TSMC expects to begin 5-nm production in 2020

TSMC is already thinking a few moves ahead of its present processes. According to DigiTimes, the company expects to start production of 7-nm chips in 2018, and it thinks its rollout of 5-nm process tech will begin in the first half of 2020.

TSMC has reportedly been researching production at the 5-nm node for over a year. The company says it's made "significant progress" in using EUV (extreme ultraviolet) lithography, and says it'll most likely deploy this know-how at the 5-nm node.

The firm also hopes to tape out customer designs based on the 10-nm process during the first quarter of this year. The foundry expects to follow that up with a new 16-nm FFC (FinFET Compact) process, a version of its 16-nm node with lower power requirements and lower production cost. TSMC expects to own over 70% of the 14- and 16-nm fab market in 2016, rising from 40% in 2015.

Last but not least, the firm is looking to start volume production of its integrated fan-out (InFO) wafer packaging in the second half of this year, for "a few very large volume customers."

Comments closed
    • ronch
    • 4 years ago

    [Nursery rhyme]

    This is the way we shrink our nodes, shrink our nodes, shrink our nodes.

    This is the way we shrink our nodes, we do it every 2 years.

    • fredsnotdead
    • 4 years ago

    “TSMC expects to own over 70% of the 14- and 16-nm fab market in 2016, rising from 40% in 2015.”

    Because everyone else will have moved on to the next node by then?

    EDIT: And now I see DrCR already said much the same thing…

    • NeelyCam
    • 4 years ago

    This article from Ashraf Eassa gives reasons why TSMC is able to accelerate the node transitions vs. Intel:

    [url<]http://www.fool.com/investing/general/2016/01/20/exclusive-the-troubling-story-behind-intels-chip-m.aspx[/url<]

      • Stonebender
      • 4 years ago

      Some take ways: TSMCs 10nm will be essentially on par with Intel’s 14nm, which has already been in production for close to two years. They haven’t come out with any 10nm production yet, so by the time they do, they will be what, three years behind Intel in that regard? Then if you extrapolate that out, their 7nm will be roughly equivalent to Intel’s 10nm. Well, I think they’re underestimating the difficulty in getting a process that requires insane amounts of multi-patterning to give acceptable yields. 5nm will require EUV, so, yeah.

      • smilingcrow
      • 4 years ago

      Interesting that TSMC now have two teams working in parallel.
      Intel do that with architecture designs so I find it difficult to believe they don’t also do that for process development when you consider how much more difficult that is and how important that is for them.
      If TSMC can afford it I’m sure Intel can.

      • chuckula
      • 4 years ago

      It’s almost like TSMC has a — shall we say — Tick-Tock model in place.

      We’ll see what they are actually producing in 2020, but it’s nice to see that they won’t wait 5 years to do another node shrink.

        • smilingcrow
        • 4 years ago

        Yeah, this announcement is like a tick-tocking bomb waiting to explode and make them seem foolish, unless …….

      • blastdoor
      • 4 years ago

      [quote<]The question I have wanted to learn the answer to for quite some time is, "just how did this happen?"[/quote<] He answers that question with "two development teams." That's not untrue, but it's incomplete, and there's a more fundamental answer. Up until recently, TSMC has not had clients willing and able to fund that level of development effort. Now they do. It's all about the economics, and the economics of mobile are changing the chip business.

      • ronch
      • 4 years ago

      What I find funny in that article is how it makes it look so easy to fill a room with people who know how to do process nodes, and bloody-leading-edge nodes at that, and how it seems to make Intel’s missteps and troubles seem a lot worse than they are. Don’t forget, Intel is still ahead of everyone and their nodes have less marketing and more actual quality. And it’s not like Intel has less ‘pull’ in the industry to attract the very best brains.

    • Kougar
    • 4 years ago

    Don’t count your chickens before they’re fabricated. I’m sure the GPU vendors in particular won’t make that mistake with TSMC again.

