[SemiAccurate] Intel kills off the 10nm process

[Arachnotronic]

I was thinking more along the lines that Intel created a special 14 nm version that's optimized for low power and high density and ported Icelake Client and perhaps the modem to it. They are now calling that 10 nm. And then they are essentially cancelling everything else.

Why?

 

[.vodka]

Some claim its due to cobalt, some claim its the difficulty due to going above quad patterning.

My verdict?: It's all speculation until its actual.



Why speculate when you have hard data?

https://fuse.wikichip.org/news/1371...ghts-reports-on-the-i3-8121u-finds-ruthenium/

Tech Insights already opened the chip and analyzed it. And yes, its using 10nm technology.

"As far as cell size is concerned, there are no surprises. TechInsights silicon measurements matches Intel’s own IEDM paper."
"Techinsights has confirmed Intel is using contact-on-active gate (COAG)."
"For the three lowest layers (poly, metal 0, and metal 1), Intel introduced Cobalt and Tungsten."

You misunderstood me. I was pointing out how simply taking 14nm and renaming it 10nm and shipping it out next year would backfire having people like Techinsights doing analysis right down at the chip level, in response to who I quoted in that post. Read that quote better.

I wasn't talking about Intel's current failure of no iGPU, 10nm i3 8121u Cannonlake chips out there or Intel's current state of 10nm at the moment.

 

[Veradun]

People, people.

If Intel's struggling immensely on the 10nm process, do you think they'll be better off skipping to a 7nm process? Technology improvements are iterative.

Agree. They'll just "relax" their 10nm node and keep calling it 10nm.

 

[IntelUser2000]

I wasn't talking about Intel's current failure of no iGPU, 10nm i3 8121u Cannonlake chips out there or Intel's current state of 10nm at the moment.

I gotcha. Sorry about that.

 

[AtenRa]

Most probably they will skip directly to 10+ or even 10++ for the 2020 products.
So actually Intel 10nm is dead, long live the 10nm+ and 10nm++.

 

[moinmoin]

What's really weird to me is how up to as close as a year ago, well maybe one and a half, they were still touting their 3yr process advantage in a way they still believed in it themselves . It really seems they got both blindsided about the competition and overconfident in their own process engineering.

It really is a perfect storm. Intel had the advantage of better architecture and a foundry operating in virtual isolation. Intel saved bigger architecture changes for the coming node which again is being perpetually delayed. Essentially within one and a half year AMD arrives with a competitive architecture which with GloFo's node licensed from Samsung turns out to be a veritable rival, and Intel's foundry soon will finally have to compete directly against leading edge pure play foundry giants as AMD goes to TSMC for the next node.

 

[Maxima1]

Most probably they will skip directly to 10+ or even 10++ for the 2020 products.
So actually Intel 10nm is dead, long live the 10nm+ and 10nm++.

That's not skipping. Cannon Lake just was never really here. Cannon Lake was the 10nm process. If presumably Ice Lake follows, that will have a designated 10nm+ (unless skipped for ++) process.

 

[DrMrLordX]

It's a given that Intel's 10nm process as seen on the 8121U is dead (except maybe for some non-Core products). The question is, is 10nm+ dead too?

 

[NTMBK]

Interesting bit of follow up by Ian Cutress:

"We’re making good progress on 10nm – consistent with what we shared at earnings. As soon as we can give you more details, we will." Official response to if it's the same 10nm as said at TMG Day last year / same 10nm as CNL.

https://twitter.com/IanCutress/status/1054420072960270338

Notice how Intel deliberately avoids saying that it's the same 10nm, despite Ian asking directly. It looks a lot like this isn't the same 10nm, going by that response.

 

[jpiniero]

Why?

Because they can't make the real 10 nm work. Which fits the need to expand the 14 nm capacity, especially with Cooper Lake being such a hog.

Perhaps they could use the 10 nm production lines and convert them to 7 to allow them to release 7 nm products earlier than needing the Arizona fab. Of course they would need to bring in EUV equipment.

 

[scannall]

Interesting bit of follow up by Ian Cutress:


https://twitter.com/IanCutress/status/1054420072960270338

Notice how Intel deliberately avoids saying that it's the same 10nm, despite Ian asking directly. It looks a lot like this isn't the same 10nm, going by that response.

