Intel’s new CEO Brian Krzanich has been acknowledged very openly that the company must capture significant share in the mobile market, or risk losing its leadership of the chip industry forever. This is not new thinking of course, but there are signs that, under the new regime, some initiatives within the company will be given far freer rein to grow; and that the firm will attack the sector via a far larger variety of routes.
In the former category falls the communications infrastructure business, which has been making steady but somewhat shadowy progress in recent years and this week unveiled its new flagship platform, Highland Forest, to take on Broadcom and others. Indeed, Broadcom is rapidly becoming as important as Qualcomm for Intel to challenge, because the former targets so many sectors where Intel needs to grow, from network backbones to smartphone SoCs, from home appliances to the internet of things. The launch of Highland Forest also signified a sort of coming-out party for the comms and storage infrastructure business under its VP and general manager Rose Schooler, who we expect to be hearing from far more frequently in 2014.
In the latter category, new lines of attack, comes Krzanich’s plan to turn Intel into a foundry, an effort which has started small, with a few deals with existing partners, but could turn the company’s hugely advanced and expensive processes into new revenues from third party manufacturing. Intel has traditionally harnessed its cutting edge processes for its own competitive edge, and has been a strong believer in investing heavily in this area during downturns, in order to sweep the market at the moment of recovery. This time around, it looks set to expand that thinking to foundry services, and could start making chips for mobile and comms players as its 22nm fabs with 3D transistor technology come online and it starts to push towards 14nm.
Krzanich said at last month’s analyst day: “We’re going to go much further. If we can utilize our silicon to provide the best computing, we’ll do that. People who can use our leading edge and build computing capabilities that are better than anyone else’s, those are good candidates for our foundry service.’ He acknowledged: “We’d become insular. We’d become focused on what was our best product rather than where the market was moving.’
While Intel may chase foundry business in the mobile world ‘ rumors persist that it will steal some of Apple’s next generation Ax processor business from Samsung and TSMC ‘ it has also been relying more heavily on third party foundries for some of its own mobile and low power chips, especially when it has been venturing into new markets. For instance, the recently announced Sofia smartphone system-on-chip ‘ a product originally developed by Infineon, which will combine Atom, 3G, GPS and connectivity in an integrated package ‘ will be made by a partner.
Intel knows well that the mobile device SoC market will be extremely tough, given the dominance of Qualcomm, its tardiness in making any real impact, and the rise of Chinese challengers. There should be better margins and room for growth in other areas of the mobile world. Cloud-RAN is an obvious and hotly pursued segment, providing a growth area for x86 servers in which Intel has the upper hand against emerging ARM-based cloud server designs. And communications infrastructure is another interesting area, given the huge increase in network capacity that will be required by continuing internet data growth; and the relatively open playing field.
Intel has been steadily widening its reach in this space, targeting many areas from super-switches to base stations. Its approach involves combining x86 processors with specialized coprocessor to offload and optimize specific tasks for a particular piece of infrastructure. It argues that using generic processors, rather than expensive specialist silicon, for many functions reduces cost and allows for easier migration and software development. That general purpose x86 processor can then be supplemented with companion chips for the most intensive tasks. The rise of software defined networking in many areas of networking will, like Cloud-RAN, boost Intel’s proposition by running far more network functions in software on x86 servers, and earlier this year the firm unveiled its SDN switch reference platform and support for carrier NFV (Network Functions Virtualization).
However, many telcos and enterprises will continue to run their networks on dedicated hardware for many years to come, so Intel also needs to find a role for general purpose x86 processors in that world. ‘This isn’t a strategy that started when SDN started to get attention,’ Schooler said on the conference call to launch Highland Forest. ‘We’ve been on this journey for a decade. We’ve been looking at the evolution for the Intel architecture franchise within communications and redefined networking to boil it down into all of the infrastructure and workloads that support a network function.’
Its recent strategy has seen its Xeon processor paired with a widening variety of companion chips – Crystal Forest for the data plane, Jasper Forest for the edge router, and so on. The latest announcements build on, and expand, the platform, targeting four key areas of network infrastructure workload ‘ application processing, control plane processing, data plane packet transport, and signal processing.
Schooler says communications infrastructure is a $16bn business, when telcos, enterprises and cloud servers are considered, and Intel has only 5% of that. It aims to increase the percentage dramatically with a processing and chipset platform to take networking equipment beyond ‘fixed function devices with customized silicon, towards more of a general purpose approach’. That pitches Intel not just against Broadcom but the habit of major vendors like Cisco to use their own customized chips.
