ARM and TSMC tape out test chip on 16-nm FinFET process

One way Intel hopes to capture more share of the market for mobile chips is by taking advange of its more advanced manufacturing capabilities. The new Atom SoCs announced today, for instance, are made on a 22-nm process with a 3D or "tri-gate" transistor structure. This structure allows the chip to operate more effectively at lower voltage levels, reducing power consumption and extending system battery life compared to chips based on traditional planar transistors.

The rest of the semiconductor industry has trailed behind Intel in the move to 3D transistors, more broadly known as FinFETs. Without this new structure, the move to process geometries below 28 nm has been slow, in part because going smaller with planar transistors doesn’t seem to pay many dividends in terms of power efficiency or costs.

Fortunately, it appears the leading chip foundry, TSMC, is making good progress toward its own implementation of FinFETs on a 16-nm process. Today, TSMC and ARM revealed that they taped out a fairly complex test chip on TSMC’s 16-nm FinFET process late last year. The chip includes dual Cortex-A57 and quad Cortex-A53 CPU cores in an asymmetrical big.LITTLE configuration similar to what a final, consumer-oriented SoC might use. This development could blaze the trail for a host of SoCs designs from customers of ARM and TSMC.

The ARM announcement claims TSMC’s 16-nm FinFET process has dramatic benefits: "designs could gain >40% faster speed at the same total power, or alternatively reduce >55% in total power at the same speed over 28HPM." Those claims are similar to what Intel has said its 22-nm tri-gate process offers compared to its own 32-nm planar tech. TSMC expects to have "more than 20 customer tape outs" on its 16-nm FinFET process in 2014.

Comments closed
    • ronch
    • 6 years ago

    Oh man. I love my FX-8350 but I can’t help but wish it’s built on a much smaller process node. 32nm is so yesterday. Hey ATIC, what’s all that oil money you’re pouring into GF good for? And please stop driving AMD into the ground with all your frickin’ one-time charges!!

    • ronch
    • 6 years ago

    Hey AMD, I hope you guys are strongly considering using this node or a similarly small process for whatever succeeds Vishera! We could all use an update on the desktop, you know.

    • HisDivineOrder
    • 6 years ago

    If only the fab industry could get out in front of Intel, maybe ARM could finally light a real fire under Intel’s ass and get them innovating again…

    …eventually.

      • Klimax
      • 6 years ago

      ARM will hit wall sooner then they’d get to challenge Intel. Although there are many options for upgrading performance none seems to be really power efficient.

      (And that is why Intel seemingly doesn’t innovate, most of optimizations are either done or not applicable and thus performance increases got much smaller)

    • UnfriendlyFire
    • 6 years ago

    Any guess of how long it’ll take for AMD to start using the 16-nm?

      • mesyn191
      • 6 years ago

      2016 according to AMD’s roadmap from a while back.

      • ronch
      • 6 years ago

      Probably not before the Second Coming.

      • Unknown-Error
      • 6 years ago

      Don’t think it works that way. This process along with GF 14XM are targeted at mobile & low power devices and not the typical x86 CPU/APU and dGPU.

      Plus, don’t forget that AMD is tied to GF until 2024 thanx to that WSA. So their incentive to fab at TSMC is getting lesser and lesser. Otherwise they have to pay GF massive amounts of cash.

    • jdaven
    • 6 years ago

    Since Intel has effectively delayed the bulk of its 14 nm chips to 2015, TSMC could ‘catch-up’ with them if they can get 16 nm chips out in the first half of 2015. This seems doable.

      • Stonebender
      • 6 years ago

      Don’t hold your breath.

        • Srsly_Bro
        • 6 years ago

        I encourage people to hold their breath.

      • NeelyCam
      • 6 years ago

      You’re joking, right..?

        • jdaven
        • 6 years ago

        I thought you were joking when you said Intel would have 14 nm chips available at retail in 2014.

