Report: TSMC set to fabricate Volta and Centriq on 12-nm process

We're still gawping at the massive performance of the GeForce GTX 1080 Ti, but here's a nugget of news about Nvidia's next big thing. According to a report from Taiwan's Commercial Times, TSMC will be fabricating Nvidia's upcoming Volta GPUs on its new "12nm" process. Currently, TSMC's most efficient process is called "16nm FFC", and given that it debuted back in April 2015, it's about time for an update.

We hadn't heard about the new process before, but it's no surprise that TSMC is moving forward, seeing as its competitors are already manufacturing 10-nm parts. Of course, the nomenclature for process technologies is a bit fuzzy on all sides, so take those feature-size numbers with a grain of salt.

Unfortunately, the report in Commercial Times doesn't actually offer new information about the Volta GPU. We still don't know much about the new parts besides the fact that they'll apparently use of stacked memory. Since the top-end Pascal parts are already using HBM2, that's hardly news. Nvidia has offered a vague "2018" launch window for Volta, and given the company's history we'd expect a full range of GPUs based on the new architecture.

The Commercial Times report also mentions that TSMC will be fabricating Qualcomm's Centriq 2400 ARM-based server processors on the same 12-nm process, so between the two bits of news, it appears that the new process may be tuned for high-performance parts rather than low-power designs.

Comments closed
    • lycium
    • 3 years ago

    What does “most efficient” process mean? Smallest feature size, lowest defect rate, most cost efficient, … ?

    Also, small typo: “apparently use of stacked memory”

      • DancinJack
      • 3 years ago

      Probably some measurement of perf/power(W).

    • Leader952
    • 3 years ago

    [quote<]Nvidia has offered a vague "2018" launch window for Volta[/quote<] Volta will available in 2017 not 2018. The 2018 number came for it being on 10nm but with Volta on 12nm it will be a 2017 release. The first 2017 Volta will be the BIG Volta for HPC much like the P100 Big Pascal was. NVIDIA Volta Allegedly Launching In 2017 On 12nm FinFET Technology [url<]http://wccftech.com/nvidia-volta-12nm-finfet[/url<] Speculation of NVIDIA Volta GPU Ramps Up in Anticipation of 2017 Debut [url<]https://www.top500.org/news/speculation-of-nvidia-volta-gpu-ramps-up-in-anticipation-of-2017-debut[/url<] [quote<]Much more of the Volta story should unfold in early May, during NVIDIA’s GPU Technology Conference (GTC), where the new architecture is expected to be introduced.[/quote<] NVIDIA 12nm FinFET Volta GPU Architecture Reportedly Replacing Pascal In 2017 [url<]http://hothardware.com/news/nvidia-12nm-finfet-volta-gpu-architecture-replacing-pascal-2017[/url<] Nvidia’s next-gen Volta GPU appears in the latest driver update [url<]https://www.pcgamesn.com/nvidia/nvidia-volta-gpu-specifications[/url<] Nvidia Volta GPU's Upcoming Debut Raises More Speculations [url<]http://www.itechpost.com/articles/79457/20170201/nvidia-volta-gpus-upcoming-debut-raises-more-speculations.htm[/url<]

      • NoOne ButMe
      • 3 years ago

      Probably professional 4Q 2017, and consumer cards 4Q2017-2Q2018 depending on how competitive AMD’s VEGA is against PascalZ

    • TwoEars
    • 3 years ago

    We’re most likely looking at a “Pascal Refresh” this summer, but still on 16nm. This will likely include a 2080 Ti card as well to keep the red team in check. Then next summer we’ll see Volta at 12nm if all goes well. Meanwhile the red team is looking at lunching Vega (big and small) this summer on 14nm. But of course – size isn’t everything, even if smaller is better. Current rumors indicate that “Big Vega” will be somewhere between a 1080 and 1080 Ti card in performance, so a lot will come down to pricing. Just like it did during the last AMD GPU launch. And of course a lot could change with drivers etc. AMD Vulkan is a factor, but not every developer will now how to use it. Or will want to spend money and time learning how to use it. Then there’s still the whole freesync vs g-sync debacle which people will have to budget for if they go one way or the other. Should be an interesting summer.

