Intel on Thursday announced that it would delay mass production of its 10 nm processors from 2018 to 2019 due to yield issues. The company has claimed to be shipping some of its 10 nm chips in small volumes right now, but due to cost reasons the firm does not intend to initiate their high-volume manufacturing (HVM) at this time. Intel executives also stated that they are confident of their product roadmap and intend to launch Whiskey Lake and Cascade Lake products later this year.

Multipatterning Issues

Brian Krzanich, CEO of Intel, stated during a conference call with financial analysts:

“We are shipping [10-nm chips] in low volume and yields are improving, but the rate of improvement is slower than we anticipated. As a result, volume production is moving from the second half of 2018 into 2019. We understand the yield issues and have defined improvements for them, but they will take time to implement and qualify.”

Intel blames a very high transistor density and consequent heavy use of multipatterning for low yields. Brian Krzanich has said that in certain cases the company needs to use quad (4x), penta (5x), or hexa (6x) patterning for select features as they need to expose the wafer up to six times to “draw” one feature. This not only lengthens Intel’s manufacturing cycle (which by definition rises costs) and the number of masks it uses, but also has an effect on yields.

Intel’s 10 nm fabrication technology relies solely on deep ultraviolet (DUV) lithography with lasers operating on a 193 nm wavelength at this time. The company’s 7 nm manufacturing process will use extreme ultraviolet (EUV) lithography with laser wavelength of 13.5 nm for select layers, eliminating use of extreme multipatterning for certain metal layers. As it appears, right now Intel executives do not consider EUV technology ready for prime time in 2019, so the company’s engineers have to polish off the last DUV-only process (again) rather than jump straight to 7 nm.

10 nm Ramp in 2019

Intel does not elaborate whether it intends to ship (in volume) its 10 nm CPUs in the first half or the second half of 2019, but only says that the company’s engineers know the source of the yield problems and are working hard to fix them. As a result, it is pretty safe to assume that the actual ramp of Intel’s 10 nm production will begin towards the second half of next year.

“We are going to start that ramp as soon as we think the yields are in line, so I said 2019,” Mr. Krzanich noted. “We did not say first or second half, but we will do it as quickly as we can, based on the yield.”

In a bid to stay competitive before its 10-nm CPUs ship in the H2 2019 – H1 2020 (production ramp takes time, bigger processors will launch later than smaller parts), Intel plans to release another generation of products made using its 14 nm process tech. This generation of chips includes Whiskey Lake products for client PCs and Cascade Lake for the datacenter, and both are scheduled for release later this year.

Questions Remain: 10 nm Slip from 2016

Intel’s 10 nm manufacturing process has a long history of delays and without any doubts this transition has been the hardest in the company’s history. Let's do a quick recap.

  Intel
First Production
1999 180 nm
2001 130 nm
2003 90 nm
2005 65 nm
2007 45 nm
2009 32 nm
2011 22 nm
2014 14 nm
2016 10 nm
2017 10 nm
2018 10 nm?
2019 10 nm!

Intel originally planned to commence shipments of its first processors made using their 10 nm fabrication technology in the second half of 2016. The first rumors about Intel’s problems with the tech started to spread in early 2015 as the company delayed the installation of equipment needed for its 10 nm manufacturing process. Then, in July 2015, the chip giant confirmed intentions to postpone 10 nm HVM from H2 2016 to H2 2017 due to difficulties incurred by multipatterning. Instead, the company promised to release its Kaby Lake products with enhancements and made using a refined 14 nm process (known as 14+ later).

Over the course of 2016 and 2017 we learned that Intel was prepping Cannon Lake (mobile, entry-level desktops), Ice Lake (higher-end client PCs, servers), and Tiger Lake chips on their 10 nm node. We also heard about various problems that Intel faced with its 10 nm technology, but the company refuses to comment on them. The chipmaker did demonstrate a system running a mobile Cannon Lake SoC at CES 2017, with a promise to release this processors late that year. Somewhere along the line, both Ice Lake and Tiger Lake slipped to 2018, which was partly confirmed by the launch of the Coffee Lake CPU made using a revamped 14 nm (14++) in late 2017.

Intel officially introduced the 10 nm fabrication process at IEDM 2017 and said it was on track to start shipments of CNL CPUs in early 2018. In January this year Intel confirmed that they had started to ship Cannon Lake processors in small volumes, but never elaborated. We have since learned that these were CNL-U parts in uninspiring 2+2 and CNL 2+0 configurations, however so far Intel has refused to state who the customer is or where anyone can buy them, despite repeated requests for this information.

Being a very large company, Intel has a multifaceted strategy that spans across product lines and generations. Right now, Intel is battling with yield issues that plague its Cannon Lake product family and the first-gen 10 nm manufacturing process. There are other 10 nm products in the pipeline that are to be made using a refined fabrication technology (such as 10+, 10++). It is pretty obvious that Intel will learn how to improve its 10 nm yields with the CNL lineup, but what remains to be seen is how significantly the delays of this product family affect launch schedules of its successors. Despite Intel's statements, there is outside discussion that Intel could decide to switch right to 7 nm, bypassing 10 nm altogether.

