Comments Locked

208 Comments

Back to Article

  • UltraWide - Thursday, July 23, 2020 - link

    Intel should concentrate on 10nm for the time being, why talk so much about 7nm when it's in such dire straits? The reason is most likely to dampen expectations for clients and investors. Best guess it that 7nm is actually much further behind than 6 months. There seems to be no end to intel's fab problems...
  • Great_Scott - Thursday, July 23, 2020 - link

    While the idea of defaulting to 3rd-party fabs is interesting, it's inherently a non-starter.

    High performance fabbing for CPUs means "TSMC". And I don't understand how Intel is going to get wafer starts away from AMD, Nvidia, and Apple. The capacity simply isn't there.
  • 0ldman79 - Thursday, July 23, 2020 - link

    $$ makes stuff happen.

    If Intel commits to a certain amount of TSMC capacity then it would be worth it for TSMC to expand to meet that demand, especially if they can manage to pull it off and have Intel pay for the brunt of it.

    It's not like Intel can just *not* make the chips. They have to or they're in serious trouble.
  • dotjaz - Friday, July 24, 2020 - link

    Intel has it's own fab, so they are not going to be a long term customer, you don't build capacity for 2-3 years worth of orders (and it takes 2-3 years to even build the fab).
    In the mean time AMD is very much committed to foundries. CPU demand is not gonna vanish. AMD will pick up what Intel can't supply, so TSMC will get the order anyway. Why help a less committed customer/competitor getting back on their feet instead of better supporting committed current customer?
  • vladx - Friday, July 24, 2020 - link

    Because TSMC doesn't take sides, they care about money first and foremost. If Intel wants to fab a good chunk of their chips at TSMC, TSMC would gladly take their money.
  • vladx - Friday, July 24, 2020 - link

    And probably they already have a deal going on as TSMC claimed Huawei ban won't affect their income even with Huawei currently representing 20-30% of their total 7nm fab capacity,
  • bernstein - Friday, July 24, 2020 - link

    that's a good point. might be that intel already picked up that (or a part of it) capacity. but it may also well be that it's destined to go to apple, amd & nvidia as they all are in need of more capacity.
  • Spunjji - Friday, July 24, 2020 - link

    @vladx - Source needed on that "20-30%" claim. I haven't been able to verify it.

    As for "probably already have a deal", lol, you're sweet. These things do *not* happen overnight.
  • vladx - Friday, July 24, 2020 - link

    You can find it in the TSMC halting shipments to Huawei article:

    https://www.anandtech.com/show/15915/tsmc-confirms...

    Just read it.
  • Spunjji - Wednesday, July 29, 2020 - link

    @vladx lol. Per the linked article: "Huawei represented the manufacturer’s biggest customer with a 23% revenue share in 2019." So - that's 23% of *revenue*, which is NOT "20-30% of their total 7nm fab capacity". It is not, in fact, *anything like that*.

    I knew that claim was phoney. Thanks for linking to the source proving as much! 😆
  • name99 - Friday, July 24, 2020 - link

    OR that non-lost income represents substantially increased AMD wafers?
    That would be my guess.

    There are other smaller (but still significant) demands out there that will also pick up that slack, like Apple Silicon. Probably also parts of Ponte Vecchio (and MobilEye, and Habana) but not yet Intel CPUs -- that's what people care about in this context, not the new/3rd party Intel stuff.
  • dotjaz - Sunday, July 26, 2020 - link

    Exactly, you just proved my point. TSMC will sell under-utilized capacity to Intel, not build capacity for them only to be abandoned once Intel's Fab is back on track.
  • Nicon0s - Monday, July 27, 2020 - link

    The demand for silicone chips continues to increase so I don't think TSMC would have a problem if after let's say 3 years Intel would just stop using their fabs. Now it's a very good time to build fabs.
  • Spunjji - Friday, July 24, 2020 - link

    Dude, it's not about "taking sides" - you can't just throw money at TSMC and expect them to "make more big fab go faster".

    I love it when people post confidently about what large corporations will do based only on naive conceptions of how these things work in concert with a buttload of motivated reasoning.
  • philehidiot - Friday, July 24, 2020 - link

    Just a thought, TSMC might well look at Intel as a short term customer in a long term business. They'll likely prioritise the manufacturers which aren't going to run away as soon as they have another option.
  • Lolimaster - Saturday, July 25, 2020 - link

    Think better, commitment is key, the other 3 big will be regular customers for next 10 years and beyond ss long as TSMC delivers. Intel at any point in time could fix their agenda snd return in house or abandon cpus, who knows. You never play with regular customers.
  • dotjaz - Sunday, July 26, 2020 - link

    You completely missed the point. I said TSMC won't build capacity for Intel ("expand to meet that demand").
    If Intel is willing to pay higher price than it's regular price AND TSMC has the capacity to spare, sure, TSMC will take the order. TSMC don't take sides, they are also not stupid.
  • dotjaz - Friday, July 31, 2020 - link

    So I was right, TSMC isn't stupid, they picked up Intel's order because they have the capacity, and they view Intel as a short term customer so they won't build capacity for them.
  • Spunjji - Monday, August 3, 2020 - link

    @dotjaz - Of course you were. It takes a particularly warped perspective to perceive TSMC failing to build out new capacity for a single customer who may not be a customer in 3 years' time ad "taking sides" - but vladx is one such unique intellect 🤪
  • bernstein - Friday, July 24, 2020 - link

    simple: TSMC might force Intel to commit to buy a certain capacity for up to a decade. which would be an expansion of TSMC's current capacity.
    Intel might have to swallow that pill, just so it will be in a position to execute on their 3rd party fab contingency plan. Given how many mm2 of die's intel currently sells, buying capacity similar to amd's (or apple's, or nvidia's) to prepare for the worst, might be a relatively painless pill.
  • HomelessHardware - Sunday, July 26, 2020 - link

    Seems an easy business arrangement. Intc has billions in cash. They pay for the new plants. Provides a 5 year agreement to purchase all chips from the plants.

    This idea that tsmc, a 3rd party manufacturing company, would turn down the largest chip designer in the world because they have it good with amd is ridiculous. Does anyone have memory of amd beyond the last 3 years?
  • dotjaz - Monday, July 27, 2020 - link

    Your logic is simply laughable. Intel has it's own capacity, its demand is not long term at all. They simply won't sign such a WSA. Only stupid people would assume Intel would sign that.
  • vFunct - Wednesday, July 29, 2020 - link

    Intel has zero fab capacity at 7nm. TSMC does.

    If Intel offered TSMC $20 billion, TSMC would gladly build a new 7nm fab just for Intel, because that's their business model: to sell wafer capacity.
  • dotjaz - Friday, July 31, 2020 - link

    What a stupid comment. TSMC sells capacity not Fabs.
  • Spunjji - Monday, August 3, 2020 - link

    vFunct and HomelessHardware here providing the thrilling armchair analysis of "have big money, get big fab built". It's not that simple, and that's not what's happening.
  • dotjaz - Sunday, July 26, 2020 - link

    I doubt they would be stupid enough to do a WSA, both Intel and TSMC are smarter than that. It will anger TSMC's other customers and also Intel will have to meet the quota like AMD had to for years.
  • JCB994 - Friday, July 24, 2020 - link

    You ever build fabs in Asia? I used to work for Samsung and if they decided to build a fab, tools are being pushed in around 6-9 months later...assuming they can get the tools that fast. Not saying the fab is finished in that timeframe but if a HEPA is in place, the tool is in place underneath. Plastic sheathing is fine for walls.
  • eek2121 - Friday, July 24, 2020 - link

    The building isn’t the issue. The equipment is.
  • FullmetalTitan - Friday, July 24, 2020 - link

    I think the limiting factor here depends on what node they are pushing to a 3rd party. If they want anything from the 7/10nm designs, they need EUV capacity which is not a next day type of order. Otherwise, yeah I have seen some fast and dirty install and ramp up work done in fabs. If they are going to move an older, defect tolerant device to 3rd party, they could get away with the condensed time frame on expansion
  • Calin - Saturday, July 25, 2020 - link

    Intel's 14nm is close to the best there is, and Intel is on the third of fourth iteration of 14nm.
    Taking everything into account, moving 14nm production to a different maker (like Samsung, TSMC, ...) would probably double their costs - not to take into account scraping good-enough production capacity.
    So, what Intel needs is 7nm (or better) capacity (and a LOT of it), not 10nm or older. Just a bit of 7nm production capacity would be counter-productive - everyone would want the BEST there is, and expect a huge discount on the current products made on the older lithography nodes.
  • HomelessHardware - Sunday, July 26, 2020 - link

    Isn’t this why Intc is deciding what to do now? So they have 2023 7nm process rolling? Or is 3 years still not enough to start volume production (with money as no object )
  • PaulHoule - Friday, July 24, 2020 - link

    Intel has a 14nm fab that works. It has a partially successful 10nm fab, and given it's recent history of dissimulation I will believe in the 7nm fab when I have parts in my hand.