    Here’s a good summary on Vulkan for developers btw: [url<]http://developer.x-plane.com/2016/03/what-vulkan-means-to-developers/[/url<]

    • Ninjitsu
    • 3 years ago

    Interesting, so nothing new, or even an update from Nvidia this year? Maybe a repeat of Kepler 2.0?

      • Voldenuit
      • 3 years ago

      There’s the 11 GT/s RAM coming to 1080, and faster RAM for 1060, but it looks like nv is sitting pretty until Vega comes out, and possibly even after.

    • chuckula
    • 3 years ago

    [quote<]The Commercial Times report also mentions that TSMC will be fabricating Qualcomm's Centriq 2400 ARM-based server processors on the same 12-nm process, [/quote<] If true then so much for this hypeline: [url<]https://www.qualcomm.com/news/onq/2016/12/07/meet-qualcomm-centriq-2400-worlds-first-10-nanometer-server-processor[/url<]

      • DancinJack
      • 3 years ago

      That’s strange.

        • chuckula
        • 3 years ago

        If Qualcomm is to be believed, then they aren’t fabbing the Centriq 2400 with TSMC at all (meaning Samsung is doing it), meaning this story got that massively wrong.

        Otherwise, if TSMC is fabbing the chip, it certainly ain’t “10nm” if it is launching this year.

          • bhtooefr
          • 3 years ago

          Unless they’re dual sourcing…

          Apple A9, for instance, was fabbed on both Samsung 14LPE and TSMC 16FF.

            • NoOne ButMe
            • 3 years ago

            A9 used LPE and FF+.
            TEMC killed 16FF to pull on 16FF+.

            I high explains a great at deal of it’s advantage over Samsung. As LPP competes with 16FF+. Not LPE.

      • Tekin1
      • 3 years ago

      The original source Commercial Times actually said that the Qualcomm Centriq 2400 will be made on a 10nm process: [url<]https://ctee.com.tw/mobile/ViewCateNews.aspx?cateid=kjmd&newsid=143146[/url<] Digitimes English version of the report: [url<]http://www.digitimes.com/news/a20170313PB201.html[/url<] [quote<]Meanwhile, TSMC also landed earlier orders for the Centriq 2400 server chips from Qualcomm. TSMC has begun volume production of the Centriq 2400 chips, using a 10nm process, said the paper.[/quote<]

    • DancinJack
    • 3 years ago

    Awwww yeah, bring it on. Volta will probably be my last big GPU upgrade for a while.

    Hopefully we can get some good compute upgrades without having to buy Titan.

    • short_fuze
    • 3 years ago

    I’ma gerbil of the peon class. So is TSMC on par with Intel, or is Intel sand-bagging at this point? All the speculation of Intel’s holding back on process shrink because of 3D Xpoint etc. (and subsequent delays because of … who-knows-what) and then suddenly announcing new tech (conveniently coinciding with Ryzen, or so it seems to me) makes me wonder how much Intel is just toying with the market.

      • DancinJack
      • 3 years ago

      They’re just *different*.

      Intel’s 14nm process is still quite a bit denser than anything else anyone has out right now. Samsung 10nm and Intel 14nm are probably pretty dang close, but they’re also specialized for different stuff. Samsung produces smaller chips that run anywhere from 100MHz-3GHz ish and mostly in SoC form for smartphones/tablets. Intel makes big ol’ CPUs, with way more logic, and that clock quite a bit higher. It’s an even bigger difference when you think about TDP of said parts. We’re talking hundreds of watts for Intel stuff, and about four, yes four, for Exynos SoC’s.

      It’s just a different ballgame. Having said all that, Intel IS moving to 10nm fairly soon. At this point though I wouldn’t be worried about manufacturing process as much as what you can get out of final products. GF does seem to be behind a bit, but it’s not like Ryzen/RX 480 are crap at the end of the day.

