Intel's Process Roadmap to 2025: with 4nm, 3nm, 20A and 18A?!
by Dr. Ian Cutress on July 26, 2021 5:00 PM ESTSidebar on Intel EUV
In all of these announcements, one thing to highlight is Intel mentioning its relationship with ASML, the sole company that manufactures the EUV machines powering production of leading edge semiconductor manufacturing.
ASML Wilton
ASML is a unique company in that it is the only one that can produce these machines, because the technology behind them is often tied up with its partners and research, but also because all the major silicon manufacturers are heavily invested in ASML. For any other company to compete against ASML would require building a separate network of expertise, a decade of innovation and design, and a lot of capital. None of the major silicon vendors want to disturb this balance and go off on their own, lest it shuts them out of the latest manufacturing technology, and no research fund sees competing against the embedded norm as a viable opportunity. This means that anyone wanting EUV specialist technology has to go to ASML.
In 2012, it was reported that Intel, Samsung, and TSMC all invested in ASML. This was, at the time, to jumpstart EUV development along with migrating from 300mm wafers to 450mm wafers. While we haven’t moved to 450mm wafers yet (and there are doubts we will any time in the next decade), EUV is now here. Intel’s 2012 investment of $2.1 billion gave them a 10% stake in ASML, with Intel stating that it would continue investing up to a 25% stack. Those stakes are now below the 5% reporting threshold, but all three of the major foundry customers are still big owners, especially as ASML’s market cap has risen from $24 Billion in 2012 to $268 Billion in 2021 (surpassing Intel).
As major investors but also ASML’s customers, the race has been on for these foundries to acquire enough EUV machines to meet demand. TSMC reported in August 2020 that it has 50% of all EUV machines manufactured at ASML for its leading edge processes. Intel is a little behind, especially as none of Intel’s products in the market yet use any EUV. EUV will only intercept Intel’s portfolio with its new Intel 4 process, where it will be used extensively, mostly on the BEOL. But Intel still has to order machines when they need them, especially as there are reports that ASML currently has backorders of 50 EUV machines. In 2021, ASML is expected to manufacture around 45-50 machines, and 50-60 in 2022. The exact number of machines Intel has right now, or has ordered from ASML, is unknown. It is expected that each one has a ~$150m price tag, and can take 4-6 months to install.
With all that being said, Intel’s discussion point today is that it will be the lead customer for ASML’s next generation EUV technology known as High-NA EUV. NA in this context relates to the ‘numerical aperture’ of the EUV machine, or to put simply, how wide you can make the EUV beam inside the machine before it hits the wafer. The wider the beam before you hit the wafer, the more intense it can be when it hits the wafer, which increases how accurately the lines are printed. Normally in lithography to get better printed lines, we move from single patterning to double patterning (or quad patterning) to get that effect, which decreases yield. The move to High-NA would mean that the ecosystem can stay on single patterning for longer, which some have quoted as allowing the industry to ‘stay aligned with Moore’s Law longer’.
| ASML's EUV Shipments | |||||||||||||||||||||
| 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | |||||||||||||||
| Actual | 2 | 4 | 10 | 3 | 4 | 5 | 6 | 4 | 7 | 7 | 8 | 4 | 7 | 14 | 8 | 7 | 9 | - | - | ||
| Target (Total) | - | - | - | 20 (18) | 30 (26) | 35 (33) | 45-50 | ||||||||||||||
| 2018 and beyond is split per quarter for actual shipped numbers Data taken from ASML's Financial Reports |
|||||||||||||||||||||
Current EUV systems are NA 0.33, while the new systems are NA 0.55. ASML’s latest update suggests that it expects customers to be using High-NA for production in 2025/2026, which means that Intel is likely going to be getting the first machine (ASML NXE:5000 we think) in mid-2024. Exactly how many High-NA machines ASML intends to produce in that time frame is unknown, as if they flood the market, having the first won’t be a big win. However if there is a slow High-NA ramp, it will be up to Intel to capitalize on its advantage.
