You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!

  • If an Intel 10nm transistor fell in the ARM forest

    Intelís news at IDF this week about partnering with ARM for foundry services on 10nm set off some wild speculation. Itís not a surprise that ARM would enable Intel Ė theyíve worked together before, ARM is an equal opportunity ecosystem partner, and ARM has publicly announced 10nm cores taped out at TSMC. Of more interest is what is actually could be in this for Intel.

    Mark Bohr, Intelís manufacturing guru, issued pretty much the same pitch heís used for a couple of years this week: our 10nm process is obviously superior to anyone elseís, based primarily on a 54nm gate pitch claimed to be several years ahead that translates to a density advantage and a lower cost per transistor. Thatís important when you need billions of transistors to get something done, like the next-generation Cannonlake microarchitecture.

    Article: Fabless Semiconductor Ecosystem Update 2012-intel-cost-per-transistor-10nm.jpg

    Why hasnít that obvious advantage translated to anything in volume but Intel Architecture processors? Zane Ball, head of Intel Custom Foundry, listed off their customers: Achronix and Netronome at 22nm, Altera and Spreadtrum designing at 14nm, and oh, wow, LG signed up for mobile parts at 10nm.

    The media outlets went nuts.

    NOW INTEL IS COMPETING FOR APPLE AND QUALCOMM BUSINESS!!
    Slow your roll, there. Last fall there were reports about the LG Nuclun 2 SoC, in a shoot off between TSMC 16nm and Intel 14nm, and Intelís benchmarks were superior. What happened? TSMC took all that business. By some reports Intelís yield on 14nm wasnít enough to meet LG demand, by others Intel pulled back indicating they needed scarce 14nm wafers for their own parts. Iím sure the pitch to LG now was weíll have plenty of wafers at 10nm, come on back.

    Why would LG not stay at TSMC? The answer is they are probably getting squeezed, either not meeting their volume commitments on the existing 16nm node (Nuclun 2 didnít make it into the G5) or unable to get any allocation on a more advanced TSMC node (read: Apple A10 and A11 all coming from TSMC). You can say youíre in the ARM mobile SoC design game, but to win you have to find and keep a fab partner filled. LG is too big to be small, and too small to be big. Intel and LG is kind of a match made via mobile dating app, after both experienced swipe lefts.

    Plenty of wafers, huh? In the early days of the PowerPC processor I persuaded my management team to take an outrageously aggressive high volume single-board computer deal knowing full well weíd only survive with that customer about 2 years, but it would up our supply chain volumes and improve margins on all of our other products. It worked, and we lasted 3 years supplying that customer profitably. Iím sure the pressure is on Intel Custom Foundry to find homes for 10nm, now, to prove out the cost/transistor story and prime the volume pump. It sounds silly they need to do that, but in the new PC reality the old Intel learning curves may not work without some help.

    Oh, but now Intel can really compete for Apple business. A couple analysts indicate there is some slim chance of that starting in about 2019. In theory, Intel could take a page from the Samsung playbook and offer Apple a smoking bundled deal on chips for Macs and iPhones and watches and anything else, making a volume manufacturing commitment. We know Tim Cook is a supply chain guy, and that Apple periodically swings the pendulum between full commit and dual sourcing just to keep everyone honest. There is also the recurring dream that ARM parts could displace Intel parts in the Mac, or that the iPad Pro just kind of takes over the Mac laptop line Ė at least Intel fabbing ARM parts leaves the door open, never say never, but as long as Intel Architecture retains its relentless improvement rate Iíd say the Mac business doesnít move.

    And, thereís the Qualcomm rumor. Hire Murthy Renduchintala and suddenly everyone believes there is a Trojan horse in play for Intel. Folks donít remember that Intel had Qualcommís mobile chipset business from 1995-1998, running 6 million chips until they decided upgrading Qualcomm from a 186 core to a 386 core was too big a risk Ė forcing Qualcomm to head for ARM. Iíd guess the organizational memory at Qualcomm hasnít forgotten that. Right now, per our own Daniel Nenni, Qualcomm is firmly entrenched with Samsung at 14nm and 10nm. Besides, Samsung and TSMC will just keep upping the process ante and take away any big gaps. TSMC in particular will be pushing hard for 7nm, probably reclaiming Qualcomm business at that point.

    Whatís left? My theory all along has been Intel could blow the industry away in IoT manufacturing volume if they can generate enough design wins, and that argument gets much stronger with Intel able to ship ARM parts from a foundry operation. Unfortunately, itís not in Intelís DNA to service a lot of smaller customers Ė their model is 10 big customers. Thereís also the problem that most IoT business doesnít need 10nm. That said, the model of 5 small-to-medium customers doesnít work at all for Intel.

    If ARM Ė or more accurately, their ecosystem partners Ė can compete with Intel for infrastructure business, Intel can and should compete for IoT edge business. Intelís low power story has always been suspect; their strengths are better in the gateway and big edge devices such as embedded vision. Canít compete with your other division? Nonsense. I was the guy who brought Intel into Motorola and we sold both architectures side by side, clearly delineating the value proposition for either choice, and we grew both lines. Itís called strategic and product marketing.

    What would need to change is the Intel Custom Foundry business would need to find a way to work with ARMís DesignStart program and its ecosystem. Whether that means multi-project wafers or compartmentalized access and support for selected IoT foundry services, I think the support problem for smaller customers is solvable. Creating the IoT value proposition for ARM on Intel 10nm might be the real challenge. GLOBALFOUNDRIES is putting a lot of energy into doing that on 22FDX, so itís not out of the question an advanced node could service IoT designs if the design cost, not just the manufacturing cost, is low. Or, Intel has to shift off the advanced node story and open up other fab space on more mature processes to foundry business.

    I would not be surprised to see something very creative, very un-Intel like, reshaping the Intel Custom Foundry business to make a jump into being a serious IoT contender now that the ARM hurdle has been cleared. Intelís IoT group might even return to designing ARM-based IoT chips Ė StrongARM didnít die because of the technology, it faded because they couldnít figure out a sustainable go-to-market model (see the 10 customers comment above). No offense intended to LG, but fighting for mobile scraps has almost no upside, itís a short-term tactic. If they stay on this course, a year from now weíll be reading the same stories about Intel struggling to be a foundry.

    Fighting for an overall Intel IoT strategy with both ARM and Intel architecture, and winning, would be worth the risk.