Is it a 10nm process? Probably. Certainly not the same one they started after though. We know they are relaxing it, a lot. And as a likely guess going back to copper. Saving the cobalt for later. Much like TSMC has done. So, it seems both Semi-accurate and Intel are right. Intel is working to get 10nm out the door. It just isn't the same one they started after. After all these years of delays, scrapping and starting over was likely the better choice.

 

[ajc9988]

Semi-accurate has never been fully accurate hence the name. Burying it behind such a massive paywall enables them to stir up the hype without the need to deliver the meat. Those silly 5 investors won't share that 'knowledge' or make a dent in the stock market.

FWIW people speculating about skipping 10nm for 7nm clearly have no clue about the consequences and the ambitious specifications of 7nm. It's more dense then TSMC's 3nm process based on what is known right now about both processes.

If you can't jump 5ft you won't solve that by jumping 10ft.

What is likely at play here is Intel relaxing the m2 pitch from 36 to 40 or the m1/2 from 40/36 to both 44 to avoid the quad patterning of those two metal pitches. Depending on what their double patterning can handle it will either be 40 or 44nm. TSMC can handle 40 on their 7nm process but it's really unclear why Intel went with 36 in the first place.

Charlie sells that as 'dropping 10nm' which is simply hyperbole from an attention whore.

Agreed. People do not understand that processes can take 3-5 years in development before reaching the light of day, and even then they may not make it to see the light. Look at TSMCs scrapping of different nodes back around the 32nm/28nm timeline. TSMC has created a new process to have partners actively involved in the development of a process node in hopes to shrink the time to market with a new node or process refinement to 18 months. Very ambitious in my opinion. But no one is above the laws of physics, and trying to find ways to make a chip takes time, especially with the issues ahead. I will mention some of those issues in responses to others below.

There was a tweet by Intel which they stated that they HAVE NOT killed off their 10nm process. BUT what someone else said that it doesn't necessarily mean they are going to 7nm right away. Perhaps they could just be jumping to 10nm PLUS [10nm+] instead?

In light of development to implementation and Intel's current timeline suggesting introduction at 2021 or 2022, I find it dubious to pull the timeline in to 2020 even for risk production. Now, everyone is forgetting Intel said Ice Lake is coming next holidays, not cannon lake which is skylake at 10nm. They are skipping to introduce the new uarch and the 10nm+ process if that is to be considered an unchanged roadmap. What 10nm+ is, we do not know other than Intel mentioning in their own slides that the transistor performance (not separating out frequency or IPC) would be slightly less than 14nm++, while stating 10nm would be 25% more performance than 14nm (Intel press day March 29, 2017, IIRC). So, they will likely have a 10nm process of some sort, or some process more dense than 14nm, to present, but we really do not have a clue what it will look like or the performance to expect from it.

Now, on the points made about the article a couple weeks ago that yields are improving and production may have moved up by 6 weeks (half a quarter), Intel now reaffirming holiday 2019 and this report suggests that what was promising may not have panned out for some reason or another. Or they may be adopting another team's idea on how to get the density, but with a different set of tweaks to the process, which was less preferred than the last set of tweaks, but is better than scrapping the node all together. Hard to tell without what is behind the paywall.

Some claim its due to cobalt, some claim its the difficulty due to going above quad patterning.

My verdict?: It's all speculation until its actual.



Why speculate when you have hard data?

https://fuse.wikichip.org/news/1371...ghts-reports-on-the-i3-8121u-finds-ruthenium/

Tech Insights already opened the chip and analyzed it. And yes, its using 10nm technology.

"As far as cell size is concerned, there are no surprises. TechInsights silicon measurements matches Intel’s own IEDM paper."
"Techinsights has confirmed Intel is using contact-on-active gate (COAG)."
"For the three lowest layers (poly, metal 0, and metal 1), Intel introduced Cobalt and Tungsten."

They might have changed the cobalt gating too so maybe that's a definite cancellation of their original process in his eyes.

So, as you mentioned, it may be cobalt. As some have espoused, Intel is working on the problems of using Cobalt and transitioning to expensive other materials intended to help deal with the problems of quantum tunneling and electron "bleeding." While they work on this, other fabs, such as TSMC and Samsung, have focused on the introduction of EUV lithography, starting with the lowest layers for Samsung and eventually moving to the entire stack next year and TSMC hoping to have EUV in risk production on 7nm+ by 1H next year (likely Q2). Intel may not introduce EUV until 7nm, expected in 2021 back in August. (https://www.eetimes.com/document.asp?doc_id=1333657).