Facebook
Twitter
RSS
326 Comments
View All Comments
JoeDuarte - Tuesday, August 3, 2021 - link
That's not how knowledge works, or how reality works. It's not possible to know the answers to these sorts of questions by a moment's reflection and opinion generation, especially when the opinions contain no real info. "smart enough to invest" is meaningless without concrete data, like how much "Taiwan" actually "invested" in TSMC, when they did so, and how this investment influenced TSMC's ability to develop new nodes, if at all.(And your second clause is probably false in that I don't think any chipmaker wastes all of their retained earnings on shareholders, and of course "wastes" is a loaded term that would have to be justified with arguments and data.)
You can't really know anything about the world around you using your method here.
mode_13h - Wednesday, August 4, 2021 - link
> You can't really know anything about the world around you using your method here.To put it another way: the term for inferencing without data is "imagination".
mode_13h - Tuesday, July 27, 2021 - link
> It would be very interesting to see some rigorous organizational research on> how some foundries are able to reliably advance with new nodes and shrinks
True. But these must be closely-held secrets. I'm sure they'd closely guard them for decades.
> it looks like Intel is simply unable to advance.
They're advancing, just slowly. They finally got 10 nm performing competitively, by the time they reached 10 nm++ (SuperFin).
And they've delivered products with Foveros and EMIB.
That's why I said above that it's not a question so much of *whether* Intel will deliver, but in what decade!
JoeDuarte - Monday, August 9, 2021 - link
Is Foveros in anything? In desktops? I haven't heard anything.I don't remember what EMIB is, but it only matters if it matters. That is, if it confers benefits to users, like speed.
I don't trust their plus sign stacking anymore, or labels like "Super"-this or that. At this point it's all going to be bottom line metrics for me re: Intel. Is it faster? How much faster? That sort of thing.
Also, it has to actually exist and be ready to buy by normal means, like Dell, Amazon, etc. If there are no desktops with SuperFin+++++ Tiger Claw, Cougar Lake, Bald Eagle Lake, Rocket Man, etc. then it's moot. I think whatever Lake is supposed to have SuperFin++++ is in desktops yet.
JoeDuarte - Monday, August 9, 2021 - link
I *don't* think...mode_13h - Tuesday, August 10, 2021 - link
> Is Foveros in anything? In desktops? I haven't heard anything.Lakefield, but I'm guessing that's why you added "in desktops". Chip development times seem to be around 4 years or so. That means you wouldn't expect an overnight transition to something like Foveros. First, it has to be proven and refined. Then, you might see chips still early in their design phase start to utilize it, but that should result in a gap between the first example and more widespread adoption. So, the fact that we haven't yet seen widespread adoption doesn't mean anything, in itself.
> I don't remember what EMIB is
It's easy to look up. It's Intel's technology for building multi-chip modules. I think Sapphire Rapids is slated to use it. Otherwise, I think Meteor Lake might be first.
> I don't trust their plus sign stacking anymore
If we were talking about 14 nm, then I'd agree. However, they've been a bit more transparent about what distinguishes 10 nm+, 10 nm++, and 10 nm+++. For details, see the article (which refers to ++ as SF and +++ as ESF).
> Is it faster? How much faster? That sort of thing.
Their foundry business should offer new transparency into their manufacturing node. Before, when Intel was practically their fabs' only customer, they didn't need to reveal detailed performance characteristics of their process iterations. However, when you're trying to attract customers to use your fabs, the standard for transparency is a lot higher. It also means that Intel needs to be more conservative about their roadmaps, since actual manufacturing contracts are going to hang by on-time and on-target delivery.
However, I agree that the end result (cost, area, power, frequency, and wafers per month) are what really matters. I'm not really bothered about dubious density metrics, and that's why I think they should just switch to a sequential version numbering scheme.
> Also, it has to actually exist and be ready to buy by normal means
Manufacturing process roadmaps are not of much interest to end-consumers, precisely because the timelines for getting end products on shelves is so far out.
Also, I'd say consumer shouldn't even think much about manufacturing tech. Good product reviews will show the relevant performance and power characteristics, and that's what most people really care about.
> I think whatever Lake is supposed to have SuperFin++++ is in desktops yet.
Ice Lake uses "regular" 10 nm, which is really 10 nm+ (but Intel doesn't call it that, because they'd rather pretend that Cannon Lake's 10 nm didn't happen). Ice Lake launched in laptops toward the end of 2019 and in servers a few months ago.
