Attending, what struck me the most was how Intel is changing from a company led by the PC client to becoming one that is much more diversified, and one that is increasingly being led by its Data Center business. This was best exemplified by the news that, in a few years, when the company is finally ready with its 7nm process, the first chips created via the process will be Xeon processors aimed at the data center. That’s a big break with tradition—for decades, Intel has brought its newest technology first to processors for clients—once desktops, now notebooks—with server products tending to follow a year or more later.
This is a big part of CEO Brian Krzanich’s plan to position Intel to address a much larger market than the traditional PC and server businesses, which together have a total addressable market of about $45 billion a year. Instead, he said, Intel is going after a much larger market, including the broader data center (covering networking and interconnects), non-volatile memories, mobile (through premium modems), and the Internet of Things—items that together represent a market with a $220 billion total addressable market for silicon by 2021.
All of these markets, he said, build on Intel’s traditional strengths in silicon and process technology. And they are all linked by a need for computing on larger amounts of data in the future, in a vision that sees data collected, moved to the cloud, used for large-scale data analytics, and then pushed back out; but with more computing needed on devices at the edge for real-time decisions as well.
As he has in a number of recent presentations, Krzanich explained that he sees the amount of data growing tremendously, noting that today the average person generates about 600MB of data each day, and forecasts that this will grow to 1.5GB by 2020. While today’s cloud is built mostly on data from people, he said, the cloud of tomorrow will be built mostly on machine data. The average autonomous vehicle produces 4TB of data a day, a plane 5TB, a smart factory a petabyte, and cloud video providers can push out as much as 750PB of video daily. Individual applications could produce even more he said, noting that the company’s “360 Replay” technology used during the Super Bowl and other sports events,
I found it interesting that Krzanich said Intel’s top priority for the year is continued growth in the data center and adjacent technologies. This was followed by continuing to have a strong and healthy client business, growth in the Internet of Things business, and “flawless execution” in its memory and FPGA businesses.
Other speakers gave details about each of these markets, including some interesting technology and market trends, as well as financial projections.
10nm Technology and the PC Business
Murthy Renduchintala, who runs the company’s Client and Internet of Things Businesses and its Systems Architecture Group, began by talking about “trying to align process roadmaps with our product roadmaps,” and explained that as an integrated device manufacturer (IDM)—in other words, a company that not only designs semiconductor products but also manufactures them—Intel has several advantages.
Renduchintala compared Intel to an “artisan baker” who not only can make bread but can also work with farmers to decide which wheat germ to plant and where to plant it. This way, the product designers can look at transistor physics three years before a product is manufactured. For instance, he said, Intel used different flavors of transistors for CPUs and GPUs even within the same chip, a level of granularity that Renduchintala said fabless semiconductor companies would find difficult to achieve. (He joined Intel about a year ago, from Qualcomm, which like most other vendors in the industry uses foundries to do the actual manufacturing of its products.)
Even though other companies are talking about producing chips on 10nm and even 7nm, Renduchintala said that Intel has a three year lead over the others. He noted that rather than focusing only on gate pitch, Intel focuses on the effective logic cell area, defined as cell width by cell height, which determines the overall area of the cell. He said Intel will maintain this lead even after competitors deliver 10nm later this year. Intel plans to release its first 10nm chips later this year as well—Krzanich showed a 2-in-1 laptop powered by a 10nm Cannon Lake processor at CES in January—and this will account for significant volume in 2018, he said.
The economic side of Moore’s law is alive and well despite rising wafer costs, Renduchintala said, noting that the company believes this will be true of the 7nm node as well. But he made a new emphasis on improvements within the process node, saying each of the three generations of 14nm technology thus far has produced 15 percent better performance using the Sysmark benchmark. He believes Intel can continue to do this on an annual cadence, with continued process improvements as well as design and implementation changes.
On the PC business, he noted that even though PC units have been falling, Intel’s profits in the segment grew significantly last year, mostly because of a focus on particular segments, such as PC gaming, where the company introduced a 10-core Broadwell-E platform with an average selling price of over $1,000; and by pushing platform technologies, such as LTE modems, Wi-Fi, WiGig, and Thunderbolt. He noted that the company has grown its mix of higher-end processors and hopes to continue that trend in 2017.
Looking forward, Renduchintala said the client group has made strategic bets on VR and on 5G modems. He noted Intel’s approach to 5G is very different from its approach to 4G, where it initially pushed WiMax, while the rest of the industry settled on LTE. He said Intel now knows it needs industry-wide standards and partners and cited a variety of companies Intel is working with on core networking, access network standards, and wireless radio standards. He said Intel is the only company that can provide 5G “end-to-end” solutions from the “cloudification of the RAN” (the radio access network) to the data center, and said it plans to be shipping samples of its first 5G global modem by the end of the year—using Intel’s 14nm technology—and plans to ship these in the millions in 2018.
Data Center Grows Beyond Traditional Server
Diane Bryant, who runs the company’s Data Center Group, focused on how enterprises are going through a period of transition, driven by the move to cloud computing, network transformation, and the growth of data analytics.
One big change for her group going forward is that it will be the first to launch on the next generation process node, meaning that Xeon products will be Intel’s first 7nm processors. In addition, she said, the data center products would also be the first on the “third wave” of 10nm products. (The first wave of 10nm, for mobile products, is due out at the end of this year, so the first 10nm servers won’t be out until next year at the earliest. Intel hasn’t yet confirmed an exact date for its 7nm process, but it seems likely that it would be in 2020 or 2021.)
