Intel's annual meeting with analysts this week didn’t answer all of my questions--detailed product roadmaps seem to be a thing of the past for chipmakers. But we did learn a lot about where Intel is headed.
This week I wrote about five big questions for Intel ahead of its annual meeting with analysts. The Webcast didn’t answer all of my questions--detailed product roadmaps seem to be a things of the past--but we did learn a lot about where Intel is headed, including some interesting changes to its strategy as it expands into new markets.
The road ahead
First, the bad news. Intel did not announce any new products or provide much more detail on its product roadmap. After a delay, the first 14nm Broadwell processors, the Core M, are starting to show up in 2-in-1s. These hybrids are still a relatively small part of the market, but Intel said there are now some 70 2-in-1 designs available (more than 10 percent of the market for smaller systems) and prices are coming down. The higher-performance (and power) version of Broadwell, the 5th-generation Core processors, will start showing up in mainstream consumer and business laptops next spring.
In the second half of next year, Intel will start shipping Skylake, which is based on the same 14nm manufacturing process but has a new microarchitecture. Kirk Skaugen, who heads up the PC group, promised a fast transition to 14nm processors, but Intel did not provide any details on how it would roll out Broadwell and its successor, Skylake, so close to one another.
At the lower end of the market, the Bay Trail-M Atom processor, branded as Celeron or Pentium, has driven a lot of growth in Chromebooks and budget Windows notebooks—some of which now sell for as little as $200. This will be replaced, in the second half of 2015, with a processor code-named Braswell, which will be manufactured on a more advanced 14nm processor (Skaugen first mentioned Braswell in the spring at IDF14 Shenzhen).
In notebooks, Intel’s focus for 2015 will be on these entry-level systems; thinner and lighter mainstream systems with the 5th-generation Core processor; premium laptops using the new enthusiast processors and Iris Pro graphics; 2-in-1s for consumers and business users powered by Core M. And in desktops, the focus will be on entry-level systems, in particular for China; all-in-ones (including lower-priced models with 4K displays); mini PCs including a new Steambox; enthusiast systems using chips like the 4.0GHz Core i7-4790K and eight-core Core i7-5960X; and more innovative business boxes.
Intel did not provide an update on its server processor roadmap, but that’s not too surprising considering the company just rolled out its Haswell-based Xeon E5-2600 v3, known as Grantley, in September. Instead Diane Bryant, the head of the Data Center Group, talked about how the company is using both standard and custom versions of Grantley to meet customers' needs.
Changes in mobile strategy
Intel is making some progress in mobile but it is coming at a heavy cost. The company is on track to meet its goal of shipping processors for more than 40 million tablets this year, compared to about 12 million last year and only 1 million in 2012. The problem is that the 22nm Bay Trail-T Atom platform was not designed to compete head-to-head with low-cost ARM-based processors in Android tablets, so Intel is subsidizing the business to gain share.
“We are aware that we are losing a lot of money in mobile,” Chairman Andy Bryant said. “This is the price you pay for sitting on the sidelines for a number of years and then fighting your way back into the market.”
A lower-cost version of Bay Trail (along with the older Clover Trail Atoms) helped narrow the gap a bit, but the real solution will be SoFIA, which will match the cost of competing chips for Android tablets and smartphones. The first SoFIA processor, with a dual-core Atom and 3G modem, is in qualification and will ship in the first half of 2015. A quad-core version with an integrated 4G LTE modem will be qualified in early 2015 and start shipping around the middle of the year.
These first SoFIA processors, which have roots in Intel’s acquisition of Infineon in 2011, will be manufactured by foundry TSMC on a 28nm process. In 2016, Intel plans to release an updated quad-core version, SoFIA LTE 2, manufactured on its own 14nm process.
For higher-performance mobile devices, Intel will continue to sell Bay Trail and Moorefield through much of 2015. Both are quad-core processors with a Silvermont CPU core manufactured on Intel’s 22nm process, but Bay Trail uses Intel graphics while Moorefield has Imagination’s PowerVR graphics, which is used in many smartphone chips. Later in 2015 Bay Trail will be replaced by Cherry Trail, a quad-core part with a new Airmont CPU core manufactured on Intel’s 14nm process.
