|Last week, we paid a visit to ARM's headquarters in Cambridge, England and sat down with the company for multiple deep dives into its structure, processor architecture, and the future of its product design. The semiconductor market for mobile and hand-held devices has changed dramatically in the past six years and ARM has had to evolve along side it. This is the first in a series of articles designed to profile different aspects of the company and its competition with Intel.
ARM's Licensing and Design Model
Most readers are aware that ARM has a very different business model than Intel. Specifically, ARM licenses a wide range of technologies in a vast number of markets. The majority of ARM's customers license specific cores -- a Cortex-A9 for one SoC, a Cortex-A15 for another, etc. The company's license model is extremely flexible -- companies can license the ARM architecture without buying a specific implementation of a chip (Qualcomm, X-Gene, others follow this route). Manufacturers can opt to license a specific core for one implementation, buy access to a chip for a period of time, or can subscribe to everything ARM has, with no need to hack out a new license for each and every product.
ARM's IP focus allows it to dedicate all its resources to building a great design rather than committing to any single process node, customer, or foundry. Architectural design and implementation is done very much in partnership with both foundries (TSMC, GlobalFoundries) and licensees like Samsung or Qualcomm. This has only grown more important in recent years. GlobalFoundries made the decision to deploy gate-first manufacturing technology at 28nm, while TSMC went with gate-last technology and multiple types of 28nm, from conventional polysilicon to high-k metal gate. ARM's goal is to create the necessary IP to enable as many manufacturing technologies as is possible.
This flexibility is one of ARM's major assets. You can build an ARM processor at multiple foundries on multiple nodes and the company works to keep architectures up-to-date. The first Cortex-A9 processor shipped on 65nm but has been ported to 28nm by at least one SoC vendor, RockChip. Comments ARM made to us in Cambridge hint that we may see 20nm variations of the processor as well.
Keep in mind that while a handful of companies command a great deal of attention from the media for aggressively pushing the semiconductor envelope, they account for less than half of the total chips sold in any given year. TSMC still earns hundreds of millions of dollars yearly on process nodes that debuted ten years ago. ARM's business model is designed to cater to companies like Samsung, Nvidia, and Qualcomm with fast refresh cycles and leading-edge node adoption, while also fitting the needs of firms with 5-7 year refresh cycles. If you're a company that built an ARM11 core in 2006 and you're eyeing a Cortex-A5 as a major upgrade this year, ARM has cores to fit your needs.
|Compared to Intel|
|The difference between Intel and ARM is more nuanced than the simple ownership of manufacturing facilities. Owning its own means of production means that Intel can tweak process technology to match the particulars of a given architecture (and vice-versa). It also gives the company far more flexibility when planning future nodes. If Intel feels that integrating Peanut Butter Silicon on Insulator (PB-SOI) is the best way to hit its performance and power consumption targets at 14nm, for example, it can make that happen internally. ARM, in contrast, is limited by the decisions of the foundry manufacturers it partners with. If TSMC and GlobalFoundries decide that PB-SOI is a bad bet at 14nm and that they'd be better off sticking with FinFET, ARM can't do much to change that.
Owning its own foundries benefits Intel in another way too -- there's no need for licensing debates or foundry payment agreements. ARM has said that it can take months to hammer out a licensing agreement with a company, particularly if the agreement is a complex multi-year arrangement. AMD's troubles with GlobalFoundries since 2011 have illustrated how yield problems and uncompetitive parts can hamper both sides of an agreement; Sunnyvale paid GF some $750M last year for the right to build Kabini and Temash at TSMC.
ARM would argue that Intel's manufacturing advantage, meanwhile, is an advantage for Intel as opposed to Intel and its customers. Again, the company has a point. Dell, HP, and the other PC OEMs are partly to blame for the miserable shape of their own margins (as low as 2%), but there's no denying that Intel's price structures and market dominance have played a part. It's a bitter pill for the PC manufacturers to swallow, particularly when Intel's margins have been above 60%, as they were before the bottom dropped out of the PC market.
Intel would argue that its manufacturing expertise has been instrumental in driving computer prices lower and enabling new form factors in desktop and mobile. Its ultimate responsibility is to consumers not to competitors. While it offers less flexibility and fewer architectures in an absolute sense, Santa Clara claims that superior technology allow it to target a wider range of usage scenarios with a much smaller number of chip designs.
Why doesn't Intel have a microprocessor even smaller than Atom? (Quark? Gluon?) Because it doesn't need to. Because it can address the markets that ARM would serve with a different processor design with a single chip, thanks to superior manufacturing and process technology.
|Subtly Shifting Strategies On Both Sides|
Which company has the "better" model depends entirely on what you're looking to build, how much control you want over the end product, and whether you believe ARM or Intel is building better chips to get you there in the long term. It's worth mentioning, however, that both Intel and ARM have taken pages from each others' playbook. Intel has done some limited fab work for other companies and some custom design work for major companies like Facebook, Google, and Amazon.
Analysts often spin Intel's limited foundry work into stories about how the semiconductor giant will throw open the doors and fab for all-comers -- forgetting in the process that Intel views its process technology advantage as its primary advantage over companies like Qualcomm, Samsung, and AMD. There's no chance at all that Intel would simply become a design firm at this stage of the game -- but the company has made changes to its core business in the name of building more flexible products.
ARM has no plans to launch a foundry, but the changes to the company's designs in the past six years can scarcely be understated. The Cortex-A8, A9, and A15 all pushed the performance and performance per watt envelopes. The Cortex-A5 and A7 improved performance at the lower end of the market but focus more on ensuring low power consumption. ARM has fielded more specific implementations in the past few years than in the decade prior and the company isn't slowing down -- the Cortex-A12, A53, and A57 are all set to hit market within the next 18 months.
Meanwhile, ARM has deepened its alliance with the foundries themselves rather than relying on its partners to carry the day. Multiple announcements on the company's timeline highlight joint agreements with TSMC and GlobalFoundries to optimize products for particular processes. This was likely an essential development, given that pace of adoption for new nodes has slowed sharply while the advantages of adopting them have declined.
It's not clear whether ARM's design focus or Intel's superior manufacturing will lead to better outcomes for the mobile market, but neither company is rushing to copy the other. Instead, both are implementing next-generation capabilities in a way that plays to their respective strengths.