Some of AMD’s next chips will arrive on GloFo’s new 12LP process

AMD's Mark Papermaster announced today that the CPU and GPU vendor will be "transitioning graphics and client products" to a new "12nm" process at GlobalFoundries in 2018. The announcement came today at the GlobalFoundries Technology Conference in mostly-cloudy Santa Clara. GlobalFoundries calls the new process 12LP, and as usual, the LP stands for "Leading Performance."

Image: Patrick Moorhead, on Twitter

Tom's Hardware followed up with Papermaster after the conference and confirmed that both Vega GPUs and a further iteration of Ryzen CPUs will be produced by GloFo on the new 12LP process. According to past rumors, AMD had intended to ship the next iteration of the Zen core on a 7nm process sometime in 2018. It's not clear whether Papermaster was talking about a possible Ryzen refresh of sorts on the 12LP process, or if the so-called "Zen 2" has been pulled back to 12nm from its purported 7nm aims.

GlobalFoundries says the new process can provide up to a 15% improvement in circuit density over its own 14LPP technology, implying that it's a bit more involved than a simple refinement. Although node names are basically hopes and dreams at this point, GloFo's 12LP should at least reflect a real advance. Not long ago, Samsung announced its own "11nm" process, and as far back as March we heard whispers of TSMC's "12nm" process, so in a sense GlobalFoundries is playing catch-up here. However, this move from GloFo is a little surprising given that the company specifically said it was skipping over any intermediate steps to move directly to its 7nm process.

Comments closed
    • renz496
    • 2 years ago

    Isn’t that AMD already deny about the AI partnership with tesla?

    [url<]https://www.streetinsider.com/dr/news.php?id=13317752[/url<]

    • BaronMatrix
    • 2 years ago

    This should help a lot with Vega’s power issues (balanced my butt)… It may also mean a 4.5GHz Ryzen 7…

    And it is reasonable since the next Zeppelin die isn’t due until Q318 for Rome… And it’s just a matter of packaging from there… That’s the beauty of the MCM approach… You only need one base die…

    • Pville_Piper
    • 2 years ago

    Interesting that they announced a next gen chip when they can’t even get the current chip to their board partners…

      • Pville_Piper
      • 2 years ago

      [url<]https://www.tweaktown.com/news/59213/amd-partners-still-vega-gpus-custom-cards/index.html[/url<] Maybe late October? Sad, really sad...

    • PrincipalSkinner
    • 2 years ago

    So can we expect that Vega on 12nm will be just a die shrink with no meaningful performance improvements?
    It sounds like R600 all over again. Not that it’s successor was bad, but that was midrange.

      • willmore
      • 2 years ago

      No, if they shrunk it, we would expect some power and performance improvements. Very modest one, but it would be unreasonable to expect it to be identical to the existing parts.

    • willmore
    • 2 years ago

    [quote<]It's not clear whether Papermaster was talking about a possible Ryzen refresh of sorts on the 12LP process, or if the so-called "Zen 2" has been pulled back to 12nm from its purported 7nm aims.[/quote<] Not to be a dead horse, but they could be doing the classic new architecture on old process, then same architecture with a dumb shrink to a new process, then new arch tuned to the new process, lather, rinse, repeat. The result of that would be an existing Zen1 on 12LP only to see Zen2 show up on it a bit later. The real benefit to this is efficiency in design as it decouples your archtiecture and implementation teams. Because the new arch is targeted to an *esisting* but not wonderfully tuned process, there are fewer dependencies on fab progress stalling chip release. On the flip side, the fab people can use the old arch to try out the new process. Nobody has to be dependent on anything new or untested. This tendency to put all of your eggs in one basket hasn't worked all that great for AMD in the past.

      • chuckula
      • 2 years ago

      [quote<]Not to be a dead horse, but they could be doing the classic new architecture on old process, then same architecture with a dumb shrink to a new process, then new arch tuned to the new process, lather, rinse, repeat.[/quote<] That's a fascinating idea. It sounds like it works like clockwork. In fact, maybe they could call it a Tick-Tock process. BRILLIANT!

        • willmore
        • 2 years ago

        *sigh* it was done this way well before Intel came up with a marketing term for it. But, since they’re no longer using it, maybe they won’t mind if AMD uses it.