    • Krogoth
    • 4 years ago

    In before the laws of physics and diminishing returns puts a swift end to this projection.

    • Unknown-Error
    • 4 years ago

    TSMC given time-frames? Oh, what could go wrong. Unless they just rebrand existing nodes. Intel obviously makes the best size improvement but Samsung also actually made progress.

    • guardianl
    • 4 years ago

    [url=http://sematech.org/meetings/archives/litho/euvl/20021015/Oral%20presentations/Oral%20111-PSilverman.pdf<]Intel was planning on EUV being needed/available in 2004.[/url<] EUV might become (economically) viable at some point, but for people in the semi industry its a joke that has long since lost it's humor.

      • chuckula
      • 4 years ago

      EUV is the technology of the future… and always will be.

    • smilingcrow
    • 4 years ago

    Donald Trump expects to be in the White House in 2016.
    Which is more preposterous?

      • chuckula
      • 4 years ago

      TSMC. Trump can always buy himself an invite.

        • smilingcrow
        • 4 years ago

        If he stopped speaking he might have a very thin chance but he’s polluted the stream so much now it’s completely toxic to a sane person.

      • ronch
      • 4 years ago

      I once viewed Trump as a refreshing break from all the politically correct politicians in America, but when I saw him speak (I only usually read the news), my hopes were dashed.

    • Fonbu
    • 4 years ago

    Wishful thinking at most!

      • Fonbu
      • 4 years ago

      After some thought (and reading), I actually retract this somewhat. Need to have faith in them to push this forward!!

    • PrincipalSkinner
    • 4 years ago

    And I expect Kate Upton will ask me out in 2017.

    • just brew it!
    • 4 years ago

    I find the prospect of chips with components that are only ~20 silicon atoms across to be mildly disturbing.

      • willmore
      • 4 years ago

      I find your lack of faith disturbing. *forcechoke*

    • brucethemoose
    • 4 years ago

    I’ll believe it when I see it. Given how much trouble everyone’s had below 32nm, I don’t think they can continue die shrinks at this pace.

    • The Egg
    • 4 years ago

    I also expect to start bagging 5 chicks per week by this summer at the latest. I think I’ve made “significant progress” with some special techniques, and expect “a few very large volume customers”.

      • morphine
      • 4 years ago

      *sings*

      Big bottom, big bottom…

    • Ushio01
    • 4 years ago

    So 28nm was first available on a retail graphics card back in January 2012, that’s 4 years ago so 5nm graphics cards if 16nm actually hit retail this year will be 2024 not that long in the future then.

    • TopHatKiller
    • 4 years ago

    murfff… odd, i’m bringing out 5nm in 2020. Huh. I guess we’ll see who’s really first.

    • WhatMeWorry
    • 4 years ago

    I would first get the 7-nm chips out the door, and then start talking about 5-nm. Sometimes the best thing to do is just keep your mouth shut.

      • deniro444
      • 4 years ago

      Well said!

      • willmore
      • 4 years ago

      It’s about communicating roadmaps to potential future customers. You and I are not TSMCs customers.

    • Hattig
    • 4 years ago

    5nm premium GPUs should provide between 38 and 60 TFLOPS computation power, assuming a conservative 50% increase in compute capacity per shrink from 14/16nm downwards. That might be too low of course (we could be over 100 TFLOPS a card in that case).

    28nm – 8 TFLOPS
    14/16nm – 14 TFLOPS +/- 4 TFLOPS
    10nm – 21 TFLOPS +/- 6 TFLOPS
    7nm – 32 TFLOPS +/- 9 TFLOPS
    5nm – 48 TFLOPS +/- 12 TFLOPS

    + = if you’re lucky / second generation (on that node) high end
    – = first generation on a node (e.g., Polaris 11 on 14nm is likely to be 10 TFLOPS, the bigger Polaris should be ~14 TFLOPS)

    The big assumption here is that TSMC will actually achieve 10nm, 7nm and 5nm on time. 10nm they must know enough about to be fairly confident, but the others are quite a way off.