Highland Forest is Intel’s third generation communications platform and the most aggressive in harnessing Intel’s power in enterprise data centers to shift the economics of networking. It is based around the Xeon E5-2600 v2 processor with the new Coleto Creek chipset, which is optimized for high end telco infrastructure, boasting the ability to process up to 255m packets per second.
Coleto Creek is the successor to the Cave Creek coprocessor and achieves 2-6 times higher performance even with the same 32nm process. That is because it uses the latest Xeon release, and some updated accelerators. Up to two Xeons and four Coleto chipsets can be combined, each consuming between 17W and 20W of power at most.
Coleto Creek is suited to handling various tasks including “applications, control and data plane traffic and to some extent signal processing’ according to Steve Price, general manager of the comms infrastructure division. Intel’s pitch is all about streamlining and simplifying the infrastructure which will have to handle such vast amounts of internet traffic, and this is not just about generic platforms which can target different tasks, but ‘a simpler tools suite and engineering, providing faster time-to-market’.
Coleto Creek will go up against offerings like Nitrox III from Cavium, but takes a different approach. Instead of using acceleration logic for regular expressions, as many network processors do, it relies on the x86 for these functions. Boosting the performance of these tasks comes from advances ‘ current and future ‘ in the AVX instruction set, plus use of caches and external memory. Intel said it had scored twice the number of design wins for Cave Creek that it had expected, and has now signed up Avantech, Adlink, Dell, Emerson, Radisys and Silicon to make boards based on Coleto Creek.
Its next plan is to produce the long awaited base station platform, built around a Xeon accelerator optimized for signal processing. Prototypes based on FPGA chips are being used in China Mobile’s huge Cloud-RAN market trial which, as Price put it, aims to ‘create synergy from big cells serving thousands of users to small cells in metro areas of public networks’.
Intel has talked about producing a full base station offering, though to date it has mainly worked with partners to get x86 processors into cell site equipment, as seen in its alliances with NSN over the RACS/Liquid Apps offering, and Cisco/Ubiquisys for the Smart Cell.
The more base station functions are virtualized on servers, of course, the more direct role Intel can play in the world of Freescale and Texas Instruments, though few doubt that specialist coprocessors and DSPs will continue to be needed for a long time. Intel, then, is likely to invest more heavily in those chips, increasing its activities in the DSP space which it once quit when it sold off Dialogic in 2006.
The cost argument for a general purpose platform will be music to vendors’ ears at a time when operators and enterprises need to collapse their cost of data delivery. But the big challenge is to convince customers that a single platform really can perform in four critical areas of the network equally well, and as effectively as specialized chips, and even Schooler admits the company will need partners to improve its offering, and gain market confidence, in some respects.
Of the four areas targeted by Highland Forest, applications processing is the one where Intel obviously scores highly, since its platform is already dominant in enterprise servers and starting to make inroads into the more rarefied world of telcos’ back end systems for billing and so on. Here, the main area of new investment will be carrier grade Linux.
As for signal processing, which is most important in the RAN, Intel has no real recent track record and Schooler admitted a partner would be required to provide special purpose IP, but as noted above, there is progress into the base station area via the Cloud-RAN developments.
In data plane processing, Intel is limited in what it can target. Its chips are not fast enough to challenge the incumbent fast router silicon of Broadcom, or the customized chips used by Cisco and Juniper. However, with Crystal Forest and now Highland Forest, Intel aims to move in on the price sensitive, and lower-speed, edge router space.
Control plane processing has some of the same issues and has traditionally been combined with data plane processing in one box, though the two are increasingly being separated, especially in virtualized environments. In both these areas, Intel will be confined to the sidelines until it can buy or acquire a very high throughput x86-based system, and the firm is remaining silent on its plans in this area, even though it has made moves in the past couple of years, such as acquiring Ethernet switch-chip player Fulcrum in 2011.
Schooler said progress had been made to upgrade x86 to get close to a specialized chip’s packet processing prowess, but as in other areas, the vision of a network that relies entirely on Intel general purpose CPUs and software is far away ‘ even in the purest Cloud-RAN and carrier SDN roadmaps, operators and vendors still see a role for dedicated silicon such as DSPs to carry out the most demanding tasks. This is likely to see Intel making further purchases like that of Fulcrum as well as using third party ASICs.
‘In the transport there will be a network processor and ASICs and the CPU will not take over all cases of throughput,’ Schooler said.