          • JustAnEngineer
          • 2 years ago

          Intel went from Tick-Tock to PAO to POOO…

    • AMDisDEC
    • 2 years ago

    AMD has announced some very significant design-ins with their new processors.
    Working with Telsa on their new AI chip is a great design win over NVIDIA and their design in of Epyc in China’s new exascale system and if China manages to hold their lead, may see AMD Epyc processors used in the world’s fastest HPC system.
    Both Telsa and China will probably want products based on the new process.

      • chuckula
      • 2 years ago

      Here’s the airplane glue.

      Tell me another one!

      Incidentally, that whole Tesla thing is yet another case of some kid at Wccftech hearing a word and jumping to a conclusion that’s not based in reality.

      Here’s a more fact-based take on what they really said: [url<]http://www.tomshardware.com/news/cpu-global-foundries-amd-tesla,35507.html[/url<]

        • AMDisDEC
        • 2 years ago

        Understanding how Musk feels about AI, especially AI beyond his understanding and control, the AMD alliance makes perfect sense.
        The only real question I have is will Musk create a scalable platform which can be used across his automotive, Aerospace and transportation product lines.
        I’m wondering if Tesla/AMD will integrate the AI engine into AMD APU.

        We’ll hear more about it as time goes on, but for now, the announcement has helped AMD stock prices rise another notch.
        [url<]https://www.cnbc.com/2017/09/20/tesla-building-an-ai-chip-for-its-cars-with-amd-globalfoundries.html[/url<]

        • Jigar
        • 2 years ago

        Elon Musk is neither confirming or denying this so thats something.

        • NTMBK
        • 2 years ago

        Impressive how you managed to get CNBC muddled up with WCCFTech. It’s almost like you’re desperately trying to discredit this for some reason.

          • chuckula
          • 2 years ago

          Impressive how you misinterpreted third-hand noise from a cable channel… that AMD’s PR has literally officially denied: [url<]https://www.streetinsider.com/dr/news.php?id=13317752[/url<] Oh, and GloFo denied it too: [url<]https://www.reuters.com/article/us-tesla-chips/globalfoundries-says-no-commitment-from-tesla-on-chip-deal-idUSKCN1BW259[/url<]

            • NTMBK
            • 2 years ago

            Please point to where I misinterpreted it 🙂 It was a rumour from an anonymous source, it was always in “pinch of salt” territory. You’re the one who utterly misrepresented its origin.

            • AMDisDEC
            • 2 years ago

            I think it’s too early to verify the results of the collaboration and we’ll hear more if/when the results are favorable.
            For now the rumor alone has given me enough to buy that new Maxima I’ve been eying.

        • freebird
        • 2 years ago

        Why link to Tom’s Hardware? Go right to the source of the article and everyone can draw their own conclusions…

        [url<]https://www.cnbc.com/2017/09/20/tesla-building-an-ai-chip-for-its-cars-with-amd-globalfoundries.html[/url<]

      • BaronMatrix
      • 2 years ago

      At this point AMD just needs a full speed DP engine… The rumor is Vega 20 will have it and Navi may expand on that… It would be cool if they did two designs (VERY expensive at 7nm)… One for AI/gaming and then an HPC only version without geometry engines or ROPs or TMUs… They say Navi will be scalable which implies modular…

        • AMDisDEC
        • 2 years ago

        It will be interesting to observe how this develops and what the architecture really looks like.
        As this EETimes article states, Tesla’s market share isn’t large, so won’t make an impact on AMD immediate revenue, however Tesla is a very good partner for AMD as it incorporates machine learning into future product lines.
        I wouldn’t be surprised if AMD embeds the AI tech into future X-Box and Playstation game console hardware. This Dr. Lisa Su is truly a gem.
        [url<]http://www.eetimes.com/document.asp?doc_id=1332331&[/url<]

        • freebird
        • 2 years ago

        As more profits/revenue comes … it allows them to expand their offerings. I don’t see split GPU designs as you mentioned unless EPYC makes substantial gains and AMD generates significant revenue from GPU Compute/AI in the next couple of years. Maybe in the early 2020s AMD could “afford” separate high end GPU gaming chip designs from GPU Compute/AI designs, but chip design masks increase in cost with the ever shrinking lithograph technology. All of that said, the HBCC of Vega currently has no use in gaming and never may be utilized, until everyone has an SSD melded with the GPU card for unlimited “open world/virtual world” computer experiences, but those are still 5+ years away. By then Vega, will be lost in the “stars” of new GPUs.