Tiger Lake uses 10 nm SF. The quad-core launched in laptops at the end of 2020 and 8-core launched a few months ago. The only "desktops" that have either are small-form-factor PCs, such as NUCs. Intel skipped selling it as a normal desktop CPU, probably due to manufacturing constraints (i.e. insufficient volume), but *maybe* also having to do with performance or power scaling.
Alder Lake should launch (for desktops!) around the end of this year and will be the first CPU on 10 nm ESF (now called "Intel 7"). It'll be the first "10 m" product they've launched for desktops.
JoeDuarte - Friday, August 13, 2021 - link
@mode_13h, Foveros isn't new though. It's old. I first read about it in 2019, or earlier. My comment rested on my unstated framing view that Foveros is old, and it's in my bucket of "Things Intel has hyped but never delivered.", like the optical Thunderbolt (Lightpeak?), Hybrid Memory Cube RAM, especially HMC2 and later, silicon photonics (as something that exists in even high-end desktops and servers), 10nm, 7nm, Phi/Knights (in computers we can buy), Panel Self-Refresh, and that optical connector embedded in functional USB ports (which would be awesome; I think Sony did something similar on a VAIO).I don't know what Lakefield is, but Googling it reveals that Intel has cancelled it, so Foveros isn't real yet, and might never be, not as something in a computer we can buy. I doubt EMIB will ever matter.
I didn't mention desktop as some sort of cynical ploy. I mentioned it because it's what I care about most, and it's also where Intel has failed to advance the most, re: 10nm etc. It's also a pet peeve of mine that desktop computing is often dismissed as unimportant because of what I call the Growth Fallacy, or the Relative Growth Fallacy. It's the fallacy that a market or industry doesn't matter, or doesn't matter much, is it's not growing at any arbitrary present. Or if some other category is growing faster, or growing a lot. Desktop is enormously important, a huge market, and where work happens. So I care most about desktop. The 7 watt stuff doesn't solve any problems for me.
mode_13h - Saturday, August 14, 2021 - link
> Foveros isn't new though. It's old.In the relative scheme of things, it's pretty new. It's also newer than the rest of the things you mentioned. And, for some of those, the issue is more that their time hasn't yet come. For instance, the OCP article (Rebecca Weekly interview) mentioned silicon photonics as an area of growing focus.
> Lakefield ... Googling it reveals that Intel has cancelled it
Not cancelled, but it reached end-of-life. There's an important difference, as one involves actually bringing a functioning product to market and supporting it.
I think Foveros is slated to appear possibly in Meteor Lake, which will be Intel's first chiplet-based desktop CPU. It's planned to come after Raptor Lake, which will follow Alder Lake. So, probably near the end of 2023. I'm not sure if they have any other products planned to use it, before then, but I wouldn't be surprised.
> desktop computing is often dismissed as unimportant
I haven't heard that, recently. For several years, the PC sector was on a slope of inexorable decline, but that changed around the time Zen launched.
> Desktop is enormously important,
It's indisputably high-profile. It was the beachfront for AMD's comeback in the CPU space, and it routinely generates lots of publicity for Nvidia and AMD's GPU products.
BTW, I'm not trying to be an Intel apologist, here. I'm just trying to present the facts as I understand them. And my understanding is that technologies like EMIB and Foveros are still in their commercial infancy. Certainty will only come with time.
AdrianBc - Tuesday, July 27, 2021 - link
To be fair, after a few years of complete stagnation, during the last 4 years Intel has improved steadily their 10 nm process.For example the maximum clock frequency has increased from 3.2 GHz in 2018 to 4 GHz in 2019, then 5 GHz in 2020 and finally 5.3 GHz (according to rumors) in 2021.
The renamed "Intel 7" manufacturing process will be the first Intel process better than the 7 nm TSMC process, if the Intel claims are true.
The problem for Intel is that this slow recovery might be too late, because meanwhile TSMC has been using already for a long time the superior 5 nm process, in high volume production (for Apple).
mode_13h - Tuesday, July 27, 2021 - link
> the maximum clock frequency has increased from 3.2 GHz in 2018 to 4 GHz in 2019,> then 5 GHz in 2020 and finally 5.3 GHz (according to rumors) in 2021.
Wow, almost like their 14 nm process!