          • NeelyCam
          • 6 years ago

          Oh… so you’re [i<]not[/i<] joking...?

            • jdaven
            • 6 years ago

            All I’m saying is that the 28 to 20 nm transition might have a lot of overlap with the 16 nm transition. TSMC might carry both processes for different segments.

            • Ninjitsu
            • 6 years ago

            Seems like a very expensive proposition, though.

      • Andrew Lauritzen
      • 6 years ago

      Are you seriously basing this on what TSMC calls “16nm” vs. what Intel calls “14nm”?

      • Ninjitsu
      • 6 years ago

      Um…unlikely. TSMC 20nm will probably enter mass production in Q3, and we *might* see 20nm based chips launch along side Broadwell’s December launch.

      Think about it, Nvidia had to introduce Maxwell on 28nm, so it’s pretty obvious TSMC are late.

      TSMC’s 16nm is unlikely before Q4 2015, as far as I can see. That might be a part of the reason why Intel shifted the bulk of its 14nm stuff to 2015, they realised that they could announce 10nm when TSMC finally manage to mass produce FinFETs.

      Also, that would mean money spread out over a longer period of time for the new fabs, and existing fabs make money for longer. The opposite may happen to TSMC, they may not be able to keep the 20nm fabs occupied once they start their 16nm facilities, if they rush it.

      • Klimax
      • 6 years ago

      Current delay is reportedly due to Haswell inventory… And believing TSMC about readiness is bit unwise.

      • Narishma
      • 6 years ago

      TSMC’s 16nm is equivalent to Intel’s current 22nm, not the delayed 14nm.

      • Unknown-Error
      • 6 years ago

      You are confusing Broadwell delay with entire 14nm process. Broadwell delay does not affect mobile device SoC using 14nm. In fact Intel is accelerating low power device fabbing. Classic CPU enthusiast will be annoyed but this is how its going to be. New priority is Atom, Quark, etc based systems. Don’t forget, even the TSMC planar 20nm is mystery right now.

    • NeelyCam
    • 6 years ago

    I thought the 16nm tapeout already took place a year ago? Haven’t they talked about that for a really long time?

    Also:

    [quote<]"Innovation is continuing at TSMC with the development of enhanced 16nm FinFET (16FF+) process that will offer an additional 15% performance improvement with no power or [b<]area penalty.[/b<]"[/quote<] Area penalty? With a shrink there sure as hell shouldn't be an area penalty. In fact, if they say that, the implication is that there is no area benefit (which has been rumored plenty, but I guess now it's official) which will come with a severe [b<]cost penalty[/b<] I guess Intel's presentation was right when it said TSMC will stop area scaling, even though TSMC was vehemently denying that and saying the slide was based on old data.

      • jdaven
      • 6 years ago

      Didn’t Intel show a test chip running on a 14 nm node at one of their conferences a year ago?

        • chuckula
        • 6 years ago

        More like a fully-functional Broadwell running a full OS and displaying complex graphics….

        [url<]http://www.anandtech.com/show/7318/intel-demos-14nm-broadwell-up-to-30-lower-power-than-haswell[/url<]

      • chuckula
      • 6 years ago

      The area penalty statement is probably comparing that uber-enhanced 16FF+ process to their “regular” 16nm finfet process. Basically they are trying to claim that they can get higher performance without having to space the transistors out more and hurting density.

      I don’t think they are comparing uber-enhanced 16nm to older processes with that statement.

      • jensend
      • 6 years ago

      The “15% performance improvement with no power or area penalty” is the 16FF+ process compared to 16FF, not 16FF compared to 20SOC or whatever.

      The tape-out of a 16nm A57 [url=https://techreport.com/news/24598/arm-and-tsmc-tape-out-64-bit-chip-on-16-nm-finfet-process<]did indeed happen almost a year ago[/url<], but that may have been just a single core and there was no mention of it being a full SoC. This new press release is saying a dual-A57+quad-A53 "BIG.little" SoC taped out in December. Much more complicated chip.