    I am glad to hear they are using 3rd party fabs because it means they have a real business plan.

    Congressfolks are right to be worried about American chipmaking. Intel seems to be failing at chip technology the same way that US Steel failed at steel (felt entitled to keep using the open hearth process when the rest of the world was using basic oxygen.) Intel is using outsize profits from 14nm to buy off the tech press, but they need to get TSMC to build a fab in the U.S. yesterday or we the U.S. will be 15 years behind of the rest of the world very quickly -- the crime is that Intel has been able to convince people that "the king is still on the throne and a pound is still worth a pound."
  • vortmax2 - Friday, July 24, 2020 - link

    It's high time American chipmakers step up their game.
  • MrVibrato - Friday, July 24, 2020 - link

    Umm... except the US wasn't holding onto the open hearth process while the rest of the world was using basic oxygen. The last open hearth steel furnace in Germany shutdown in 1993. China in 2001. Russia 2018. Ukraine and India still use open hearth to this day.

    And the US stopped using the open hearth process by 1992. Yeah, if your analysis of American chip making is as thorough as your analysis of American steel making...
  • PeterCollier - Friday, July 24, 2020 - link

    Because this isn't about some kindergartner's sense of morality and doing right about one customer, but about money.
  • azfacea - Friday, July 24, 2020 - link

    TSMC capacity is not an issue for intel anymore than it is for any one else. If i am TSMC tho why would I want to help rescue intel ?? why not help AMD/ARM destroy intel. TSMC does not owe anything to intel. its their product and their IP they can sell it to whoever pleases them the most.
  • Nozuka - Friday, July 24, 2020 - link

    Because Intel might never get their shi** together and then they have another huge customer.
  • azfacea - Friday, July 24, 2020 - link

    hate to break your bubble, but who will be whose customer down the road may have a bit more to do with what options exist in say 2025, not so much about intel's feelings getting hurt or not in 2020
  • dotjaz - Friday, July 31, 2020 - link

    Then TSMC can leave the decision until Intel decides to scrap their Fab. The market isn't going anywhere. AMD will pick up the loss which TSMC will manufacture. TSMC isn't losing anything.
  • medi05 - Friday, July 24, 2020 - link

    Intel is a juggernaut, so if anything but a handful of niche products is to be manufactured at 3rd fabs, it's a huge problem.
    And then, if it is only a handful of chips, why even bother.
  • Skeptical123 - Friday, July 24, 2020 - link

    ya at full scale this is going to be a snag, but I would think the margins on high-end dies and benefits provided by smaller transistors means Intel can justify paying a premium to get them made. Which would give them priority... but the same is true for nvidea high end gpus... but TSMC might just want to stick it's foot in the door of more gov/super computer stuff via Aurora.
  • JayNor - Friday, July 24, 2020 - link

    Intel was building the NNP training chips at TSM. Perhaps they have an option to substitute other chips, since it was canceled.
  • name99 - Friday, July 24, 2020 - link

    The questions that matter are
    - is Intel interested in profits today, or long term viability?
    - what does Intel consider to be its core competence/value?
    + x86 lockin?
    + CPU design?
    + sales contracts with large firms?
    + running fabs?

    Point is, nothing else matters until these are answered. If Intel believes profits today are primary, then they simply will not invest enough for the future. Discussing plans based on grand new fabs are a non-starter.
    Likewise if they consider their core competence to be selling to large companies and creating lock-in (cf eg IBM) then they don't much care about solutions that involve serious new technology.

    Personally I suspect at this point the best they could do is hire TSMC (or Samsung) to run their fabs. Saying "we'll switch to TSMC or Samsung" is not good enough by itself because
    - what do you do with those very expensive already existing Intel fabs?
    - TSMC (certainly) and Samsung (I think so) are already full up with existing customers.

    IF Intel is willing to accept the humiliation of saying "we screwed up utterly and completely", a solution is to let TSMC or Samsung completely take over their fabs. Note what this means
    - they need complete control to fire all the Intel people they want
    - and to replace all the equipment
    - and to tell Intel designers "these are the new design rules. get used to them".

    Half measures likely won't work. Will TSMC license to Intel? Maybe, for enough money, and some pretty serious clauses on what Intel can and can't use going forward. But is that good enough if the same management that have screwed up 14nm (oh yes, even 14nm was delayed, go back and look at the tape), then 10nm, then 7nm, stays in control?

    In a sense TSMC taking over Intel's fabs (call it joint development, call it renting, call it outright purchase) is probably the best possible outcome for the world as a whole. It gets us much more TSMC goodness, at much higher capacity, spread across the world (less China politics or natural disaster risk), and with process futures controlled, at least for another decade or so, by people who actually understand engineering.

    Would Intel accept that? My guess is Bob Swan and friends will not. If you own Intel stock and care about its value, start making your opinions known to the board. There ARE alternatives available that allow Intel to remain a viable tech company. But every day this is delayed is a day wasted... Switching those Intel fabs to TSMC or Samsung processes will not be immediate, neither will redesigning Intel cores to new design rules.

    So, what are you waiting for? The same crowd that have lied to you for 5 years and achieved very little in all that time are somehow going to perform a miracle tomorrow? Get freaking serious!
    GET THE FABS IN THE HANDS OF PEOPLE WHO KNOW HOW TO USE THEM...
    Or go broke. Your choice...
  • Eliadbu - Saturday, July 25, 2020 - link

    The capacity is there for who ever pay the most, you want to have x amount of waffers reserved for you - pay accordingly.
  • 0ldman79 - Thursday, July 23, 2020 - link

    It isn't an advertisement or something, this is an earnings report.

    Intel can't just ignore 7nm issues and hope investors don't ask. They've have issues on 10nm and continue to, tho they are improving.

    Pretty sure at this scale if they lied about 7nm progress they could be held liable in court. The people that make $$$ happen for Intel want to know what they're money is doing. They can't just pretend that they're not six months behind on a project and hope investors don't notice.
  • anonomouse - Friday, July 24, 2020 - link

    Tell that to 10nm in the year 2017.
  • Spunjji - Friday, July 24, 2020 - link

    "Pretty sure at this scale if they lied about 7nm progress they could be held liable in court"
    They did that with 10nm for 2 years straight. 🤷‍♂️

    As long as they keep raking in the cash and making their investors happy, the legality of it is strictly theoretical.
  • WaltC - Friday, July 24, 2020 - link

    Translated: "As long as Intel stock holders don't sue"....;) Intel has no choice but to be pragmatic because they'll never catch AMD with their current attitudes. Interesting to think that if not for AMD Intel wouldn't even be talking about 7nm--or anything else very interesting, for that matter.
  • YB1064 - Friday, July 24, 2020 - link

    Is it a problem with EUV machines or their process itself? Is there a link to an overview of the whole fabbing process I can read up on? Help a brother out Anandtech readers!
  • bernstein - Friday, July 24, 2020 - link

    since intel is using the same machines as everyone else (those from ASML) it's definitely their process.
  • peevee - Friday, July 24, 2020 - link

    Do you think with the prices like that they are not customized for every client?
  • Santoval - Saturday, July 25, 2020 - link

    What do you mean by "customize"? Optimize for TSMC or, er, "unoptimize" for Intel? The former might be an option, with the potential input or even active contribution of TSMC. The latter makes no sense though. Even if Intel wanted to"unoptimize" their EUV scanners out of sheer incompetence ASML would point out that what they want or ask would cause them this or that issue. Therefore I strongly doubt Intel's issue is with their EUV step-and-scan machines.
    Also bear in mind that Intel are not using EUV at all for their 10nm node. It is entirely DUV based with SAQP printing. Intel's 7nm parts are their first (production) parts with any EUV layers.
  • Speedfriend - Sunday, July 26, 2020 - link

    If you know anything about an EUV machine, you will know that physically is is not customised. It is the single most complex piece of machinery ever made
  • HomelessHardware - Sunday, July 26, 2020 - link

    Questions:
    1) if both tsmc and Intc are using the same machines, why can one push out 7nm and the other cannot? Is Intc adding more complication to their design?

    2) shouldn’t ASML help Intc fix whatever issues there are so they don’t lose the second biggest fan in the world?
  • Spunjji - Wednesday, July 29, 2020 - link

    Put simply: there's a lot more to a silicon manufacturing process than just the lithography machines.