          • msroadkill612
          • 2 years ago

          “the HBCC of Vega currently has no use in gaming and never may be utilized, until everyone has an SSD melded with the GPU card for unlimited “open world/virtual world” computer experiences, but those are still 5+ years away.”

          pure speculation. Not even good speculation – HBCC could well be fantastic for gaming.

          The “melded” ssd exists now on the radeon pro ssg, & the port is standard on all? vega architecture.

          With hbcc, a poor mans Vega SSG is available on any Vega, but using the pcie bus instead of superior Fabric to access an ssd array.

          Pretty soon, a triple (3x nvme ports are common on tr mobos) array of affordable`128GB nvme ssdS using tr’s plentiful lanes, should yield 384GB with raw bandwidth of 12GB/s read & 9GB/s write – similar to gpu<> system memory now. Many gamers say 8GB/s is fine for a gpu. Even 2/3 of the above speeds would serve well.

          The above is a little optimistic now, but only a little. The next model samsung ssd needs only a small bump to get to 4GB/s max of the bus.

          HBCC offers nothing less than ~unlimited, slower (but not nearly as slow as u think), memory. It takes all kinds, but personally I think thats exciting, and perhaps most popularly, to gamers.

      • ronch
      • 2 years ago

      Jim Keller now works for Tesla. You think that has anything to do with AMD winning them over?

        • AMDisDEC
        • 2 years ago

        “winning them over”?
        That sucking pacifier must be real old.

        Clearly you don’t comprehend “Business”.

          • ronch
          • 2 years ago

          And clearly you don’t comprehend civilized conversation.

            • AMDisDEC
            • 2 years ago

            Only when I hear/read it.
            You are confusing civilized, with Dotards.

            • ronch
            • 2 years ago

            Maybe, but then who often gets a ton of downthumbs around here? Can’t be fun talking with you. Thank goodness this isn’t actual face-to-face conversation.

            • AMDisDEC
            • 2 years ago

            Thumbsdown? LOL, you are just making my case for me.
            Does TR write you a check or give you a car or AMD stock for ThumbsUp?
            You Dotard!

            • ronch
            • 2 years ago

            Impressive.

            • AMDisDEC
            • 2 years ago

            More precisely. Out of sync with drone-goosestep.

      • Anonymous Coward
      • 2 years ago

      With this news, things are certainly looking up for AMD.

      They have a lot of debt to manage, but I don’t think any other company is in such a strong position to deliver semi-custom high performance parts. (Also the touch of Musk is not bad for reputation.)

      Over at nVidia they have very solid GPU and AI-specialized processing, but they are weaker on the CPU side (and probably system integration). Over at Intel, they have very solid CPU cores, but are weaker in GPU and AI-specialized processing. AMD is likely easier and cheaper to work with than either nVidia or Intel, to top it off.

      Aiming for 90% as good but with much better terms might take them a long ways.

        • AMDisDEC
        • 2 years ago

        Well said.
        The Sun is shining brightly on the face of AMD and this time around they have qualified management to soak up the rays.
        Musk is concerned about the AI and computational resources needed for the Space-X Mars mission as well as Tesla, The Boring, hyperloop, and Gigafactory. This could be a sweet source of R&D funding for AMD.
        The release states that Tesla will own the AI IP, but I wonder in what form, as subset or superset of what AMD develops.

        • msroadkill612
        • 2 years ago

        ” (Also the touch of Musk is not bad for reputation.)”

        Hmmm?

    • Mat3
    • 2 years ago

    Until this article, I had always just assumed that “LP” meant “Low Power”.

      • Wirko
      • 2 years ago

      For people old enough, “LP” means and will always mean something else. So much more when the number twelve is around.

        • smilingcrow
        • 2 years ago

        I bought over a thousand albums on vinyl but no longer think of LP in that context as I stopped buying them almost 30 years ago and sold my collection over 20 years ago.
        It seems like another lifetime having to deal with vinyl.

    • deruberhanyok
    • 2 years ago

    This would be an interesting move as a mid-cycle refresh. Like, before “Zen 2” they release “Zen +” which is mainly lower power use or hits higher clockspeeds. Same if it were used to improve on Vega’s power draw / heat for future Vega chips (whether they be refreshed versions of the big ones or used for smaller ones to succeed Polaris).

    • ronch
    • 2 years ago

    Could this be the answer to Ryzen’s clocking blues? Can’t wait to see how the next wave of Ryzen chips stack up.