We will need to do both as we quickly approach the issues related to node shrinks moving forward. Which is the best, focusing on the materials science side or on the lithography to reduce patterning defects and costs associated with quad patterning? Time will tell. But, Intel was burned on the promises of EUV being ready since 2015, a large reason that we are now sitting here, years later, with no 10nm chips from Intel.

Those issues with EUV still have not been resolved fully, such as mask problems with deformation, pellicles, etc. In fact, Samsung is moving use forward without some of that equipment, and TSMC is trying to do the same. They are pushing forward to make use of the EUV equipment while hounding the companies providing the equipment. But, they came to this point literally years after Intel was banking on its use in production, so Intel moving to the materials side, something not being done in manufacturing at the other fabs this early, instead them waiting for 5nm or 3nm to introduce cobalt, ruthenium, or other similar materials makes sense.

We also must remember that Intel's 7nm and the 5nm or 3nm processes at other fabs spells the end of finFETs as currently envisioned, instead potentially moving to nanowire and nanosheet Gate All Around transistors, which introduce their own difficulties in implementation.

Also, Intel may be choosing to less aggressively target pitch, as mentioned by Spartak (I agree with this), could be using cobalt less for contact or linings, etc. Time will tell.

 

[jpiniero]

Now, on the points made about the article a couple weeks ago that yields are improving and production may have moved up by 6 weeks (half a quarter), Intel now reaffirming holiday 2019 and this report suggests that what was promising may not have panned out for some reason or another.

I think it's more like the Fake 10 nm is progressing nicely. The Real 10 nm is what got cancelled, and Charlie is also implying that even the loosened Real 10 nm is cancelled too.

Charlie had also said that Icelake Server was cancelled and replaced by yet another 14 nm product without saying what arch (publicly at least). That's part of the reason why I think the Fake 10 nm is 14 nm tooling based.

What the density is, we won't really know until Icelake Client ships I reckon.

 

[Maxima1]

Charlie is also implying that even the loosened Real 10 nm is cancelled too.
.

He definitely is. He was already saying the relaxed 10nm was "12nm" (really more like 10.5nm).

 

[Dayman1225]

Charlie had also said that Icelake Server was cancelled and replaced by yet another 14 nm product without saying what arch

Where?

 

[jpiniero]

Where?

https://semiaccurate.com/2018/09/11/intel-puts-out-another-14nm-2020-server-platform/

Most of it is of course behind the paywall but you can see why Icelake Server being cancelled is implied. Plus there was that roadmap that didn't have Icelake Server on it, only "Future" (in 2021)

 

[ajc9988]

I think it's more like the Fake 10 nm is progressing nicely. The Real 10 nm is what got cancelled, and Charlie is also implying that even the loosened Real 10 nm is cancelled too.

Charlie had also said that Icelake Server was cancelled and replaced by yet another 14 nm product without saying what arch (publicly at least). That's part of the reason why I think the Fake 10 nm is 14 nm tooling based.

What the density is, we won't really know until Icelake Client ships I reckon.

I actually mentioned the idea of a different loose 10nm being used when I said "they may be adopting another team's idea on how to get the density, but with a different set of tweaks to the process." What I mean is Charlie's reporting on what was gutted at 10nm initially resulted in his later article discussing their mitigations being better than thought, so that the damage from the original gutting was not that bad. The questions then come in whether the mitigations had unforeseen consequences requiring a further redesign on process. So we are in agreement there.

As to his reports on Icelake Server being cancelled, I find this dubious. We already know Cascade-X/SP should be dropping in December on 14nm++. We know they have a planned 14nm++ refresh for sometime next year called Cooper. There is a chance Cooper will be pushed out to get more time for whatever is produced in 2020, but I think it is a bit early to say there will definitely be one more 14nm chip after those.

It is like press mentioning that the new 9th series HEDT chips, excluding the overclockable Xeon, will be on Skylake-X but 14nm++. Intel has renamed EVERY 14nm chip with the process refinement, which makes me think that the new HEDT 9980XE and company are just soldered versions of Skylake-X chips on 14nm+. My theory also involves their capacity being constrained, and that the 14nm++ process requires some changes to production lines, so they used the 14nm++ for the Coffee refresh on mainstream, which already seems to have severe supply constraints limiting availability and driving the price to between $520 and $580, and to ramp up supplies for the upcoming Cascade Server chips, which have higher margins, thereby getting the most out of the capacity they have for manufacturing. If I'm correct, you only get minimal increase in clock speeds, same IPC, and will see the CPUs clocked closer to their limits than on the 7th generation chips for Skylake-X. But that is just my theory.