    ASML aren't involved in those parts of the process. I'm not sure they *could* help.
  • serpretetsky - Friday, July 24, 2020 - link

    I believe 7nm is Intels' EUV process. I dont know how dependant their 7nm EUV is on 10nm. It might not have any of the same problems as 10nm (probably a bunch of different problems though), and they probably have separate parallel teams working on it. If that is the case then there isn't as much incentive to necessarily get 10nm done as soon as possible. 7nm is just as important if not more so.
  • Lolimaster - Saturday, July 25, 2020 - link

    Because even dumbed down their 10nm is still broken and unfixable.
  • Gondalf - Saturday, July 25, 2020 - link

    Hard times for AMD.
    A big contract with Intel means nearly zero 5nm silicon for AMD.
  • Spunjji - Wednesday, July 29, 2020 - link

    Gondalf posting FUD as per usual. Guess who already has orders down for the 5nm process before Intel even began considering this? I'd give you a list, but it's shorter just to say: not Intel. 😂
  • Santoval - Saturday, July 25, 2020 - link

    "Intel should concentrate on 10nm for the time being, why talk so much about 7nm when it's in such dire straits?"
    Because Intel's 10nm node is a fundamentally flawed node. It is also highly complex, even for today, and that complexity in turn leads to lower yields and thus higher costs and lower profit margins. In particular it is too complex for DUV lithography with SAQP printing. It should be OK with EUV (and maybe SADP), but EUV was not ready when Intel started developing their 10nm node (even today it's not 100% ready).

    Intel had a lot of trouble in particular with their switch to cobalt (from copper) of the two lowest layers (M0 & M1) of their metal stack and their introduction of COAG (Contact Over Active Gate). I heard they were thinking of ditching one of them in their + and ++ node variants, but I can't recall which one.

    They also decided to relax the transistor density of their highest density 10nm cell variant* in these subsequent node variants, in an effort to increase yields and drive down costs, even if that means slightly larger dies. I strongly doubt they have ever released anything with that quoted ~101 million transistors per mm^2 density except, perhaps, that poor Cannon Lake i3 they have been trying to forget. Even that was clearly not fully fabbed at the 10nm node's highest density, since that would make no sense. Perhaps the only Intel part that *was* fully fabbed at that node if the 10nm logic die of Lakefield, since that would increase its energy efficiency.

    *Intel's 10nm node is employed in *three* different cells, ranging in density from 67 MTr/mm^2 (highest performance - lowest density, with 12 fins) to 101 MTr/mm^2 (lowest performance - highest density, with 8 fins). All three cells are used for different parts of the same processor, depending on the best fit of each. In other words the quoted ~101 MTr/mm^2 transistor density of Intel's 10nm node is largely a marketing fabrication (which is "not even wrong", except it really is). That density is only used in cells for things like IO and uncore parts, *not* for the CPU and iGPU cores, which require the lowest density (with the possible exception of Lakefield). More in the link below.
    https://www.anandtech.com/show/13405/intel-10nm-ca...
  • HomelessHardware - Sunday, July 26, 2020 - link

    Hello, you seem to know more about this than 99% of the people here. What’s your take on the Intc 7nm delay? You think it’s really just a 6 months delay (an issue they have now fixed) or is this just a repeat of 10nm, and we can expect to hear about more delays in future conference calls? Any insight?
  • Spunjji - Wednesday, July 29, 2020 - link

    Lakefield definitely doesn't hit the ~100MTr/mm^2 density either 😬
  • HomelessHardware - Sunday, July 26, 2020 - link

    I couldn’t agree more. Look at aapl. They never disclose what they are doing until a product reveal. I don’t see why Intc couldn’t just talk about ramping up 10nm in this call. If the 7nm delay is really only 6 months, then why bring it up. Just push that news to the
    “We had an issue and we fixed it.”

    Instead it’s “we have a 7nm issue, we are pushing our entire product schedule back to 2023 and we are open to all options including outsourcing.

    It sounds like you don’t have any confidence that you can fix the issue, Bob. Great messaging.
  • fogifds - Monday, July 27, 2020 - link

    Intel should drop 10nm development. We're all over it. It's okay enough to ship laptop chips. The next two chips for desktop are 14nm, so that's all we're getting anyway.
  • YB1064 - Tuesday, July 28, 2020 - link

    Already happening:
    https://finance.yahoo.com/news/intel-now-ordering-...
  • JayNor - Thursday, July 23, 2020 - link

    While not in 7nm, and never planned for 7nm, Sapphire Rapids was announced during the cc to be sampling in 2H. It will support DDR5, PCIE5/CXL, AMX matrix operations on bfloat16, as well as avx512 bfloat16. Those are also technical advances, appearing in Intel chips far in advance of the AMD competing chips.

    Also, AMD is not currently using a tsm process that uses EUV, even though TSM offers it.
  • anonomouse - Friday, July 24, 2020 - link

    Yeah, and Cannon Lake was sampling in 2015, and "on track" in 2017. How did that product do?

    You can bet that by the time that Intel releases something it calls Sapphire Rapids, AMD will absolutely be on a more advanced process.
  • Rudde - Friday, July 24, 2020 - link

    Sapphire Rapids will not be 7nm. Ponte Vecchio is the chip that will suffer the fate of Cannon Lake, if any.
  • Santoval - Saturday, July 25, 2020 - link

    Sapphire Rapids is not of course impacted by the delay of Intel's 7nm node but it has its own separate issues. I mean, of course, the atrocious yield issues of Intel's 10nm node that have forced Intel to make two dual node releases (Ice + Comet Lake and Tiger + Rocket Lake), with the 10nm parts being released in far lower volume while Alder Lake (their first "top to bottom" 10nm release, to be fabbed at 10nm++, which is the node variant Intel hopes will *finally* resolve their 10nm node yield issues..) gets pushed back to 2H 2021, which as we all know means "no earlier than Q4 2021"..
    Sapphire Rapids will also be fabbed at 10nm++ but judging by Alder Lake's delay Intel are still not very confident they are going to resolve their yield issues. For Sapphire Rapids that doesn't really matter though : due to its obscene price yields are not going to matter much ;)
  • medi05 - Friday, July 24, 2020 - link

    AMD on 5nm in 2022 is very likely.
    On the other hand, given that transistor density of Intel's 10nm is higher than TSMC's 7nm, what Intel calls 7nm might well be along TSMC's 5nm.
  • Spunjji - Friday, July 24, 2020 - link

    FFS, when will random accounts stop posting this lie?
    Ice Lake on 10nm is ~50M transistors per mm²
    Renoir on TSMC 7nm is 63.33M transistors per mm².

    Intel's transistor density was *supposed to be higher* on 10nm, a claimed 100M transistors per mm². They relaxed it - twice - to meet yield targets. The world's most dense process is useless if you can't use it to manufacture a working chip.
  • DigitalFreak - Friday, July 24, 2020 - link

    But, but, I read it on WCCFTech. It has to be true, right? /s
  • Spunjji - Friday, July 24, 2020 - link

    I think people just copy/paste the first result for "Intel 10nm density", which is of course their marketing claim. God forbid you try to find out what Ice Lake's density actually is.

    Even Lakefield - which relegates most of the components that scale poorly to the 22nm die - is pretty rubbish compared with what AMD did with Renoir on TSMC 7nm.
  • regsEx - Saturday, July 25, 2020 - link

    Zen 2 4-core logic is 37 sq mm.
    ICL-U 4-core logic is 31 sq mm.

    Don't remember for sure, afair Zen 2 has more cache. But Intel is more feature-rich, so should have more transistors.
  • Spunjji - Wednesday, July 29, 2020 - link

    Are you quoting Matisse or Renoir for Zen 2? I was specifically referring to Renoir, which was laid down after AMD's first round of experience with manufacturing on 7nm and has shown significant improvements in density over Matisse; this is despite it also having to integrate all of the un-core, which Matisse does not.
  • regsEx - Saturday, July 25, 2020 - link

    And Intel never disclosed transistor count in ICL.
  • Spunjji - Wednesday, July 29, 2020 - link

    Jim Keller circa 2019 begs to differ.
    https://www.pcgamesn.com/intel/next-gen-cpu-archit...

    You would think they'd be boasting about it if it had actually hit their projected density, though.
  • Santoval - Saturday, July 25, 2020 - link

    "They relaxed it - twice - to meet yield targets".
    They do not need to relax their density in a technical sense, they just need to modify the layout of the three different 10nm cells in each processor. As a brief reminder (there's a more detailed one in the link below) what Intel calls "10nm node" refers to cells with three different densities :

    The 100.8 MTr/mm² is the high density cell library, but that is commonly only used for I/O and uncore parts of the processor - except perhaps when very high efficiency (at the expense of performance) *and* very small size are required, as in the case of the Lakefield 10nm logic die. I suspect that Intel fabbed the entire die except perhaps the Sunny Cove core at that density. This cell variant is also perfect for misleading marketing targeted at the general public (while at IEDM they disclosed the truth..).

    There are also the "Ultra High Performance" cells, with a density of just 67 MTr/mm² (intended largely for CPU cores and iGPUs) and the "High Performance" cells with an intermediate density of ~81 MTr/mm², for every bit that requires a moderate density and possibly for the cores of the mobile -U and -Y SoCs.