A few different factors will make this change possible, Bryant said. First, the Data Center now has enough volume, as it takes a significant number of wafers to bring up a new process. But just as important is Intel’s new use of a packaging solution called EMIB (for Embedded Multi-die Interconnect Bridge), which lets the company cut up a Xeon die into four pieces, each of which can be debugged independently, and then connected via this 2.5D package, so it functions as a single chip. (The new package was actually first announced in 2014, but the company gave more details at this week’s ISSCC conference, and this looks like its first major use.) Until now, a server die was just too big to be used for
Bryant noted how Intel’s overall data center business grew 8 percent last year, but enterprise and government sales were actually down 3 percent, while cloud server provider sales were up 24 percent and communications service providers were up 19 percent. Enterprise sales accounted for 49 percent of the business last year, the first time this business was less than half of the group’s sales.
Bryant said that enterprises continue to need more compute—growing at 50 percent per year—but said that some workloads are quickly moving to the cloud, while others are mostly staying on premises. For instance, she said, collaboration workloads grew 15 percent in the cloud last year, but actually shrank 21 percent on-premises. On the other hand, she said, high-performance simulation and modeling require extremely low latency, so it is almost entirely run on-premises. Overall, 65 percent of workflows are now run on-premises, a figure she expects to level out at about 50 percent by 2021.
Broadly defined, artificial intelligence applications account for about 7 percent of today’s servers, Bryant said, with the majority running classical machine learning algorithms in applications such as recommendation engines, stock trading, and detecting credit card fraud. But, she said, deep learning—the neural-network approach used in the prominent image recognition and voice processing applications—accounts for 40 percent. In this area, Bryant talked about how GPGPU instances have gotten a lot of attention, but that overall these still impact only a small percentage of the overall server market: 20,000—30,000 servers out of 9.5 million.
Bryant noted Intel’s intention to serve all parts of the AI market with a series of processors, including the next-generation traditional Xeon servers; packages that combine Xeon with the firm’s FPGAs (through its Altera acquisition); Xeon Phi (with many smaller cores in a new version called Knights Mill that allows lower-precision calculations); and Lake Crest, which includes a chip specifically designed for neural networks, a result of the acquisition of
Another change is Intel’s increased focus on what it calls “adjacencies”—products that surround the server, including its OmniPath interconnect used in the high performance computing market; silicon photonics, including an on-chip laser providing 100Gbps now, with 400Gbps on the roadmap; 3D XPoint memory DIMMs; and its Rack Scale Design proposal for denser, more energy-efficient server racks. Bryant talked about the increasing importance of the networking market, where Intel is working to convert communication service providers from ARM and custom processors to the Intel architecture, as part of a move to SDN and Network Functions Virtualization. She said she expects 5G to be an “accelerant” in that effort. Bryant also said Intel is now the leader in network silicon (counting both its data center products and the Altera FPGAs, although the slide she showed indicated it is still a highly fragmented market).
3D NAND and 3D XPoint Memory
Rob Crooke, who runs the company’s non-volatile memory group, talked about why now is “a great time to be the memory guy at Intel,” and addressed the company’s plans for both 3D XPoint and 3D NAND flash memory.
I was a bit surprised to hear relatively little on the Optane drives, which Intel is preparing using the 3D XPoint technology. These drives are arriving a bit later than originally expected, but Crooke said that they have begun shipping the first units to data centers, and said the company has a clear path for three generations of this technology. He seemed to be positioning them more as eating into the market for high-performance memory (DRAM) than for the traditional SSD storage market, at least initially, but in the long run, both Crooke and Krzanich sounded very optimistic on Optane, and not only in the data center, but in enthusiast PCs as well, with Krzanich saying that “every single gamer” will want Optane in his or her system.
Crooke said this would be “an investment year” for Optane, with the company expecting such drives to account for less than 5 percent of total storage revenue.
Crooke was extremely enthusiastic when talking about the firm’s plans in 3D NAND. He explained that he thinks Intel has a competitive advantage with its 3D NAND products
To illustrate how fast density is improving with this technology, Crooke first held up a 1 TB hard drive, and then showed how the first generation 1 TB SSD was a bit smaller. Then he held up the 1 TB module currently shipping, which looks to be about the size of a stick of gum, and then showed the module Intel will be shipping later in the year, a single thumbnail-sized package. To illustrate how this will impact the density of a data center, he held up a thin 32 TB module designed for a server and said that using this module you could now get 1 petabyte in a thin 1U server, instead of a full rack server, which would be required with hard drives.
Internet of Things & ADAS
Doug Davis, who has been running the firm’s Internet of Things group and is now focusing on the advanced driver assistance systems (ADAS) group, talked about both of those areas.
On IoT, he said Intel’s interest is primarily in the value that data has when moving through the network to the cloud, and the application of data analytics, as well as analytics on the edge. He said the difference between IoT and earlier embedded systems is primarily about connectivity and using open platforms. Davis cited a Gartner study that said there were 6.4 billion connected things at the end of last year, an increase of 30% over 2015.
In particular, Davis focused on the retail, transportation, industrial/energy, and video markets, including network video recorders and data analytics moving to cameras and video gateways.
Davis’s biggest focus was on autonomous driving, which he said would be the most visible AI application in the next 5 to 10 years. He talked about how this will require connections back to the cloud and said that while today’s cars use $100 to $200 of silicon (much of this for the infotainment system), by 2025 the silicon bill of materials may increase to 10-15 times that number. Davis said Intel is involved in a number of autonomous vehicle tests, including a 5G trial platform, and has a partnership with BMW and Mobileye for the next generation of such vehicles.
Michael J. Miller is chief information officer at Ziff Brothers Investments, a private investment firm. Miller, who was editor-in-chief of PC Magazine from 1991 to 2005, authors this blog for PCMag.com to share his thoughts on PC-related products. No investment advice is offered in this blog. All duties are disclaimed. Miller works separately for a private investment firm which may at any time invest in companies whose products are discussed in this blog, and no disclosure of securities transactions will be made.