not everything can benefit from that as well.

There are things you can not shrink much further especially if we are talking about Analogue stuff and driver stuff that drives signals over the PCB.

[Bananasplit_00]

Wow that's strange, guessing they hit a snag somewhere and unlike Intel they decided to just give up I guess

[cj09beira]

50 minutes ago, Stefan Payne said:

Ever heard about TSMC??

You know that Samsung also does Chip manufacturing?
 

AMD doesn't need GF, never did.

GF always needed AMD.

 

See the Problem?

 

No, there is still SAMSUNG.

samsung is too focused on low power mobile chips for it to be useful for a high performance server/consumer cpu 

[leadeater]

43 minutes ago, Stefan Payne said:

There are things you can not shrink much further especially if we are talking about Analogue stuff and driver stuff that drives signals over the PCB.

Pretty much all those are on totally different fab processing type though, you won't see a MOSFET being made on a modern TSMC process or even at TSMC at all. But you could see a PCIe switching chip for Broadcom being made at TSMC.

[Mihle]

If their 7nm was as close to being complete as it seemed like they where, I don't know why they just finished the development of 7nm and then stopped to shrink. Only thing that I can see if they litterally had no way at all to finance it.

 

Are they sure they litterally stopped 7nm or is it actually just spending less money and not rushing it so much, take the time it needs to save money?

[leadeater]

2 hours ago, S w a t s o n said:

 

Honestly, I feel like they cut their losses too early. 1st gen 7nm was so close to being ready and they basically just burned all that money spent on development. They could have at least revised the roadmap to be much slower or something, the volume AMD/IBM would have provided probably could have made it viable to run with 1st gen for awhile

Maybe if the need arises they'll just license Samsung's 7nm process or something, that might be a cheaper option if 7nm customer prospects are low.

[Quadriplegic]

9 hours ago, Princess Cadence said:

Now where are all those Intel haters making jokes about 10nm at? two sides of the coin eh?

Boi, did your attempt at joke not go well or what?

[Stefan Payne]

35 minutes ago, cj09beira said:

samsung is too focused on low power mobile chips for it to be useful for a high performance server/consumer cpu 

TSMC is as well as those are the ones that demand the best manufacturing and highest efficiency.

 

And those are OK for modern high performance parts.

You "just" loose a bit of clockrate, so not that big of a deal.

 

Pretty much all those are on totally different fab processing type though, you won't see a MOSFET being made on a modern TSMC process or even at TSMC at all.

No, of course not.
And there is a bunch of shit where lowest cost is the most important thing. If its manufactured in 28nm, 45nm or even 90nm, doesn't matter much as the Chip itself doesn't consume much anyway.

 

The Cost of the silicon is the most important part for thas. There are for example Infineon LLC-Resonant Mode Converters, PWM Controllers, SUpervisiors, +5V Standby Chips with integrated MOSFETs, Diodes, Trannys and so much other shit that is in high demand but lowish cost.

 

33 minutes ago, leadeater said:

But you could see a PCIe switching chip for Broadcom being made at TSMC.

I don't know as most of it is PHY.

Though I haven't seen a DIE Shot from one of those Switches (yet). that might be interesting.

 

But tbh, I doubt that they benefit from 14nm and especially 7nm.

[Froody129]

GloFo no longer GloGo 

[leadeater]

12 minutes ago, Stefan Payne said:

But tbh, I doubt that they benefit from 14nm and especially 7nm.

They don't, TSMC uses a larger node when making those. Though with such good yields and PCIe 4 and 100/400Gb networking the GloFo 12nm/14nm might actually be decent options for those chips.

[Trixanity]

35 minutes ago, Mihle said:

If their 7nm was as close to being complete as it seemed like they where, I don't know why they just finished the development of 7nm and then stopped to shrink. Only thing that I can see if they litterally had no way at all to finance it.

 

Are they sure they litterally stopped 7nm or is it actually just spending less money and not rushing it so much, take the time it needs to save money?

Because the costs to continue were too high. They'd still need to develop 7nm further and work out kinks. They'd have a capacity problem so they'd need to shut down 12/14nm entirely and convert it to 7nm losing a valuable investment and high-yield profitable node. 

[Doobeedoo]

Too bad though. But good AMD 7nm lineup of products is not affected at least. Cause it's a really important node and in general and Zen shift to it will be exciting to see. Their new consumer GPU eventually as well. 

[Curufinwe_wins]

Just a reminder everyone that their 7nm process is closer to intel's 10 than to intel's 7. Turns out 7/10nm is hard.

[Lurick]

Just now, Curufinwe_wins said:

Just a reminder everyone that their 7nm process is closer to intel's 10 than to intel's 7. Turns out 7/10nm is hard.

Hush, we don't need your logic up in here!

 

[S w a t s o n]

12 hours ago, Taf the Ghost said:

R&D covers a lot. You can do all of the design work for a multi-billion dollar project, but the big cost is when you go to start physically building it. The big bill comes when you have to build the fab to move to high volume. Some of the scuttlebutt is that they'd have needed to do it at Fab 8 in NY, which probably meant no more 14nm customers. Though it really just leaves us with a lot more questions than answers.

 

Is IBM moving somewhere else?

Are they still taking delivery on EUV machines?

Are they going to show back up with 5nm in 2022?

Will they continue the research Fab until they can license the technology to trailing Fabs?