        • frenchy2k1
        • 3 years ago

        GF is on Samsung’s 14nm so far.
        They’ve already announced that they will skip 10nm and target 7nm directly.

        The reason process is confusing lately is due to technology.
        On one hand, you have what is currently being used, deep UV (193nm lights) in immersion with multiple patterning. This is the tech that got us to 14nm and can be pushed (with more patterning) to 10nm or even 7nm. The down side of it is that each patterning adds steps and time to the process (a patterning is an imprint from the a mask, then clean up and etching). So, this is a known devil by now but growing more and more expensive. It is also isotropic, meaning that it allows those very dense lines only in one direction (orthogonal to the repeats)

        On the other hand, you have extreme UV. This is a brand new technology that can do 10nm circuits in any directions, but its problems are many. This is almost in the x-ray light and can only use mirrors (no lens), changing the whole process. The light sources are the main problem, lacking power so far.

        So, each manufacturer needs to choose which tech to implement for 10nm and beyond. Intel is edging their bet, but as each is very expensive, few can do that. GF is waiting for 7nm for this reason, letting its competitors debug the new process before jumping to extreme UV directly.

        TLDR: process is currently a very expensive mess. It will take a few years to stabilize.

        Also, intel is NOT moving to 10nm soon, their next gen consumer will still be on their 14nm process.

          • DancinJack
          • 3 years ago

          That’s not what my response, or his/her question was about.

          Anyway, your idea of “soon” is clearly different than mine. Just because Intel has another CPU on 14nm doesn’t mean 10nm won’t be out this year, or more likely next year. In terms of CPU manufacturing process life cycle, that’s soon IMO.

          • tsk
          • 3 years ago

          There’s gonna be a 10nm cannon lake 5w part out in Q4 of this year, guaranteed.

            • NoOne ButMe
            • 3 years ago

            Yes. And OEMs will have tens and tens, maybe hundreds, of units to sell.
            It’s a 2017 product like how Broadwell was a 2014 product.

      • DavidC1
      • 3 years ago

      They are all fudging, including Intel. Ryzen’s L3 cache is smaller than KBL’s L3. That tells me plenty. Intel’s real world implementations don’t live up to their density claims. Also even if your foundations and tools are the best you can’t expect the final result to be. IF, what Intel is saying is true, they’ve got best 1 year lead. Who cares though? It’s not like we get to compare TSMC Kabylake versus Intel Kabylake.

      As scaling slowly stops expect more of this fudging. Contrary to some beliefs, Intel and others are doing the best they can.

        • NoOne ButMe
        • 3 years ago

        Intel’s claims are true. Intel has the cheapest process. Intel has the densest process. Intel has the highest performance process. Et Cetera. Now, how often Intel claims to have the cheapest [i<]and[/i<] densest [i<]and[/i<] lowest power process at the same time? 😉 Anyways, quick research shows GPUs have approached to up to twice as dense on the same node as CPUs. Pascal (and Polaris) are in this range. Note moving to HDL for AMD got them around a 30% density increase over Kaveri.

        • Klimax
        • 3 years ago

        “14nm” by foundries use 20nm metal backend. You can try to have one of the most regular structures in CPU smaller, but it will cost you something. Like cache latency. (See review)

      • NoOne ButMe
      • 3 years ago

      For high performance, low power CPUs, Intel dominates.

      For pretty much everything else TSMC and CPA are nearly as good or better.

    • Hattig
    • 3 years ago

    16FFC is fairly recent, and aimed at lower-cost 16nm designs and mobile.
    Otherwise it’s 16FF+ which is what GPUs are using IIRC, and that is indeed from late 2015.

    TSMC’s 12nm appears to be 16nm FEOL and BEOL (16FF has 20nm BEOL).

      • tay
      • 3 years ago

      FEOL? BEOL? HAALP!!!

        • ArdWar
        • 3 years ago

        Front End of Line: Semiconductor layer
        Back End of Line: Interconnect layers

Pin It on Pinterest

Share This