      • chuckula
      • 2 years ago

      Sometimes I wonder what I’ma gonna do.
      Kuz there ain’t no cure for the Ryzen clock blues.

      • willmore
      • 2 years ago

      Will they adopt a classic “process then tune” cadance like the industry has used for a long time? I.E. will we be seeing better tuned libraries for the next gen of Zen?

        • cygnus1
        • 2 years ago

        “the industry” ? I think you just mean Intel.

          • willmore
          • 2 years ago

          No, DEC did this decades before Intel started doing it.

            • cygnus1
            • 2 years ago

            Ahh, didn’t realize that. I guess that does make it ‘classic’ as in old, but I don’t think it’s something other companies in the industry actually do that often or at the very least they don’t talk about their products in enough detail to discern if that’s what the process they’re following.

            • willmore
            • 2 years ago

            Not a lot, sure. You have to either own the fab and the design or at least have them in very close contact. So, outside of Samsung, Intel, AMD/GF, (historically) DEC, and (historically) HP, not many got a chance to do it.

    • ronch
    • 2 years ago

    Mark Papermaster!! How long has it been since we kept in touch?!

    Anyway, 12nm!! Intel is doooommmed!!

    Gotta catch up, Mr. Other Mark from a certain other chip company!! (That would be Mark Bohr from Intel. – Eddie)

    • chuckula
    • 2 years ago

    It’s interesting that they advertise a 15% improvement in circuit density over 16nm “industry” solutions since GloFo never used a “16nm” process but instead had “14LPP” as stated on the slide. We’ll see if it’s more than 15% or less than 15% better.

    Additionally, there were rumors that AMD GPUs were jumping to TSMC next year. These statements seem to cast doubt on those rumors.

      • jihadjoe
      • 2 years ago

      Funny thing is TSMC’s 16nm has higher density than GloFo’s 14.

        • chuckula
        • 2 years ago

        I’ve typically heard the opposite. [Edit: And Anandtech has official numbers that agree with me: [url<]https://www.anandtech.com/show/11862/globalfoundries-weds-finfet-and-soi-in-14hp-process-tech-for-ibm-z14-cpus[/url<]] At least that was the story when Polaris launched.

          • Beahmont
          • 2 years ago

          See, I’ve heard that GLFO had the better density, but worse switching speed than TSMC. The difference between density and switching speed of the two processes was still supposed to favor TSMC because the density differences were relatively minor, but the difference in switching speed of the transistors was very large.

      • Mat3
      • 2 years ago

      Strange comparison to make on the slide. I want to know how it stacks up against GloFo’s own 14nm process.

        • freebird
        • 2 years ago

        If you read the actual GF press release, they state those performance improvement figures compared to industrial wide 16/14nm offerings.

        “The new 12LP technology provides as much as a 15 percent improvement in circuit density and more than a 10 percent improvement in performance over 16/14nm FinFET solutions on the market today. This positions 12LP to be fully competitive with other 12nm FinFET foundry offerings. The technology leverages GF’s expertise at Fab 8 in Saratoga County, N.Y., where its 14nm FinFET platform has been in high-volume production since early 2016.”

        So, not sure why the slide differed, but…

        Maybe slide was made earlier and they just didn’t change it to match, an intern did it or 16 didn’t sound good enough in the press release and they changed it to 16/14nm in the press release.

        or maybe this was just a planned 14nm+ process, but GF thought they could make more sales/charge more money by calling it 12nm to compete with other foundries with a “12nm” process.

      • msroadkill612
      • 2 years ago

      They well have had contractual obligations resulting from the fab sale that they would have to buy GF out of if they switched.

      • tygrus
      • 2 years ago

      16nm direct shrink to 14nm (all else being equal) = 30% improvement in density and maybe 10% performance @same power. 14nm to 12nm should be 36% and another 10% perf. They should spend more money on R&D instead of marketing BS.

    • Hattig
    • 2 years ago

    It’s using a 7.5T library instead of a 9T library to get the density improvements.

    I’m guessing that the methodology is going to be extremely close to Samsung’s “11nm” process.

    So 12nm products mid-2018, and 7nm mid-2019 feels like the roadmap at the moment.

      • freebird
      • 2 years ago

      Yeah, I agree with you, I was initially hoping for something to “appear” CPU or GPU using 7nm by Jan 2019, but with this it now looks like 7nm won’t be here until mid 2019 at earliest…unless they do a high-end Navi with associated Professional Level pricing.