These will be followed by two new 14nm processors in 2016: the quad-core Broxton for high performance and SoFIA MID with integrated LTE for mid-range mobile devices.
The transition from Bay Trail to more cost-effective chips, in particular SoFIA, will help narrow losses in mobile next year, but Intel still doesn’t expect the business to be profitable until some time in 2016. In addition, the tablet business won’t keep growing at the same rapid rate; Intel said that moving forward it expects to grow shipments at about the same rate as the growth in the overall tablet market.
Intel also seems to be changing its strategy in smartphones. When Intel first announced its tie-ups with Rockchip and Spreadtrum earlier this year, most of the coverage focused on breaking in to China’s market. That’s certainly important, but it now appears those partnerships are part of a broader strategy to rely on partners to build chips for low-cost and mid-range devices sold throughout the world.
The development of a “synthesizable” Atom core—one that can easily be integrated into other chipmakers designs like ARM’s CPU cores—was a big step here. Rockchip is already selling a processor that combines a dual-core ARM Cortex-A5 with Intel’s 3G modem, and in the first half of 2015 it will swap out the ARM cores for a quad-core Atom with SoFIA 3G-R. Spreadtrum’s quad-core Atom with integrated LTE will start shipping in mid-2015. This way Intel can reduce its own investment in developing low-end mobile chips but still increase the presence of the Intel Architecture in mobile.
Moore’s Law is alive and well
Despite formidable challenges, Moore’s Law remains at the heart of the company’s strategy nearly 50 years after Intel co-founder Gordon E. Moore first observed that over the long run transistor density doubles approximately every two years (technically, he initially postulated a faster rate). “It’s our job to keep it going as long as possible,” said CEO Brian Krzanich. After a lengthy delay, the 14nm yields are in a “healthy range” now and will catch up with where 22nm was at the same time in its production ramp by the first quarter of 2015.
William Holt, who heads the company’s Technology and Manufacturing Group, talked about how Broadwell was a “true shrink,” meaning the pitch or distance between major features on the chip are still scaling at the expected rate. The reduction in the distance between fins, the raised channels behind the name “FinFETs,” is right on target (0.70x), the distance between the gates on top fell just short (0.78x), but the interconnects scaled further than expected (0.65x) thanks in part to the introduction of a new feature, airgaps, that reduces cross-talk between interconnects spaced so closely together.
The result is that overall logic area scaling continues. In comparison to its predecessor, a 14nm Broadwell dual-core chip is 37 percent smaller and has 35 percent more transistors (1.3 billion compared to 960 million for Haswell)—a 2.2x increase in transistor density.
The reason this is so important is that the cost of designing and manufacturing chips is increasing. So chipmakers need to pack more transistors in a given area to reduce the overall cost per transistor. For the first time Intel showed a slide indicating it expects this trend to continue at 10nm and even at 7nm. Holt said that Intel will be able to achieve this with our without Extreme Ultra-Violet (EUV) lithography, a tool still under development and not yet suitable for volume manufacturing.
It was interesting that Intel barely mentioned its “tick-tock” cadence, the metronomic strategy in which a physical shrink one year is followed by a new microarchitecture in the following one, perhaps because of the 14nm delay. But I wouldn’t make too much of it. The reality is that Intel’s product development has grown more complex as it expands beyond PC microprocessors into all sorts of standard and custom products for different markets so it’s probably smart to remove those shackles.
Furthermore Intel is now on its third-generation of chips using FinFET transistors (it began production with Ivy Bridge in late 2011, followed by Haswell in late 2012 and now Broadwell) while competitors TSMC and Samsung/GlobalFoundries are still making 28nm and 20nm chips with planar transistors and won’t start production with FinFETs until mid-2015. That means Intel maintains a lead in process technology and manufacturing of 3.5 years.
IP reuse and the foundry business
All semiconductor companies develop intellectual property that they use in their own products or license to others. But Intel now seems to be placing more emphasis on this aspect of the business.