Related Reading

Sources: Intel, SeekingAlpha

POST A COMMENT

155 Comments

View All Comments

  • FunBunny2 - Sunday, April 29, 2018 - link

    "I think it depends on whatever eventually works out, and what future discoveries are made that lead to that technology winning out. For some technologies they could potentially grow the circuits instead of draw them, in which case photolithography would no longer be the key technology. "

    manufacturing tech isn't the issue. the engineers will always figure out some ways to make the widgets. the issue is at what feature size (and if that is one size fits all feature types, so let's just say minimum) does that feature go Heisenberg? no. quantum computing won't support deterministic problem spaces. until someone shows confirmatory math.
    Reply
  • Yojimbo - Monday, April 30, 2018 - link

    Manufacturing technology is a big issue. Thousands and thousands of man-years has gone into understanding and refining photolithographic idioms. Billions of dollars have been spent on the photolithographic infrastructure. Whenever someone comes up with a new technology they are thinking about bringing to mass market a big question he must ask himself is "Can this be made with standard CMOS manufacturing techniques?" Of course if a technology shows enough promise, or if there is no suitable alternative, such a barrier can be overcome. But that's a far cry from there not being an issue there at all. Reply
  • FunBunny2 - Monday, April 30, 2018 - link

    "Manufacturing technology is a big issue."

    at this very moment, that's true. even using known, and knowable, tech engineers could produce feature size at 2nm or 3nm, I suppose. but that's meaningless if such features go Heisenberg. the point is, of course, that Heisenberg is closer than it appears in the mirror.
    Reply
  • peevee - Monday, April 30, 2018 - link

    You look like a person who knows something.
    Why the huge jump from 193nm to 13nm? What happened to intermediate wavelengths? Say, 100nm would suffer much (~60 times) less from shot noise compared to 13nm, but would eliminate good deal of multipatterning already.
    Reply
  • Yojimbo - Tuesday, May 1, 2018 - link

    I don't know, sorry. I'd guess that the issue is one of materials. They need a light source, optics, masks, and resists for any wavelength. Perhaps they couldn't find any set of suitable candidates until they got down to EUV wavelengths. They were looking at X-rays, too, but in the end it was EUV that showed the most promise. I know that they were developing a 157 nm system. It was beset with problems, but I think they might have solved them. However, the industry settled on immersion lithography at 193 nm instead. Reply
  • Yojimbo - Tuesday, May 1, 2018 - link

    Oh I found something. It's a a journal article by someone from Bell Labs of Lucent Technologies published in "Material Today" in 1999. It says:

    "Extension of conventional refractive optical lithography to wavelengths below the F2 excimer (157nm) is problematic due to absorption in the refractive elements. It is also very difficult to design and construct reflective optical systems with sufficiently large numerical apertures to allow printing at or below the wavelength of the illumination source. The idea of EUV lithography is to use small numerical aperture reflective optical systems at wavelengths much shorter than the circuit dimensions. The reflective elements for EUV use multilayer Bragg reflectors such as MoSi pairs to produce reflectivities up to nearly 70% at 14 nm. The mask in an EUV system is reflective and also uses the same type of multilayers. A plasma or synchrotron based source is used to illuminate the mask which is imaged by a system of mirrors onto the resist-coated wafer with a reduction factor of four. The optical systems require mirrors with unprecedented tolerances with respect to figure and finish.That is, the shape of the mirror must be correct in addition to the surface being smooth. The specifications are in the angstrom and in some cases, sub-angstrom range posing serious challenges for mirror fabrication, coating, and mounting."

    So, according to that, they have problems with refractive optics at wavelengths shorter than 157 nm and in order to use reflective optics they need much shorter wavelengths.
    Reply
  • Spunjji - Friday, April 27, 2018 - link

    This x1000 Reply
  • Santoval - Friday, April 27, 2018 - link

    Jesus, so that means that Ice Lake is moved back to a Q1 2020 release at the earliest, probably Q2 for truly high volume. And Intel will have spent at least 5, perhaps even 6 whole years (2014 to 2019) releasing various 14nm, 14nm+, 14nm++, 14nm+++ etc CPUs.. If the 7nm yields of TSMC, GloFo (and Samsung) turn out to be much better Intel is going to be crushed next year by AMD, and ARM based CPUs will close even more the performance gap with Intel CPUs. Reply
  • Kevin G - Saturday, April 28, 2018 - link

    Ice Lake is still looking like a 2019 launch, though probably later in the year. The design has taped out so the masks are done, just waiting on yields to improve. This was before Meltdown/Spectre. So even if
    Intel were to revise them, they'll still be waiting on the process to mature.

    What this does mean is that Cannon Lake is vapoware.
    Reply
  • jrs77 - Friday, April 27, 2018 - link

    I guess it was only a question of time before we hit the wall with the process. Sure there are products made in 7nm allready, but they're nowhere near as complex. So I don't think that intel will bypass the 10nm process and go straight to 7nm, as we've hit the current limit imho, as to how small we can go with the current materials for highend CPU manufacturing. Reply

Log in

Don't have an account? Sign up now