Can their research Fab handle IBM's volume?

Are they just going to license future nodes anyway?

 

HVM & research was canceled, but 7nm at GloFo is probably not dead. So, we'll just have to wait and see. These things have weird ways of playing out in tech.

I highly doubt 5nm by 2022 will be possible since they would need EUV and I doubt they are going to take on any more than the 2 EUV machines they already have since they just killed the research budget. If anything I would expect them to try to sell them or something. This choice effectively kills off EUV which in my mind kills off 5nm

[S w a t s o n]

5 minutes ago, Curufinwe_wins said:

Just a reminder everyone that their 7nm process is closer to intel's 10 than to intel's 7. Turns out 7/10nm is hard.

Again, 7nm was not hard for TSMC or GloFo, GloFo was on track and was not suffering delays. Intel's 10nm on the other hand is 3 years late and has to be re-done

[GoldenLag]

1 minute ago, S w a t s o n said:

Again, 7nm was not hard for TSMC or GloFo, GloFo was on track and was not suffering delays. Intel's 10nm on the other hand is 3 years late and has to be re-done

not to mention, redone at a lower density goal. 10nm intel is probably going to be great, but how great and not to mention how late?

[Taf the Ghost]

2 minutes ago, S w a t s o n said:

I highly doubt 5nm by 2022 will be possible since they would need EUV and I doubt they are going to take on any more than the 2 EUV machines they already have since they just killed the research budget. If anything I would expect them to try to sell them or something. This choice effectively kills off EUV which in my mind kills off 5nm

It seems the actual requirement to stay in Leading Edge is to get EUV working. (What a strange statement to say "get EUV working" when you have to pay 100s of millions for the machines.) That might actually be the real "nut" that GloFo didn't want to pay to crack, which would actually explain the decision making a lot better. 

[Taf the Ghost]

2 minutes ago, GoldenLag said:

not to mention, redone at a lower density goal. 10nm intel is probably going to be great, but how great and not to mention how late?

Really freaking late.

 

I'm still not convinced we're going to see Icelake on Desktop. Icelake-SP is Q2 2020, which normally puts Desktop in Q4 2019. The Icelake team would have needed to redo chunks with a new tapeout to be ready for that. Intel will launch server products at 25% yield, but they can't do that for the volume Mainstream brings.

[Taf the Ghost]

Also, I think Cobalt is going to end up being the cause of most of the woes at 7nm/10nm. Maybe not as stable of a metal layer as everyone thought.

[S w a t s o n]

4 minutes ago, Taf the Ghost said:

 Intel will launch server products at 25% yield, but they can't do that for the volume Mainstream brings.

Quote

25% yield

[mr moose]

2 hours ago, Taf the Ghost said:

Also, I think Cobalt is going to end up being the cause of most of the woes at 7nm/10nm. Maybe not as stable of a metal layer as everyone thought.

I've been hearing this too, but most of the people saying that are also saying major news/proof will come out tomorrow. Still waiting which makes me suspicious of they actually know anything or are just making an educated guess. 

 

My money is on density, I reckon they just went too high too fast.  If they can lower their density and drop GPU's from the die they could be looking at a very 7nm competitive 10nm chip.   But I don't know how easy it is to do that nor how many design stages they have to go back if that is even the answer.

 

EDIT: here is an interesting perspective on Cobalt from  the start of this year:

 

https://spectrum.ieee.org/semiconductors/materials/cobalt-could-untangle-chips-wiring-problems

 

For those not interested in the article heere are (IMO) the most intersting parts :

 

Quote

Cobalt has three times as much inherent resistivity as copper but is far less prone to electromigration. For that reason, manufacturers are switching to cobalt for the metal layers that make up short-range connections within and between transistors.

Quote

“The biggest issue is where to insert a new technology,” says Dan Hutcheson, CEO of semiconductor consulting company VLSIresearch. “If you insert early, you incur a lot of costs. Intel is willing to pay a premium for performance, and they have the scale to debug new materials.”

So cobalt is better chemically and from an engineering perspective,  and Intel is essentially the only company positioned to work out the kinks in actual manufacturing.

Quote

GlobalFoundries also described a switch from tungsten to cobalt in its 7-nm process at the December meeting.

It appears that last conclusion has born out to be true to some degree.

[Ryujin2003]

Where are all the NVIDIA and Intel conspiracy theories?

[S w a t s o n]

3 hours ago, mr moose said:

So cobalt is better chemically and from an engineering perspective,  and Intel is essentially the only company positioned to work out the kinks in actual manufacturing.

It appears that last conclusion has born out to be true to some degree.

I've been reading that the combination of quad patterning, cobalt problems, COAG and density are what caused Intel's 10nm to shit the bed, they just pushed it too far and the cobalt was apparently developing micro-fractures along with quad patterning giving massive numbers of defects in general (possibly related to COAG) as even intel wasn't using EUV. TSMC and GloFo were both planning to use cobalt in gen 2, TSMC should still be so I'm not sure what you mean by that. Intel isn't working out any kinks in manufacturing they are just re-doing it.
 

Basically Intel went super hard on density, tried cobalt, tried COAG, and used quad patterning without EUV all at once for a single node.

[leadeater]

55 minutes ago, Ryujin2003 said:

Where are all the NVIDIA and Intel conspiracy theories?