 

[Dayman1225]

https://semiaccurate.com/2018/09/11/intel-puts-out-another-14nm-2020-server-platform/

Most of it is of course behind the paywall.

I've seen whats behind it and it certainly isn't a new 14nm Server CPU arch. Its a 4s Server platform named Cedar Isle for Cooper Lake and potentially Icelake

 

[dark zero]

Most probably they will skip directly to 10+ or even 10++ for the 2020 products.
So actually Intel 10nm is dead, long live the 10nm+ and 10nm++.

Even more, I see the 10 nm being canned for desktop solutions and goes directly to 10nm+ for next year.

 

[moonbogg]

Where is my 8/16 chip at 5ghz + 25% ipc for $350? That's all I want to know.

 

[Vattila]

Even more, I see the 10 nm being canned for desktop solutions and goes directly to 10nm+ for next year.

Ok. So, on one end of the spectrum, we have a camp here that actually sees good news: Intel solves the problems with a 10nm+ process next year and fulfils its promises for products on shelves by holiday season 2019 with 10nm+ Ice Lake-U/Y. If you are in this camp, do you expect 10nm+ Ice Lake-S for desktop in 2019 also? Early 2020? Mid 2020?

Then you have the more sombre camp at the other end of the spectrum (where I'm at), which thinks this is an ominous sign that Intel has given up on the 10nm roadmap altogether, except for their limited commitments to "12nm" Ice Lake-U/Y by the end of 2019, with all their 10nm fab conversion plans scrapped and product roadmaps redone for 7nm (their own or third-party) two or three years out into the future. Until then they are expanding 14nm++(+) capacity and implementing plan-B products to keep the business going and compete as best as they can.

 

[krumme]

Its going to be really tough on the server end...
Its a huge possibility for amd if their zen2 delivers and come on time. A bit of luck. Tsmc on track Intel side tracked.

 

[Maxima1]

Ok. So, on one end of the spectrum, we have one camp here that actually sees good news: Intel skips ahead to 10nm+ next year, fulfils its promises for products on the shelves by holiday season 2019, skipping Cannon Lake and accelerating the move to Ice Lake. If you are in this camp, do you expect 10nm+ Ice Lake-S for desktop in 2019 also?

Why would anyone think they're putting out Cannon Lake during 2019, which wouldn't even have an IPC uplift to mitigate the clock advantages of their mature 14nm process? Cannon Lake was mainly just a shrink to 10nm. That process isn't workable, and Intel has stated many months ago Ice Lake is slated to be on 10nm+ This isn't good news but what has been expected.

Then you have the more sombre camp at the other end of the spectrum (where I'm at), which thinks this is an ominous sign that Intel has given up on the 10nm roadmap altogether, except for their limited commitments to "12nm" Ice Lake-U/Y by the end of 2019, with all their 10nm fab conversion plans scrapped and product roadmaps redone for 7nm (their own or third-party) two or three years out into the future. Until then they are expanding 14nm++(+) capacity and implementing plan-B products to keep the business going and compete as best as they can.

I see Charlie has tweeted he expects Intel will try to bring closer 7nm, so he obviously thinks 10nm is dead. I don't see how they could get this any earlier than 2021, though. That's a long time.

 

[Vattila]

Why would anyone think they're putting out Cannon Lake during 2019

I did, silly me. But with the recent leaks it is clear that Ice Lake-U is in the works. I'll correct my post.

 

[Vattila]

Intel has stated many months ago Ice Lake is slated to be on 10nm+

I've seen this mentioned before, but do you have a reference for that? Just saw this recent tweet by David Schor (Wikichip):

"We can speculate [that Ice Lake is] on 10nm+ but AFAIK that was never confirmed officially. Nonetheless, they should both be using the same feature sizes and more or less identical design rules."

https://twitter.com/david_schor/status/1054794367997747201

Edit: Never mind. The reference actually follows in the Twitter conversation:

https://www.anandtech.com/show/11722/intel-reveals-ice-lake-core-architecture-10nm-plus

PS. Note that this is an article from August 2017. A lot can have changed since then.