    Since Intel already developed three distinct cell libraries of three different densities they wouldn't need to relax either of them, they would just need to use more low and mid density cells in each processor and I'm quite certain that's how they "relaxed" their transistor density. The end result is the same but it must be much simpler (and thus cheaper) for them to "relax" transistor density in that way.
    https://www.anandtech.com/show/13405/intel-10nm-ca...
  • Spunjji - Wednesday, July 29, 2020 - link

    Their total density still doesn't meet the original spec for "Ultra High Performance" cells alone, though. It's hard to be sure - because they're being cagey about transistor counts and die size - but based on Jim Keller's statements about Ice Lake's rough transistor count vs. its measured die size, they're way down from where they'd be if the vast majority of their chip hit 67 MTr/mm². The best information I have suggests that even with Lakefield's compute die, which excises lots of the uncore to the "22nm" interface die, they're still in the ~60MTr/mm² region.
  • Santoval - Saturday, July 25, 2020 - link

    p.s. By the way, do you have a source for these two densities? Ice Lake's density, in particular, sounds extremely low. Even if Intel used their lowest density 10nm cells for the *entire* die then its transistor density should not be less than 67 MTr/mm². So where does this 50 MTr/mm² value come from?
  • Spunjji - Wednesday, July 29, 2020 - link

    I'm gonna have to come back to you on this. as I cannot for the life of me find the source, and I've been trying for ten minutes now. I really should bookmark this stuff.

    Again, though, that number is from Intel's original 10nm specs, back when they were still pretending Cannon Lake was actually shipping. I'm fairly certain they did not hit those projected densities.
  • Rudde - Friday, July 24, 2020 - link

    Intel was sampling Ice Lake in Q2 2019, with general release later this year.
  • ZoZo - Friday, July 24, 2020 - link

    *later _that_ year
    (what you wrote means that Ice Lake will have its general release later in 2020)
  • dotjaz - Friday, July 24, 2020 - link

    AMD is not using N7+ because it's a lame process. Nothing but a learning curve for TSMC. IP pool for N7+ was never good enough and it's not compatible with N7. TSMC also never had the capacity, so N7+ production hovered below 10k wafer per month with report of yield issues (<70% IIRC).

    By the time TSMC get more EUV gears, N6/N5 would be ready. Plus AMD uses a bit different version of TSMC's N7, it requires taller cells and more complex routing/metal layers, so it would stand to reason AMD can just take TSMC's N7+ and use it either.
  • eek2121 - Friday, July 24, 2020 - link

    TSMC stated a while back that yields for N7+ were above 92%.
  • extide - Monday, July 27, 2020 - link

    Yield percentage is a useless metric as it is totally based on die size. The number you want is defects per sq cm.
  • name99 - Friday, July 24, 2020 - link

    Why do people continue to post this nonsense about TSMC having "short" or "learning" processes? EVERY two or three years we see this, about 20nm, then about 10nm, now about 7nm+.
    Maybe if Intel bothered to learn from TSMC, and the value of constant small iteration rather than trying to make grand huge leaps, they wouldn't be in the mess they are in.

    Yes, N7+ is for learning. Just like 20nm. Just like 10nm.
    SO FREAKING WHAT?
    Learning is how you develop a 5nm process that comes in on time and on budget. If you think learning is lame, try the alternative...
  • Santoval - Saturday, July 25, 2020 - link

    "(<70% IIRC)"
    You mean dream yields compared to the yields of Intel's 10nm node, right?
  • TeXWiller - Friday, July 24, 2020 - link

    If I remember correctly, Sapphire Rapids was to be the product used in the Aurora. Those technical advancements are helpful on reaching the targets on power/performance that Aurora needs.

    AMD's Milan successor should start appearing soon after, at least on the basis of the systems already contracted or planned. It would be a good candidate for an implementation of DDR5/PCI5e support. The Frontier system (at least) is reported to be based on a custom chip anyway.

    Supercomputing is going console. You may quote that as necessary. ;)
  • JayNor - Friday, July 24, 2020 - link

    AMD FAD 2020 slides had no mention of DDR5/PCIE5. I recall that they need to move to version 3 Infinity Fabric to enable bandwidth of PCIE5.
  • Spunjji - Friday, July 24, 2020 - link

    Not sure what point you're trying to make here? Sapphire Rapids will follow Ice Lake SP by "roughly a year" - that's the same time-frame we're looking at for Zen 4, which will also have AVX-512, DDR5 and PCIe 5 support. It will also be based on an EUV process.

    But we'll see what each company has to offer *at that time*.
  • Santoval - Saturday, July 25, 2020 - link

    They are called "TSMC", not "TSM". AMD are introducing AVX-512, DDR5 and almost certainly PCIe 5.0 and bfloat16 with Zen 4. bfloat16 (without AVX-512) might even be introduced with Zen 3 if AMD gets many customer requests for it. I think they also plan to add CXL to Zen 4 (specifically for Epyc).
    AMD's Zen 3 based CPUs and their Navi 2 GPUs almost certainly have already started being fabbed on TSMC's 7nm+ node (which has 6 EUV layers out of 65 in total, while the first iteration of their 5nm node has 14 EUV layers out of a total of 59 - see link below), and are to be released by the end of the year.
    As for those "technical advances" you need to be more specific. Wikipedia editors would mark that sentence as "Avoid weasel words" if you wrote it in a Wikipedia article :)
    https://semiwiki.com/semiconductor-manufacturers/s...
  • Great_Scott - Thursday, July 23, 2020 - link

    This is literally insane. I don't like to subscribe to all of the various 10nm conspiracy theories, but really, small lithography and staying ahead (manufacturing chops, really) is all that Intel is good at.

    And now Samsung is ahead of Intel fabs? Global Foundries should have stayed in the game. If only they had known how things would turn out...
  • 0ldman79 - Thursday, July 23, 2020 - link

    I read more about Global Foundries, their investor pulled the 7nm equipment. They didn't so much choose not to run 7nm as they had the decision made for them.

    They may yet get back in the game. There's certainly a need for 7nm and beyond capacity.

    I don't think 10nm failure is complicated at all. They scaled better than the norm with 14nm tho they had a rough time at the start, Broadwell is essentially a mobile only generation. They tried to scale even better with 10nm and failed. They had to pull back on the density a bit to get it going. I don't know all of the details but the density has gotten a bit worse and they admitted they loosened the specs a bit to get 10nm functional and yields improved.

    One reason Intel is having so much trouble is they're trying to go a half node better than everyone else with each jump. That's why everyone says that Intel's 14nm is more like TSMC's 10nm, Intel's 10nm is more like TSMC's 7nm... thing is physics dictates it's only going to get harder as everything gets smaller. You can't keep expecting to gain more at each step when the reality of physics says that gains will actually be more difficult exponentially moving forward.
  • dotjaz - Friday, July 24, 2020 - link

    They knew Intel had troubles. Yet they pulled out anyway. GloFo's 7nm wasn't all sunshine and rainbow either.
  • Eliadbu - Saturday, July 25, 2020 - link

    Samsung is ahead of Intel? With 8nm they are definitely not and their 7nm remains to be seen. Samsung is also experiencing issues with yields. getting to 7nm and below is really difficult and GloFo simply did not have the capital nor the skillful personal to get 7nm into profitable process.
  • shadowjk - Thursday, July 23, 2020 - link

    how the mighty have fallen...
  • shabby - Thursday, July 23, 2020 - link

    HAHAHAHAHAHAHAHAHAHAHAHAHAHAHA omg HAHAHAHAHAHAHAHAHAHA
  • Lord of the Bored - Friday, July 24, 2020 - link

    Took the words right outta my mouth.
  • brucethemoose - Thursday, July 23, 2020 - link

    So, how is Samsung doing these days? Seems like they're one stumble away from giving TSMC a rather remarkable market/geopolitical position.

    Also, I agree: Intel's advanced packaging will be a *huge* factor going forward. They can make up for some efficiency deficits by going slow and wide, and do it more economically than they could with monolithic mega-dies.
  • name99 - Friday, July 24, 2020 - link

    The public face of Samsung is troubling. They produce a lot more slideware, a lot more promising where they will be multiple years from now, than TSMC.
    This suggests that the rot, the control by MBAs and marketing rather than engineering, has already started...

    There's still time for them to pull back, but right now I'd say they're on the wrong track and so far there's no evidence that the uppermost levels of Samsung (or the SK government) understand the trajectory that Samsung Semiconductor is on and where it ends.