      • NoOne ButMe
      • 4 years ago

      If Polaris 11 is over 8TFLOP, which is already a high estimate, I would be shocked.

      Polaris 10 appears to be a Pitcairn replacement. Polaris 11 is likely replacing Hawaii, possibly killing Fiji with architectural improvements.

      • ronch
      • 4 years ago

      I’m assuming you’re getting your 50% improvement from one node to the next based on transistor density, but what about power? If they can reduce power draw by 33% for every shrink which would allow power draw to stay the same from node to node while putting in 50% more transistors in the same die area, then that’s feasible. But of course we know maintaining 33% reduced power from node to node is a bit optimistic.

      • BlackDove
      • 4 years ago

      You taking into account die space for DP and SP cores in that calculation?

        • NoOne ButMe
        • 4 years ago

        Given he rates 28nm as 8TFLOP he’s following AMDs card at stock I think he is talking about AMD.

        FP64 on GCN takes up a lot less area than the ~40% incease that each SMX in GK110 has over GK10x due to FP64 units and datapaths/etc.

        AMDs card DO take an area hit and I think AMD would be better off at only 56/60 CU and being 1/2FP64 instead of doing 1/8 FP64.

    • NoOne ButMe
    • 4 years ago

    So 5nm will be the next “cheap node for everyone to use” after 14-16nm class.

      • orik
      • 4 years ago

      I was reminded of GPU manufactures announcing they were skipping 20nm with this announcement.

      “10nm and 7nm are too hard, we hope 5nm will be better”

        • NoOne ButMe
        • 4 years ago

        20nm sucked for what a GPU is aimed at, shrink it a little and add finFETs makes it not suck. If 10nm sucks they’ll skip.

        • ronch
        • 4 years ago

        Going to Mars is too hard. Let’s go to Uranus instead.

    • DeadOfKnight
    • 4 years ago

    How about they focus on getting things out on time like they haven’t been able to do in a very long time before building the hype for stuff on the horizon? No fabs have locked down EUV or 450mm wafers or any of this we’ve been talking about for 10nm processes and below, and TSMC doesn’t have a good track record for pushing things on the bleeding edge of technology.

      • NoOne ButMe
      • 4 years ago

      No on company locks down the technology unless they want to beat the full cost of early adoption.

      TSMC saying EUV is 5nm pretty much forces Intel and Samsung to aim 5nm if they were planning for 7nm before.

        • kuraegomon
        • 4 years ago

        I don’t think TSMC is capable of forcing Intel _or_ Samsung to do much of anything. Intel in particular has plenty of history of being a solo early adopter on the process front.

          • NoOne ButMe
          • 4 years ago

          And now that the costs have skyrocketed Intel has stopped.

          Talk to ASML If you want. It’s cheaper for everyone to change-over at once. If only Samsung and Intel do it, either they pay more, or ASML makes* (not pays) less.

          Intel wants 450mm wafers. Intel also doesn’t believe going solo is worth it anymore. It also lets then save a few hundred million dollars a year at minimum.

            • Stonebender
            • 4 years ago

            You’re clueless if you think Intel has stopped anything. At this point *EVERYTHING* revolves around EUV, and Intel is pushing hard. They didn’t give ASML $5 Billion just to wind up sitting on their hands. And you’re not going to see 450nm wafers from anyone anytime soon. The cost is astronomical, and not worth it.

            • NoOne ButMe
            • 4 years ago

            They stopped spending money to get there first, at least not as early. And 450mm wafers are dead until further notice.

            R&D costs to research and develop equipment has gone up the further node shrinks and the number of playesr who end up on that node has gone down. Now we’re at 3-4 players. Intel does not want to spend a few billion extra dollars to get tools only to have TSMC and Samsung walk up 1-2 years later and not have to spend that R&D money. This cost of billions of dollars is why the industry needs to change overall all at once. ASML would force both sides to pay a premium for tools if they support both sides. And no one wants that.