    • Milo Burke
    • 6 years ago

    Is this going to benefit AMD’s ability to compete on the mobile and desktop CPU front?

      • chuckula
      • 6 years ago

      No.

        • Milo Burke
        • 6 years ago

        Doesn’t AMD use TSMC?

        I thought AMD was stuck on 28nm because the fabs they use weren’t advanced enough. But if the fabs are already gearing towards 16nm production that can increase performance by 40% or decrease power use by 55% in a single generational improvement, can’t AMD work this fabrication into their existing architecture to whip out a mean contender to Intel parts?

        I’m not suggesting it is easy or instantaneous or a utopian solution. I just wonder if this fab evolution is something AMD can use, and if not, why not.

          • Stonebender
          • 6 years ago

          A tape out is a far cry from production. I wouldn’t expect to see any 16nm chips from TSMC until maybe 2016. By then Intel will already be moving to 10nm.

            • Milo Burke
            • 6 years ago

            Thanks for the clarification, Stonebender and chuckula.

          • chuckula
          • 6 years ago

          AMD uses TSMC for gpus but the only cpu/apu parts that TSMC makes for AMD are kabinis. Those chips need to be very very cheap. That means that when AMD finally gets future parts out, that 16nm process won’t be cutting edge. It could be nice for discrete GPUs though.

      • Srsly_Bro
      • 6 years ago

      AMD doesn’t have the ability to compete.

      • ish718
      • 6 years ago

      These 16nm chips are going to be super expensive lol

    • anotherengineer
    • 6 years ago

    Sigh now how long do I have to wait for a 16nm Radeon?? 😉

    *waits for Neely’s comments*

      • Srsly_Bro
      • 6 years ago

      he/she has already posted.

        • ronch
        • 6 years ago

        I have a feeling Neely is a man. Not as manly as men with fabs though.

          • NeelyCam
          • 6 years ago

          Huh? You have feelings?

            • ronch
            • 6 years ago

            Hey, what do you think I am?! A microprocessor?!!

            • Klimax
            • 6 years ago

            No, android…

          • Srsly_Bro
          • 6 years ago

          It could be possible Neely him/herself doesn’t know.

    • jjj
    • 6 years ago

    You should be clearer that this is not the first ARM on 16nm tape out..they did just a A57 in april last year.

    • ronch
    • 6 years ago

    Ok, so theoretically, will chips using this process tech be comparable in size, switching speed and energy efficiency to the same chips theoretically built using Intel’s upcoming 14nm node?

      • chuckula
      • 6 years ago

      No. More like 22nm given their own advertising numbers. The 20nm process that we will hopefully see in the second-half of this year is still planar. The “16nm” process doesn’t do much to improve transistor density beyond 20nm but does introduce finfets, which should have a positive impact on power characteristics.

      • Duct Tape Dude
      • 6 years ago

      I would think so.
      Differences between 16nm and 14nm using the same techniques (ie: FinFETs) will boil down to architectural efficiencies rather than manufacturing process. As always, Intel also has a time advantage, meaning they probably have tuned their own FinFET power consumption/performance more finely.

      I don’t know what materials Intel or TSMC are using, so perhaps there are some additional minutia, but for the most part they should be directly comparable.

        • MadManOriginal
        • 6 years ago

        The nm designations of different processes are no longer a useful direct comparison. One company’s 16nm is another’s 22nm if you were to try to compare them in the traditional ways, but even then those comparisons are becoming harder. I think processes should have codenames instead of feature size names going forward given that feature size isn’t an especially useful number any more.

          • Liron
          • 6 years ago

          They kind of -are- using codenames, since the feature size of TSMC’s 16nm is mostly 20nm.

    • chuckula
    • 6 years ago

    Yeah.. to everyone who is predicting how Intel can’t get 14nm working, it looks like the “16nm” process development has just reached the stage where Intel was with 14nm… in 2012 or so.

Pin It on Pinterest

Share This