    But single advanced fab for the entire world? Probably not yet.
    While Samsung seems determined to copy every MBA-driven failure of Intel, SMIC is just as determined to copy every engineer-driven success of TSMC; and so far they've done pretty damn well, a whole lot better than detractors were claiming just a few years ago. And having Huawei (I assume) ready to flood them with orders and cash, won't hurt that on-going advance...

    https://www.kitguru.net/components/cpu/christopher...
  • pikunsia - Thursday, July 23, 2020 - link

    The big delay has not been 7nm FinFet but 10nm lithography. I think this is very clear for everyone. 10nm release date? How to know it if it is delayed everytime.
  • Xex360 - Friday, July 24, 2020 - link

    How could they mess up so badly, they had all the money and time on the world, ever since Sandy Bridge in 2011 they didn't do anything, money kept pouring.
  • Zingam - Friday, July 24, 2020 - link

    Maybe the smartest people don't want to work for them.
    Or just maybe too much office politics and PMs?
  • PeterCollier - Friday, July 24, 2020 - link

    It's called affirmative action
  • name99 - Friday, July 24, 2020 - link

    Easy target. Wrong I think.
    Intel made the DELIBERATE choice to delay EUV and simultaneously throw ten different unproven technologies at 10nm. See my long comment on this further down.
    Affirmative action and other politics doesn't explain this choice; a finance (rather than engineer) mindset is what led to it.
  • PeterCollier - Saturday, July 25, 2020 - link

    Diversity initiatives caused a loss of morale as well as allowed objectively less talented minorities advance to positions of power and stymie the decision making processes that lead to the 10nm fiasco and now the 7nm fiasco.
  • Spunjji - Wednesday, July 29, 2020 - link

    You have no evidence for this assertion. It's based on your own rampant confirmation bias. Please don't bring this sort of crap here, it's just not necessary.
  • PeterCollier - Saturday, July 25, 2020 - link

    The Boeing 737 Max project is another example of a diversity fail.
  • Lord of the Bored - Saturday, July 25, 2020 - link

    No, that's the end result of letting McDonnell-Douglas take over the company when they merged back in 1997.
    MD-D SOMEHOW came out of that merger with all their guys in top management positions despite the fact that they were a company run by accountants instead of engineers, had been struggling for years, and ... oh yeah, Boeing had just bought the entire company. So the new Boeing management started running Boeing like they'd run MD-D: into the ground by prioritizing short-term profits at the expense of engineering expertise. And with passenger jet replacement cycles being fairly long, it took some time for the full damage done to become evident(though the 787 was a catastrophic cock-up, it didn't actually kill people so it flew under the radar when it flew at all).

    https://finance.yahoo.com/news/1997-merger-paved-w...
    https://medium.com/swlh/boeings-culture-crash-c150...
  • Kamen Rider Blade - Monday, July 27, 2020 - link

    And this is why we need HEAVY Anti-Trust regulations.

    And specific regulations against "Bean Counters"
  • FunBunny2 - Saturday, July 25, 2020 - link

    riiiiiiiight. and our previous Black president was really born in Kenya?

    the problem with the Max was obvious with the NG, previous 'engine upgrade', to a 1964 airframe. the -100 was designed to fly out of the way small airports without ramps, tubes, and such to terminals; i.e. land and passengers walk out the underbelly extended stairs, just like in larger business jets today. this put the engines, JT-8s, in the wings and with sufficient clearance. this is, technically, a by-pass engine, but not enough that you'd notice. Boeing had to fabricate a 4 foot extension tail piece in order to mount reversers that would actually work. the landing gear has, so far as I know, never been redesigned to accommodate high by-pass engines, so the Max is a failure. all CxO level decision(s) (they had multiple opportunities to fix the issue, particularly the NG when the writing was on the wall.)
  • Spunjji - Wednesday, July 29, 2020 - link

    It's an example of the problem of trying to engineer a product backwards to meet a marketing goal. Nothing more, nothing less.

    Your comments here are examples of an educational fail, not to mention a diversity fail in the sense that you have clearly spent too much time around people with certain kinds of political beliefs. We used to call them "Nazis" but they cry if you call them that now; I think "identitarian" is the new hotness.
  • twtech - Saturday, July 25, 2020 - link

    Sure, but where do you think these decisions come from?

    First, there are the executives themselves - who are they, how qualified are they - but if they're worth their salt, their engineering decisions are informed by what they are hearing from the engineers.

    Then there's execution. Hitting a target or not is not a foregone conclusion when the target involves developing something new. Yet again, this boils down to the efforts and competency of the engineers as a group.

    Sure, a single exec can muck up the works by making bad decisions, not listening to the employees - even deciding that the company no longer primarily cares about competency is an executive decision.
  • Spunjji - Wednesday, July 29, 2020 - link

    Oh cool, a racist moron showed up to the chat.

    Thanks, moron.
  • yeeeeman - Friday, July 24, 2020 - link

    Bad management
  • DigitalFreak - Friday, July 24, 2020 - link

    Yep, that's usually what it comes down to. Rarely is it the people in the trenches that are the problem.
  • twtech - Friday, July 24, 2020 - link

    Intel began a new hiring initiative in January of 2015, and it wasn't to hire the best employees they could get.

    Maybe it's just a coincidence, but I don't know - that's right around the time they went off the rails.
  • peevee - Friday, July 24, 2020 - link

    Both hires and promotions.
  • Spunjji - Wednesday, July 29, 2020 - link

    Great, the remnants of 8chan spilled out into the AT comments section. 🤢
  • vladx - Friday, July 24, 2020 - link

    They failed on risk managament, thinking they could pull same scalability as with previous nodes.
  • nicolaim - Friday, July 24, 2020 - link

    From PC Gamer:

    "Intel is likely to be keen on adopting a new standard metric as it tends to come out on top in raw transistor density as it stands today. Intel reports a density of 100.76MTr/mm2 (mega-transistor per squared millimetre) for its 10nm process, while TSMC's 7nm process is said to land a little behind at 91.2MTr/mm2 (via Wikichip)."
  • dotjaz - Friday, July 24, 2020 - link

    So what? TSMC's N7 is actually in HVM while intel can't even build anything beyond a 15W mobile part on their 10nm+. And you do realise Zen2 is actually denser than ICL, right?
  • Guspaz - Friday, July 24, 2020 - link

    For that matter, TSMC's N5 is already in HVM... and they're expecting N5P to enter HVM before the end of the year.
  • Guspaz - Friday, July 24, 2020 - link

    And, I should note, TSMC's N5 is estimated to be 171.3 MTr/mm2
  • JayNor - Friday, July 24, 2020 - link

    Intel shipping Agilex FPGAs and P5900 family 5G base station parts in 10nm.
  • peevee - Friday, July 24, 2020 - link

    And yet Intel has failed to achieve 100M with commercially viable yields, and is about 50M in reality, with limited frequencies even at that.
  • Geef - Friday, July 24, 2020 - link

    Haha maybe Intel should start working on 13nm instead. They can at least show they are still in the game! Forget about trying 7nm. I guess Delay Delay Delay does still work for investors as long as the company sells chips.
  • anonomouse - Friday, July 24, 2020 - link

    With the lead time that it takes to actually design, tape out, and fab chips, including validation, you don't decide _now_ to possibly move silicon components for a late '21 product to a different fab. If they're going to ship on time, or even close to it, they will have already decided and done design work by now, and possibly even done early tapeouts.

    Another report has quoted Swan as saying parts of Ponte Vecchio were "always" planned to be fabbed across externally, so it'd be good to get actual confirmation on that.
  • Ryan Smith - Friday, July 24, 2020 - link

    I can confirm that Swan said that. But I suspect he's counting the memory; Intel was never going to make the HBM stacks.
  • dotjaz - Friday, July 24, 2020 - link

    If we are talking about old Intel where architecture is tied to a certain process, then yes you are correct. But what they can do is port parts of the design to 3rd party design, even dual sourcing to buy the 3rd party version some time. GPU is far less complex to port than CPU.
  • anonomouse - Tuesday, July 28, 2020 - link

    Even the design cycle of same RTL -> physical design + implementation + integration -> tapeout -> validate and find bugs -> re-tapeout -> validate in product config -> release is still quite long. Longer than you'd realistically manage from now until end of '21.
  • supdawgwtfd - Friday, July 24, 2020 - link

    WANTED!

    A Basic Editor.

    Seriously. How does shit like this get through?

    There is no proof reading done at all is there?

    Just relying purely on spell check yeah?

    A 6 year old would tell you something is wrong with this sentance.

    "Intel has found the root cause of the issue and is moving to fix it, stating that that the company doesn’t believe that there are not any fundamental roadblocks in their 7nm. "
  • Ryan Smith - Friday, July 24, 2020 - link

    "There is no proof reading done at all is there?"

    Sadly, copy editors went away with the Great Recession, so no. But thank you for the quick check!
  • RSAUser - Friday, July 24, 2020 - link

    There is such a thing as Grammarly.
  • damianrobertjones - Friday, July 24, 2020 - link

    I once installed a Grammarly trial... I stated that I'd plagiarised the text from somewhere else etc. I'd 'literally' just written the 2 page post. Fail.
  • PeterCollier - Sunday, July 26, 2020 - link

    All it does for the plagiarism check is a simple phrase match search. It's not unexpected that one might use similar phrases as other authors, but the app won't know better.