            450mm wafers are on hold because only Intel wants them really bad. And maybe Apple. If TSMC, GlobalFoundries and Samsung saw the same value in 450mm wafers that Intel did they would be willing to invest, which would put the 5 billion plus dollars into small amounts per each company.

            edit: 5 billion is very optomistic, looking around for more accurate R&D numbers, 20-30 billion is more accurate for 450mm wafers.
            450mm wafers also won’t reduce costs for small volume parts : [url<]http://www.electronicsweekly.com/blogs/viewpoints/450mm-delayed-till-2023-450mm-euv-light-source-2014-03/[/url<] For TSMC and to a lesser extent Global Foundries and Samsung 450mm would be an extra cost which wouldn't allow them to reduce costs. And for small customers would lead to increased costs. all for 5-15 billions dollars of total investment in R&D and building a fab each.

            • Stonebender
            • 4 years ago

            You misinterpreted my post. Intel gave ASML $5 Billion to support EUV research, it had nothing to do with 450mm wafers. You’re nuts if you think Intel isn’t pushing hard on EUV. Somebody has to get their first, and they are damn well sure it’s going to be them. And the tool wise cost hasn’t been as high as you’d imagine (at least from a lithography perspective), 10nm is using essentially the same tools that have been in use since 22nm.

            I can’t even imagine the cost of switching over to 450mm wafers. Nevermind the cost of the tooling, the fab downtime would be a killer.

            • NoOne ButMe
            • 4 years ago

            Ah. Okay. When did Intel do this 5 billion? Their stock and other capital investment was for EUV and 450mm wafers was a little over 4 billion last I read it. 450mm wafers was effectively canned for now so I understand nearly all the money Intel put in going to EUV.

            Intel might be first to EUV, but they will be a close first. Instead of being as early as 4 years they will be within a year, probably within months of implementing a process. So the R&D will fall less on them than previously. TSMC push for 10nm indicates that. Even if TSMC 10nm in characteristics is closer to Intel 14nm than Intel 10nm, the tools are the same I believe.

            +1

    • DrCR
    • 4 years ago

    [quote<]TSMC expects to own over 70% of the 14- and 16-nm fab market in 2016, rising from 40% in 2015[/quote<] Is that only because Intel will have moved on by then?

      • NoOne ButMe
      • 4 years ago

      I do my believe Intel is considered as they do pretty much all in house. Their largest 14nm customer was going to be Altera. And Altera is not quite a customer anymore.

      • orik
      • 4 years ago

      because Intel doesn’t use their fab for other people’s chips much? (I think, let me know if I’m wrong)

      • maxxcool
      • 4 years ago

      My guess is the plan on outliving GloFlo and price cutting them to death.. *IF* they can keep their customers happy that is..

      • Zizy
      • 4 years ago

      If all Samsung 14 nm production is counted, they surely made more chips than TSMC in 2015 (either measured by area or number; probably they aren’t comparing revenue/profit).
      Total PC market (in units) is about as big as S6+Note5+iPhone combined.
      Unless Intel sells way fewer 14 nm chips than the old 22 nm ones, something I doubt, Intel cannot be included in this comparison.

    • USAFTW
    • 4 years ago

    Are they starting to overtake Intel or is it just air?

      • chuckula
      • 4 years ago

      Oh it’s 5nm alright!

      For a rather large value of “5”.

      • ronch
      • 4 years ago

      Intel calls their cadence ‘Tick Tock’.

      TSMC’s is called ‘Kableem Kablam Kabloom’. So yeah, it’s meant to overtake Intel.

        • chuckula
        • 4 years ago

        In honor of David Bowie, GloFo is now on the “Wham Bam thank-you ma’am” cadence.

      • Mr. Camel
      • 4 years ago

      Maybe they are using “air gaps” in their 5 nm interconnect design.

