Signs point to GlobalFoundries making AMD’s 28-nm APUs

After publishing yesterday’s story about TSMC possibly manufacturing AMD’s 28-nm APUs Krishna and Wichita, an interesting e-mail appeared in our inbox. The e-mail was from GlobalFoundries, and it explained that, while AMD hasn’t made its 28-nm manufacturing plans official just yet, AMD executives have “consistently reaffirmed their commitment to Gate First HKMG.” An exchange in AMD’s Financial Analyst Day question-and-answer session, which I’ve transcribed below, was cited as evidence:

Dirk Meyer: And Chekib… one of the clarifying questions… I did say we would have initial revenue shipments of 28-nm products before the end of next year. Chekib, do you want to put some color on the 28-nm node?
Chekib Akrout: Sure. So, first of all, the answer is yes. We are going to use the high-k metal gate gate-first for 28. 28 nm is really a derivative; it’s a half-a-node move from 32, and we are happy about what we’ve seen in the high-k metal gate gate-first at 32, including, by the way, what we’re expecting from [a] performance perspective.

GlobalFoundries advertises its 28-nm HKMG gate-first process right on its website. TSMC, meanwhile, stated last year, “We developed a gate-last approach for TSMC’s 28nm high-k metal gate family.”

I guess that settles that. Remember, kids: always take your rumors with a pinch of salt.

Comments closed
    • spigzone
    • 9 years ago

    Theotard strikes again.

    This rumor was maybe started and definitely amplified by Theo the Retard at BSN and it sprung from a roadmap slide that had AMD’s two foundries logos across the top, with Globalfoundries first and TSMC second. SomehowTheotard concluded from this TSMC was going to manufacture the 2nd ‘half’ of the roadmap.

    §[<http://www.brightsideofnews.com/news/2010/11/10/amd-moves-cpu-architecture-into-tick-tock-mode3b-schedules-20nm2c-14nm-apus.aspx<]§ My BSN site comment at the time: >> Spigzone 6 days ago "Note that products based on 22nm and 16nm processes by GlobalFoundries are MIA" Based on what? ... the placement of the TSMC logo over the right half of the chart ... ??? << Not to excuse TTR, as this rumor was easily disproven.

    • esterhasz
    • 9 years ago

    Maybe you should send an email to the people at GP out of charity – perhaps nobody told them.

    No, but honestly, you’re right, gate first *does* have important risks and issues to solve but the available workarounds to the thermal problems are not looking completely crazy either. The question is whether the additional cost for the tricks that can lower threshold voltage offsets the savings in die size and design cost. This remains to be seen. But there should be less problems with a > 2Ghz low voltage part like the Bobcat. Llano may be a different question though.

    edit: should have been a reply to #3, me no clicky good…

      • NeelyCam
      • 9 years ago

      GF is stuck with IBM – they don’t really have the capability to develop the process without IBM, and IBM says ‘gate first’.

      Meanwhile, TSMC has been in process development by themselves for years. Both TSMC and Intel (gate last proponents) have been hugely successful. AMD doesn’t have to go with the good’ol fabs they used to own – now they have a choice.

      I understand the allure of gate first – easy design rules and potentially tighter layout obviously sound great. But if it comes with yield issues, increased variability, higher Vt etc., it might not be worth it. Also, AFAIK, gate last should enable better strain engineering, resulting in better transistor performance.

      If GF can pull it off and save a few bucks, of course that would be awesome for them. I think they are taking a big risk, though, betting on something that the top two CPU/GPU manufacturers have decided is a bad idea.

        • esterhasz
        • 9 years ago

        Well, put this way, I mostly agree. There are risks. But the worst case scenario is probably that the potential benefits of gate-first don’t materialize and that the whole thing ends up more expensive than a gate last approach would have been. Going with a slow / low voltage part first will surely help though.

        It’s a tough choice, the problem with decisions of that kind is that you cannot know whether it’s going to work before you do it. Sure, you can make all kinds of projections based on established knowledge + projective modeling but complex manufacturing processes have way too much fickle in them to be controlled this way. When GF made their decision, they had to be confident that a gate first strategy will produce certain thermal characteristics for a certain price, and part of that is “gut” (as the contrary decision would be) because in order to know, you have to do it…

          • NeelyCam
          • 9 years ago

          Yep. This stuff is hard; everybody makes bets, and hope they pay off.

    • willmore
    • 9 years ago

    This seems an odd choice at first, but when you look at GFs 28nm process, you see that there are really three of them–SLP, HP, and HPP.

    The first of them makes complete sense. It could allow AMD to produce a small die size very low power device that could completely overshadow even the MID destined atom processors. If the design rules are similar enough, a quick recompile of the processor and you’ve got a much higer performance part–good for notebooks and low end desktops.

    This is assuming that Krishna and Wichita are going to be synthesized cores like their predacessors.

    • NeelyCam
    • 9 years ago

    OMG is AMD really going to make this mistake?

    Gate first = DOA. Good luck.

      • Hattig
      • 9 years ago

      Swallowed the Intel spin hook line and sinker eh?

        • NeelyCam
        • 9 years ago

        Yes. I’d say Intel can be considered to be a bit better in silicon manufacturing than Global Foundries… wouldn’t you agree? If Intel says gate first is a mistake, there’s a good chance that this is the case.

    • kamikaziechameleon
    • 9 years ago

    Here is to hoping that this shapes up to be a good partnership.

    • BoBzeBuilder
    • 9 years ago

    Deja Vu.

Pin It on Pinterest

Share This