    That being said, the spelling, grammar, and style checks that Grammarly performs are superior to Word.
  • DannyH246 - Friday, July 24, 2020 - link

    Send your complaint to marketing@intel.com they write most of Anand's articles these days.
  • catavalon21 - Friday, July 24, 2020 - link

    Riiiiiiiiiiiiiight. If that was the case, this article wouldn't be here.
  • Lord of the Bored - Saturday, July 25, 2020 - link

    I thought when the article was about Intel falling flat on their faces, the official word was that marketing@amd.com wrote most of the articles.
  • Spunjji - Wednesday, July 29, 2020 - link

    Depends which brand of toolbox is writing the comment 🙄
  • Lord of the Bored - Friday, July 31, 2020 - link

    That's what I was saying, albeit less veiled with sarcasm. It is rare for someone to complain about the site being an Intel shill in an article about Intel's hilarious failures.
  • Oxford Guy - Friday, July 24, 2020 - link

    Pragmatism is a rather vacuous concept in the world of technological business.

    Big business is always pragmatic. Corporations, by definition, are opposed to idealism. They exist to extract profit by selling things for a bit (or a lot) more than it costs to make them (albeit without environmental costs and labor costs often factoring in enough).
  • Carmen00 - Friday, July 24, 2020 - link

    Interesting article. I wouldn't have seen this coming even last year! This newfound pragmatism has the fingerprints of Bob Swan on it, I suspect. That's the same Bob Swan who doesn't have as deep a technical background as Lisa Su, and would be quite happy to NOT create the highest-performance (and most expensive to produce) chips, as long as he can sell to the majority of the market. That is also a pragmatic decision, but is it the right one? Which cutting-edge engineer wants to work for a company that is dedicated to producing middle-of-the-road chips?

    The last few x86 years have been revolutionary - Ryzen came out of nowhere and a resurgent AMD is making inroads everywhere. The next few x86 years look like quite the roller-coaster ride, too!

    (by the way, small typo: "and baring further issues" should be "and barring further issues". I'm sure that Intel wouldn't want any more of its issues to be laid bare!)
  • serendip - Friday, July 24, 2020 - link

    Middle-of-the-road Intel chips will have less performance than AMD's while being less efficient than ARM's. If I were an OEM product planner, I would buy AMD for x86 compatibility and ARM for thin and light devices with long battery life. The only way Intel can survive when squeezed like this is to drop their prices below AMD.
  • sonny73n - Sunday, July 26, 2020 - link

    “ The only way Intel can survive when squeezed like this is to drop their prices below AMD.”

    This really cracked me up. I think Intel won’t be able to get down from their high horse. Instead of looking for a proper solution, it’ll be much easier for them to stick the old dirty tactics - payouts to OEMs to gimp AMD products or it could be long term contracts that are still in effects. Take a look at the current Dell XPS line-up for instance. Dell have a handful varieties of 10th gen Intel Cores but no Ryzens at all. The Ryzen 4000 series clearly beat the craps out of all the latest Intel models but somehow Intel gets to be in premium products, exclusively. Not to mention the price - Intel laptops cost about $250 more than AMD equivalents.
  • sseemaku - Friday, July 24, 2020 - link

    Looks like American companies are done with advanced silicon manufacturing with Asian companies taking the lead.
  • Peskarik - Friday, July 24, 2020 - link

    that's what happens when you hire MBA TALENT into CEO positions.
  • Spunjji - Wednesday, July 29, 2020 - link

    Yup. I love that we have bona-fide dinguses blaming "diversity" for the problems, totally bypassing how it's the good-old-boys-club conservative business-school lads leading the enterprise right into the gutter.
  • yeeeeman - Friday, July 24, 2020 - link

    There is a high chance that Intel fabs will go the same route as AMD fabs.
    As for tsmc plans, I can tell you that Intel will use tsmc 6nm process for their chipsets in 2022.
  • Quantumz0d - Friday, July 24, 2020 - link

    No. Intel fab capacity is too much for TSMC is one, and Intel losing their US flagship Semi to a foreign country is not happening. That will be a big blow to US. Intel will keep their shit together once that 10nm gets optimized. Time will only tell what is future.
  • catavalon21 - Friday, July 24, 2020 - link

    Intel already produces wafers outside the U.S. They have wafer fabs in Israel (2 locations), Ireland, and China. Intel claims their "most advanced" manufacturing facility is in Israel.
  • yeeeeman - Saturday, July 25, 2020 - link

    I know tsmc cannot take over fabbing Intel chips, but I can tell you that at least the PCH will be made on tsmc 6nm for meteor lake (2022)
  • TristanSDX - Saturday, July 25, 2020 - link

    definitelly not. Chipsets do not require high speed, and dies are small, yield is always great
  • boozed - Friday, July 24, 2020 - link

    Hah!
  • ksec - Friday, July 24, 2020 - link

    @Ryan - Did Bob explicitly said 7nm coming in 2023? ( Which is somehow widely reported on the internet but not quoted from anyone at Intel ) 7nm was suppose to be late 2021, 6 months delay only make it 2H 2022.

    Having its yield 12 months behind Internal schedule is something different. And Intel is giving its engineers additional 6 months to hopefully catch up.
  • Wilco1 - Friday, July 24, 2020 - link

    2H 2022 means late 2022 for mobile parts. Volume and desktop parts would be in 2023 if there are no further delays. So unless you want to be disappointed, assume 2023 for 7nm.
  • Rudde - Friday, July 24, 2020 - link

    From the article:
    "The first 7nm client CPUs are now not expected before late 2022 or early 2023. Meanwhile the first 7nm server part is not expected until the first half of 2023."
  • ksec - Friday, July 24, 2020 - link

    I read that as an opinion. Not something directly from Intel.
  • Ryan Smith - Saturday, July 25, 2020 - link

    He explicitly said late 2022/early 2023. To quote Bob Swan's statement:

    "We now expect to see initial production shipments of our first Intel-based 7nm product, a client CPU in late 2022 or early 2023."
  • Duncan Macdonald - Friday, July 24, 2020 - link

    When Intel go the chiplet route to try to match AMD in the number of cores in a package, it will inevitably have a detrimental effect on inter core communications. Their current monolithic chips have faster inter core communications than AMD has between cores in different chiplets. (Access to data in the L1 or L2 cache in a different chipset requires at a minimum 4 hops between chiplets and the I/O die (requesting core - I/O die - core with the data - I/O die - requesting core). The monolithic current Intel CPUs just have 2 hops (requesting core - core with the data - requesting core).) How much of an impact this has in real world performance remains to be seen.
  • schujj07 - Friday, July 24, 2020 - link

    We seem with the Ryzen 3 3300X vs Ryzen 3 3100 how much of a difference it makes having all CPU cores in one CCX. Supposedly Zen 3 CCXs are not going to be split between 2x 4 core clusters and instead be a single 8 core cluster.
  • peevee - Friday, July 24, 2020 - link

    The solution is to have all cores and their personal caches on one chiplet (optimized for speed), and common LLC (and maybe some other components like memory controllers) on another chiplet (optimized for density and low leak current). In one over another configuration, probably, as LLC would not be too hot.

    Personal core caches might need to be made bigger if latency to such LLC increases. But cache coherency between them will be on the CPU chip.
  • mildewman - Friday, July 24, 2020 - link

    Yikes - 7nm was supposed to help put out the 10nm dumpsterfire.
  • name99 - Friday, July 24, 2020 - link

    And you believed that?
    Did you know that there's no such word as gullible in the dictionary?
  • coburn_c - Friday, July 24, 2020 - link

    All those times they talked about how their process was more advanced than the competition.. you almost feel bad for them.
  • Wilco1 - Friday, July 24, 2020 - link

    They very well knew it was lies all along. All this talk about having best transistor density when they can't achieve even half that in real chips. 10nm Lakefield gets 49MT/mm^2 vs claimed 100+MT/mm^2.
  • Peskarik - Friday, July 24, 2020 - link

    Swan wants to outsource manufacturing? Well, there you have a TALENTED manager. Outsource, cut and fire, the is all the TALENT could do, typical MBA waste.
  • Spunjji - Wednesday, July 29, 2020 - link

    100%. All the Intel shills are banging on about how this will kill AMD, but really, it's going to gut Intel's talent pool and vapourise their biggest advantage: vertical integration.
  • poohbear - Friday, July 24, 2020 - link

    The conference call was an absolute disaster! Intel stock dropped 20% in one day today...its worst one-day performance in its HISTORY. When Bob Swan said 7nm might be delayed by up to one year and they might have to outsource their manufacturing, analysts went ballistic on him! Bob Swan needs to be fired, his leadership is horrendous. Intel needs new leadership, the past few years have been a complete train wreck. Look what Lisa Su did to AMD in the past 5 years, a complete 180 turn around and AMD stock has reached an all time high today.
  • ExarKun333 - Friday, July 24, 2020 - link

    Now we know why Apple said 'see ya!'
  • name99 - Friday, July 24, 2020 - link

    *Now* we know?