      • not@home
      • 4 years ago

      5 to the base 20

      • Hattig
      • 4 years ago

      It’s mostly air. Check the table at the bottom of [url<]http://www.eetimes.com/author.asp?section_id=36&_mc=RSS_EET_EDT&doc_id=1328711&page_number=4[/url<] Intel's 14nm has smaller fin pitches and gate lengths than TSMC and Samsung's processes. Samsung's 14nm LPP appears to be a minor shrink again however, but probably only reduces the gate length a bit more towards Intel's gate length. That doesn't mean they are bad, even when compared to Intel's process directly. There are a lot of other factors at play.

    • ronch
    • 4 years ago

    Whoa that’s THREE Generations down the road from 14nm in just four years. My skepticism is killing me.

    Sure makes my 32nm FX chip look like a relic.

      • Deanjo
      • 4 years ago

      [quote<]Sure makes my 32nm FX chip look like a relic.[/quote<] It was a relic when it debuted.

        • Meadows
        • 4 years ago

        Wasn’t *that* hopeless back then, but an additional 3-4 years of not getting any better certainly left its mark.

          • xeridea
          • 4 years ago

          The uArch improved but last 2 generations were only in APUs and then laptops. Sticking on 32/28nm certainly hurt.

        • chuckula
        • 4 years ago

        Ahem… we prefer the term “classic” thank you very much!
        — AMD marketing.

          • maxxcool
          • 4 years ago

          lol, before the comment expanded when I clicked it I thought you were referring to our age 😉

      • blastdoor
      • 4 years ago

      I don’t think it’s quite that radical, is it? They are talking about “starting production” which probably means volume is the next year.

      So in terms of volume production, that means:

      2015 — 16 nm
      2017 — 10 nm
      2019 — 7 nm
      2021 — 5 nm

      That’s basically Moore’s Law continuing until 2021.

      So long as Intel doesn’t nab Apple as a foundry customer, TSMC ought to have a customer willing to fund the investments needed to meet this schedule.

        • Rand
        • 4 years ago

        I think we can rule on Apple using Intel’s process. Their ARM chips are indirectly competing with them. The worse Apple performs the easier it is to sell Android where Intel can compete. And there is always that small chance Apple ports OSX to ARM.
        Safer just to keep Apple out.

          • nexxcat
          • 4 years ago

          [quote=”Rand”<]I think we can rule on Apple using Intel's process. Their ARM chips are indirectly competing with them. The worse Apple performs the easier it is to sell Android where Intel can compete. And there is always that small chance Apple ports OSX to ARM. Safer just to keep Apple out.[/quote<] There are a lot of what-ifs here, and this is where things get complicated. I can certainly see one scenario going the way you wrote, but here's another scenario, based upon the following premises: [list=1<] [*<]Intel chipsets do not score a major mobile design win. [/*<][*<]Apple iPhones continue the current market share. [/*<][*<]ARM core speeds continue to lag behind x86-64. [/*<][*<]Intel has excess fab capacity. [/*<][*<]The world continues its transition to smaller devices. [/*<] [/list<] I think 1-3 are relatively safe premises. I have no idea about #4, but if #1 holds, all the capacity forecasted for Intel chipsets will remain idle. At some point, that capacity would need to be used, and because of #5, Intel won't have the large x86-64 products to soak up the excess. At that point, I'm not sure Intel can [i<]afford[/i<] to not court likes of Qualcomm, Broadcom, etc., and of course, Apple. It would be very, very interesting if excess capacities in Intel fabs coupled with relative slackness in Intel-designed chips dictate that they either need to downsize their fabs or they sell excess capacity to fabless chip designers. If that happens, I can see Intel courting Apple if their excess capacity is large enough, since Apple will bring very steady, large volume.

      • Stonebender
      • 4 years ago

      Especially since they haven’t even caught up to Intel’s 14nm, which has been out for close to two years.

Pin It on Pinterest

Share This