Intel spends significantly more than competitors on research and development (R&D), and CFO Stacy Smith talked about how 75 percent (or roughly $11 billion per year) of this spending goes to foundational technology (process development, CPU and core design, SOC integration, and validation and testing) or shared platform investments (3G and 4G modems, connectivity chips and graphics) that is used across all of its products.
For example, the 3G and 4G technology isn’t just for smartphones and tablets; Intel is seeing increased attach rates in laptops, especially Chromebooks, and it is critical for other categories such as the Internet of Things and wearables. So while Intel may scale back development of its own application processors for smartphones, it isn’t giving up on wireless where, after Broadcom exits, only six players will still be fighting it out (HiSilicon, Intel, Marvell, Mediatek, Qualcomm and Samsung).
During the Investor Day, Intel said the wireless and SoFIA team that is currently part of the Mobile and Communications Group will become a standalone Platform Engineering Group (PEG) that can work with customers in all of these areas. The other part of MCG that develops Atom processors for mobile devices will be merged with the PC business into a new Client Computing Group (CCG) under Skaugen.
The development of IP also helps Intel with development partners (Rockchip and Spreadtrum) and with its growing foundry business, which now has six public customers (Achronix, Altera, Microsemi, Netronome, Panasonic, Tabula). Today these customers are bringing their own designs using their IP--supplemented with technology from companies like Cadence and Synopsys--but eventually Intel plans to offer its own IP to these customers.
As semiconductor manufacturing grows more difficult and expensive, and the number of players dwindles (from 18 a decade to just four today in advanced logic), Intel’s manufacturing edge grows more compelling to other companies. In the past Intel has said it wouldn’t make chips for companies that compete with its own products; now it is talking about opening its foundry business to “any customer that is able to utilize its leading-edge technology.”
The bottom line
This year has turned out better than expected (Intel said sales will be up about 6 percent) and the company expects revenues to grow in “mid-single digits” again next year. The PC market has stabilized, and for now Intel is outgrowing it (PC processor shipments grew about 9 percent this year). Next year Intel expects to ship about the same number of units with overall sales down a bit.
The big drivers will continue to be Chromebooks and cheap Windows PCs at one end, enthusiast PCs at the other and more innovative designs such as 2-in-1s and better all-in-in-ones in between. Intel says there are some 600 million PCs out there older than four years that will need to be replaced, and the release of Windows 10 next year ought to give this refresh cycle a lift.
The real growth is in the data center, which is already a $14 billion business expected to continue to grow at an average of 15 percent a year at least until 2018. Enterprise IT hardware is still big and important, but other customers include the cloud service providers, telcos and technical computing. The growth drivers are the cloud architectures, software-defined networks (SDN) and network-functions virtualization (NFV) for communications, high-performance computing, and of course Big Data and analytics.
Bryant said Intel has now developed 35 customer Xeon processors for customers such as Amazon, HP, Microsoft and Oracle optimized for these kinds of applications. And it is getting an increasing share of its revenues from other data center technologies including storage controllers and Ethernet switches, silicon photonics to replace fiber and copper; fabrics and switches (TrueScale now and OmniPath starting next year) and communications gear (with the Axxia and Mindspeed acquisitions).
Intel’s other businesses--Internet of Things, Software and Services (McAfee), and memory--sometimes get lost in the shuffle but each is on track for more than $2 billion in sales this year. Although gadgets like the MICA bracelet, Basis Peak smartwatch and SMS Audio smart earbuds get lots of press, the IoT business is growing quickly due to applications such as retail, automotive and manufacturing.
The software business is growing more strategic and Intel said it believes that mobile devices based on its silicon will be the first--other than the Nexus-branded ones--to get the Android 5.0 Lollipop update. The memory business, a joint venture with Micron, is shifting from selling memory chips to developing client and enterprise SSDs, which are increasingly important to close the widening performance gap between CPUs and hard drive-base storage.
Next year Intel will start production of its first 3D NAND memory, a 32-layer 32GB chip, which starts showing up in SSDs in late 2015 and will enable drives with more than 10TB within about two years. Samsung is the only company currently manufacturing 3D NAND and selling SSDs using it, though Hynix is also expected to start production by the end of the year and Toshiba/SanDisk should start production sometime in 2015.