    Did you ever think that those of us predicting a great future for ARM, and for Apple switching to ARM Macs, even back say three years ago, were not the "cranks" and "fanboys" we were insultingly labelled, just paying closer attention than everyone else?
  • Spunjji - Friday, July 24, 2020 - link

    "The first 7nm client CPUs are now not expected before late 2022 or early 2023"

    I reckon that means 2H 2023 at the earliest. I certainly wouldn't bet on them managing anything more than a pipe-cleaner before that, a-la Cannon Lake.

    The total bullshit-fest that was their communication strategy over 10nm has really dented my faith in their ability to predict future products with any reliability.
  • mrvco - Friday, July 24, 2020 - link

    Call me cynical, but it sounds like Intel plans to 'level the playing field' by high-balling 3rd Party fabs for short-term capacity which will increase AMD's costs and/or reduce their ability to ship product.
  • Spunjji - Monday, August 3, 2020 - link

    I don't think they could do this even if they wanted to. They'd have to actually manufacture something with that capacity, and they'd have to justify to their shareholders why they were paying so far over the odds to do it - which means it could never be more than a sideline for them as long as they still have their own fabs.

    Meanwhile AMD's entire capacity comes through 3rd party fabs, so they're always going to be able to place larger orders.
  • name99 - Friday, July 24, 2020 - link

    "A side-effect of this plan would have been that 10nm was to be a relatively short-lived process, allowing Intel to get off of the troubled process rather quickly and on to the more reliable 7nm process.
    "

    Why do you phrase this all as though it was random outside forces that imposed all this on Intel. EVERY ONE OF THESE STEPS was a deliberate Intel choice which was criticized from by outsiders for precisely the reasons we are now seeing.

    10nm was chosen not to be EUV because Intel though it was more important to goose profits than to move to (and LEARN ABOUT) new technology. End result
    - no 10nm
    - no higher profits
    - no learning about EUV -- meaning that now 7nm is delayed.

    Second aspect is 10nm was chosen to be super-dense compared to 14nm. Again, why? Apparently because it looked good on marketing slides. Intel could have chosen a slow deliberate path to 10nm, one step at a time (quadruple patterning as one step, COAG as another step, cobalt as another step, SDB as another step) learning with each step, backing out when something didn't work. But rather than copying TSMC (which does precisely this) Intel would rather mock TSMC as taking baby steps each year ("their 16nm still uses 20nm BEOL! hah hah hah!")...

    Third aspect is WTF does Intel KEEP thinking it needs to tell the world what it plans to do five years from now. STFU!!!
    Haven't you learned, from at least 2015, that you HAVE NO IDEA WHAT THE FUTURE LOOKS LIKE??? You aren't psychics. Your predictions ("more reliable 7nm process"???) are worthless. So STOP TALKING ABOUT THIS! Stop running the company based on marketing. Talk about what you have to ship today, and what you hope to do over the next six months. THAT'S IT.

    AMD isn't telling us crazy plans about what they'll be doing in 2025. Apple isn't doing this. ARM isn't doing this. TSMC isn't doing this. nVidia isn't doing this.
    Put a damn adult in charge (ie NOT Bob Swan) and fire these pinheads who seem to think you can run a company by blathering about how your 2027 technology will be the greatest on earth rather than shipping today. Intel is way deep into high on their supply territory; cut off the source, and cut it off with extreme prejudice.
  • FunBunny2 - Saturday, July 25, 2020 - link

    "You aren't psychics."

    and apparently, aren't physics, either.
  • Quantumz0d - Saturday, July 25, 2020 - link

    I usually do not agree on your ARM notions or such but you mentioned it well. That Bob Swan and their Investor greed is killing this company. They all should shut the fuck up on the future bullshit. And actually focus on existential technology and R&D with ROI only. Also their pathetic side garbage plans are useless, IoT, Mobileye and all it's not their core Bread and Butter they wasted a lot of talent on them.

    And next Intel invested billions into 7nm, I hope that stupid "pragmatic" trash is not real and they stick to their guns and achieve it even if it's late and get the recoup on the uarch side or price.

    And on the political aspect. America needs to get back into this Semi race, they are losing a lot. If this falls it's a big thing and bad thing for US.
  • Spunjji - Wednesday, July 29, 2020 - link

    All of the above!
  • phyar - Friday, July 24, 2020 - link

    Do you think that Intel will anyway be able to compete on price, thus spoiling everybody's business, in particular if third party fabs don't allocate enough capacity to it and at the right price?
  • KCballer - Friday, July 24, 2020 - link

    Where is that Intel schillz Decidium at? I was waiting for him to spin this as a win from Intel.. lmao
  • arashi - Sunday, July 26, 2020 - link

    AyyMD is going to get no chips cause Intel will buy ALL the capacity from TSMC! Sad!

    /s
  • Spunjji - Wednesday, July 29, 2020 - link

    Nothin' but tumbleweeds from our favourite spinner. Sad times!
  • eastcoast_pete - Friday, July 24, 2020 - link

    I can't help but wonder if and how that correlates to the fact that recent Intel CEOs had less and less of a background in relevant scientific and engineering disciplines. This reminds me of what happened at Boeing when their headquarters moved to Chicago, and became more a company run by lawyers and MBAs. Intel headquarters are still in Silicon Valley, but their leadership is just as divorced from the hands-on expertise.
  • eastcoast_pete - Friday, July 24, 2020 - link

    The other, I believe interesting what-if here is what if Intel's next core designs would be fabbed on TSMC's then-newest process. How would that affect AMD, if they would loose the process advantage they currently enjoy? All speculation now, but imagine 16 Golden Cove cores, but now in highly efficient TSMC 5 nm EUV - we'd finally have a real battle for the crown!
  • GeoffreyA - Monday, July 27, 2020 - link

    Yes, I would love to see a battle of pure microarchitecture against microarchitecture.
  • wrkingclass_hero - Friday, July 24, 2020 - link

    Their gross margin is going to take a big hit when they start using 3rd party fabs.
  • zodiacfml - Friday, July 24, 2020 - link

    Sounds like they have calculated that they will make more money if they outsource, similar to what they did with sticking to 14nm parts for years. Logical decision but with implications in the far future. I reckon, in 5-10 years, Intel will lose the fab business like AMD/Globalfoundries did and leave it to TSMC/Samsung
  • zodiacfml - Friday, July 24, 2020 - link

    Now I feel bad in the stomach. My fears that started with the 14nm debacle becoming true, executives and stakeholders are just cashing in, little interest for the company's far future. Expect a fabless Intel soon.
  • lmcd - Saturday, July 25, 2020 - link

    Why are people acting like Intel has never worked with TSMC before? They've prepared their Atom CPU cores to be manufactured on TSMC processes before, indicating they have experience working with TSMC custom blocks. AFAIK some of their wireless products are also fabbed there and/or their SSD controllers.

    Intel will assuredly bring their lithography back up to speed. Their teams are just too used to hyper-efficient designs from top to bottom and need to be less aggressive.
  • way2funni - Saturday, July 25, 2020 - link

    This whole article reads like a 'really bad news media management' situation.

    They are giving you a heads up on bad news (6 month delay) which nobody believes because after the 14nm desktop cycle that has lasted SIX YEARS ALREADY and intel is still not expected to drop a desktop 10nm CPU until H2 of 2021.

    IMHO they need at least a year or two to monetize 10nm on desktop node and get the revenue out before moving on - they can't just write it off after all this , rake the hit and jump straight to 7nm amirite?

    Then they leave the door open and essentially ANSWER the unspoken question hanging in the air "well, 6 months is fine - the economy is ficked anyway - but what if you DON'T make it and it ends up being SIX YEARS AGAIN?

    I dunno, maybe intel figures evan at single digit yields, the big server chips that retail 5k and up can still be profitable on 7nm.