    • ptsant
    • 2 years ago

    If Ryzen 2xxx is 10-20% faster in freq, without any other change it all, it would be immediately much more competitive with Intel in single-threaded. If they also manage to fit AVX512, then there will be almost no meaningful difference with high-end Intel.

    If rumors of a Q1 2018 launch are true the pressure on Intel will be huge.

    I might even be tempted to do an in-socket upgrade…

      • chuckula
      • 2 years ago

      [quote<]If rumors of a Q1 2018 launch are true the pressure on Intel will be huge. [/quote<] I don't know where those rumors came from but if the picture in this article is true then they've been totally shot down. The slide expressly says "Risk production in 1H 2018". "Risk production" occurs well before you have a launched product. I would say that if the 'risk production' proceeds smoothly you'd expect to see commercial products by late summer or early fall of 2018.

        • ptsant
        • 2 years ago

        Interesting. I have no idea what “risk production” means, but some people in other sites have been very optimistic based on the 1H 2018 figure.

        Anyway, even summer 2018 would be nice. 2019 would be too much, though.

          • chuckula
          • 2 years ago

          Well it’s a vague term but look at how the terms “risk production” and “high-volume production” are used within the confines of that single slide in this article.

          GloFo clearly states that their 14LPP process has been in “high-volume production” since early 2016. What was the very first product that AMD launched using that process? That would be the Rx 480 Polaris GPUs that TR reviewed on June 29, 2016. So there was a delay of multiple months from the “early 2016” launch of high-volume production to when you could actually buy a product.

          • Spunjji
          • 2 years ago

          “Risk production” usually means you /might/ get a working product out at low yields, but more likely you’re just going to be validating your architecture against their process and perhaps generating a few engineering samples for qualification purposes.

      • Anonymous Coward
      • 2 years ago

      Outsized ambitions will doom AMD, as they have before. I know its boring to accept the status quo, but AMD needs wisdom and not testosterone.

        • AMDisDEC
        • 2 years ago

        Ambition wasn’t AMD’s failure, but bad management and a terrible strategy made by those with low testosterone and imbalanced Dopamine to Serotonin levels.
        There’s a new sheriff in town!

          • msroadkill612
          • 2 years ago

          AMD was being run by the same types running intel now.

        • msroadkill612
        • 2 years ago

        They have sure been ambitious, and their success somehow makes their cheek seem worse, but to be fair, their odyssey also includes an incredibly pragmatic biz model, cast in the crucible of financial adversity.

        Big time success from such an ‘umble start point seems fraught, til you look at the details, and see just how scalable all is, from the products to production.

        Even distribution & marketing.

        In PCs, they let credible enthusiasts do their excellent marketing campaign free, keeping it spiced with an ongoing treasure hunt of teaser clues.

        In servers, it seems (they cant make enough epyc’s atm i hear, & its not even sold retail yet) if they can sell to the ~10 main cloud providers, they have much of the market and the rest will follow.

        Fabulously fabless, they have very flexible production, and a hival product to absorb the cost of queue jumping if needed.
        They seem to have negated much of the incumbents market power.

      • freebird
      • 2 years ago

      I doubt AMD will devote much Ryzen silicon to AVX512 (unless they feel they need to for the Server Market somehow). I believe that AMD views the GPU as a better source of “compute” services going forward (especially with their limited resources) and keeping the CPU as a more general purpose (x86 based) processor. CPU & GPU are really losing their descriptive nomenclature, perhaps, when you start putting MMX to AVX in CPUs and Compute/AI/Deep Learning features in GPUs, then they really become something else entirely, right?

        • msroadkill612
        • 2 years ago

        I hear intel’s fuss about avx is BS. If u need avx, u r doing it wrong. u should use way better gpu compute.

    • Helmore
    • 2 years ago

    I’m feeling more and more in favor of using different metric to express different process nodes. The use of talking about half-pitch sizes is starting to seem more meaningless every year.

      • DPete27
      • 2 years ago

      Worked well when you’re splitting trees. Not so much when you’re splitting hairs.

      • freebird
      • 2 years ago

      This article does a good job of comparing process technology by the different fabs, but only compares other Fabs 7nm tech to Intel 10nm tech.

      [url<]https://www.semiwiki.com/forum/content/6879-exclusive-globalfoundries-discloses-7nm-process-detail.html?s=b5434b9d52086fca5a21bc5c9098420b[/url<] Enjoy!

Pin It on Pinterest

Share This