    Then mobile. if 10nm on desktop arrives by H2' 21 , they can push 7nm desktop to H2 2023 which accomplishes all the goals. UNless AMD eats any more share (which they will)

    They may be figuring if AMD can execute on desktop via 3rd party fabs, so can they, but I dunno if I see intel opening up their IP portfolio to 3rd party fabs. The ip they have locked up under patent may be the only ace they have going forward but if they can still use their own fabs to make the chips that make the big servers go and then all the mobile stuff that still has a healthy margin, and farm out the less profitable stuff, even if they continue fucking up their nobes and yields, they may survive the 2020's.
  • Santoval - Saturday, July 25, 2020 - link

    "I dunno, maybe intel figures evan at single digit yields, the big server chips that retail 5k and up can still be profitable on 7nm."
    Do they really have single digit yields? I thought, at an absolute worst, they had low double digits. I suppose you refer to yields of large Ice Lake-SP dies, not small Ice/Tiger Lake ones, right?
    So they have seen no yield improvements from the latest iteration of their 10nm+ node variant which they are fabbing Tiger Lake with and whichever 10nm+ variant they are using to fab Ice Lake-SP, and just *hope* things will be fixed (somehow) with their 10nm++ variant for Alder Lake? Yeah, good luck with that. The low yields of Ice Lake-SP can be tolerated due to their obscene prices, but the same does not apply to Ice & Tiger Lake, hence Comet & Rocket Lake, respectively. As far as I know though they are not planning a dual node release of Alder Lake with yet another tortured 14nm node part.
  • Santoval - Saturday, July 25, 2020 - link

    It took a little while but the first completely non unexpected delay of Intel's 7nm node was announced. And I am certain it is only the first one. How much delay are Intel going to face in total? I bet a system based on a 16-core Ocean Cove-S fabbed at TSMC's 3/4nm node (:p) that it's going to be 3 x 6 months, i.e. a delay of an extra 12 months or 18 months in total. So, mark my words and prepare monies for that Ocean Cove system (with at least 64 GB of DDR5 please) : the first product fabbed on Intel's 7nm node is not going to be released before Q4 2023 - Q1 2024. To be more specific let's say December 2023, a few weeks before Christmas*. With the possible exception of Ponte Vecchio, since yields for a $500 million supercomputer are much less of an issue.
    *In low-ish, Ice Lake - level volume, along with some 10nm part which will comprise the bulk of the release. In high volume not before Q2 2024.
  • aperson2437 - Saturday, July 25, 2020 - link

    Maybe Intel should just focus on getting to 3nm using EUV lithography. I suspect that TSMC might shock everyone and they'll be selling industry leading 3nm chips in 2021. The Apple A14 chip for the upcoming iPhone 12 this year will have TSMC 5 nm chip technology. Intel better put the pedal to the metal. Google this ... TSMC 3nm 'risk production' in 2021 paves the way to 2022 mass production
  • AnGe85 - Monday, July 27, 2020 - link

    Currently, TSMCs roadmap for 3nm ist HVM in 2HY22, meaning smartphone SoCs (small chips), meaning no x86-CPUs until 2023. Additionally you can expect something like a 5nm-Zen at the earliest in 2HY23, because Zen4 is scheduled for 2022 and for example Epyc will be launched over various models throughout the whole year, so it is unlikely to see 5nm from AMD already early 2023.
    Additionally Intel has shifted Granite Rapids SP (7nm+) to 1HY23 and stated, that the first 7nm-CPU will be a client CPU, meaning this has to arrive considerably before Granite Rapids SP, so the first 7nm-CPU from Intel could appear by the end of 2022 (I would guess it will be a mobile chip rather than a desktop chip).
  • AnGe85 - Monday, July 27, 2020 - link

    Correction: "like a 5nm-Zen" should have been a "3nm-Zen"
  • Vitor - Sunday, July 26, 2020 - link

    By the time their 7nm get in the market, Apple will have their 3rd gen of desktop chips made in 3nm.

    I feel the sun is setting for Intel, even if slowly like a long summer day.
  • ijdat - Sunday, July 26, 2020 - link

    Spending a fortune on your own fab works so long as the savings on wafer prices outweigh the cost of the fab, and your process is either better than the foundries or at least as good.

    The problem for Intel is that the cost of building a bleeding-edge fab has been rising exponentially -- much faster than their overall sales -- so the first condition no longer applies. The second one is gone -- at 22nm they were first to FinFET (better than the foundries), 14nm was pretty good (and optimised for CPUs), but 10nm was an over-ambitious car crash and doesn't give them a solid foundation to build 7nm on -- which was late anyway compared to TSMC 5nm, and is now even later, assuming Intel can get it to yield resonably quickly like they didn't with 10nm. Meanwhile TSMC use the entire world's chip sales to pay for their R&D and fabs, and by doing "baby-steps" have already ended up ahead of Intel, and are pulling further ahead month by month.

    All this makes it well-nigh impossible for Intel to catch up on process technology in the short-term, which in turn makes it completely impossible in the long-term -- they've fallen behind and there's no way to fix this. The only thing they can do that might work is to divert all their huge process R&D expenditure into chip R&D and try and overtake AMD (and others) that way -- in other words, acknowledge defeat on the process front and pull out, and try and win on IC design (which they're perfectly capable of doing without the process millstone round their neck).

    Unfortunately this would mean such a massive loss of face -- admitting that their "world's best process" strategy was dead -- that I doubt if Intel management has got big enough balls to do it.

    On another subject, there's now only one 5nm process at TSMC, N5+ (improved performance, same design rules) has been effectively renamed N5, that's all that the fabs are making from now on.
  • Duncan Macdonald - Sunday, July 26, 2020 - link

    The statement by Intel could have been more accurately put as "When Intel was ahead of AMD, its management became complacent and did not press its R&D for several years. When AMD came out with its Ryzen chips the management had no good way to respond. Intel's multiyear inertia has left AMD on a more advanced manufacturing node (and about to transition to a yet more advanced one) and it will take Intel years to catch up to AMD. Before that happens much of the customer base is likely to move to AMD from Intel. Expect profits to decline for the next several years."
  • Igor_Kavinski - Monday, July 27, 2020 - link

    Look. Let's stop beating about the bush. What's the real source of the problem at Intel? When did all these process manufacturing troubles start? Who are the incompetent people who failed to execute their plans and have been giving nothing but excuses for the past few years now? Intel shareholders need to take action now before it's too late. Get rid of the problematic people ASAP before the cancer grows even more and metastasizes. Search out the tumors and give them the boot in the most unceremonious way, because that is what they deserve for taking their salaries and bonuses and still failing to keep up with their promises. This isn't a joke anymore.
  • Farfolomew - Monday, July 27, 2020 - link

    I agree whole-heartedly with this. I know friends who work at Intel l who, quite frankly l, have the work ethic of an overpaid middle class, part-time government worker (which they happen to be too, in the National Guard).

    Yeah they’re not the high-level engineers that are designing chips, but hearing about the company, it just feels like a pseudo-government operated dinosaur, not at all like the seemingly agile Apples, Googles, and Amazons that dominate the rest of the tech industry.
  • Spunjji - Thursday, July 30, 2020 - link

    The problem with this approach is that at a company as large as Intel, the only visibility you have of any problems is mediated through several layers of management - and to all intents and purposes, *they are the problem*. But when have you ever heard of middle-managers firing themselves?

    All you'd achieve is motivating a bunch of ass-covering incompetents to point fingers. After the axes fall, you're left with a bunch of cowardly idiots and none of the otherwise-competent staff they sacrificed to retain their careers.

    It's really not clear how you solve a problem like this. After a few generations of having incompetent cowards at the top promoting equally-useless people to positions below them - and the ensuing rounds of empire-building, recrimination and backbiting that result - you've baked the problem into the heart of the company. Jim Keller's departure illustrates pretty well how you can't solve a problem like that just by hiring a few good managers.
  • Silma - Monday, July 27, 2020 - link

    Remarkable! After a 5-year delay, 10 nm is officially there at Intel, but the foundry is incapable to produce en masse, or to increase GHz, limiting it to sub 30 W mobile processors.

    Now Intel is promising that it will reach mass production in 7 nm in 2023. Do you even believe it Anandtech?
    But assuming its true, everybody else will already have moved on 5 nm.

    As AMD has shown, 2 years of 7 nm advance was all it needed to best Intel in all categories of processors from mobile to datacenter.

    Now imagine an additional 2-3 years more advance for AMD, ARM-based processors, Apple & RISC-V.
    And imagine how much capacitiy will be left for Intel production when it will call TSMC.
  • Spunjji - Monday, August 3, 2020 - link

    Failing any unforeseen events (not to be dismissed), by 2023 everyone else will be beginning their 3nm transitions.
  • Farfolomew - Monday, July 27, 2020 - link

    This seems rather simple supply and demand economics for TSMC: they’re now even MORE in demand, so they should be charging a pretty penny for the luxury of manufacturing with them.

    Again, TSMC, would appear to holding the world’s chip makers’ nuts in a vice, they’d be silly not to squeeze
  • Oxford Guy - Saturday, August 1, 2020 - link

    Can't they go to Samsung if TSMC asks for too much?
  • Spunjji - Monday, August 3, 2020 - link

    We'll soon know the answer to that when we see what Samsung were able to produce for Nvidia. My understanding was that they're a little behind TSMC, but probably not so far behind that it'd be worth overpaying.
  • paul.atwork - Tuesday, December 15, 2020 - link

    Is this the start of RISC supremacy? Seems that now both Apple and IBM have released their 7nm into the market. (hot chips announce .. https://www.anandtech.com/show/15985/hot-chips-202...

Log in

Don't have an account? Sign up now