Two years ago, Intel invited journalists to Santa Clara for a series of in-depth discussions and demonstrations of the company's first serious 32nm cell phone, codenamed Medfield. Unlike the flash and pomp that had characterized other mobile phone promises from Intel, the company went to great pains to keep its statements on message and believable. Medfield, Intel promised, would compete against midrange smartphones. It would demonstrate Intel's ability to compete in mobile. Future products, like Clover Trail+ and Merrifield, would then really open the market to x86 devices.
Obviously it hasn't shaken out that way. Early launches, like the Xolo X900 and Orange San Diego were followed by a few anemic Clover Trail+ products -- and not much thereafter.
Market misalignment
Intel's mobile problems can be traced to several issues. First, the company's radio timeline has been anything but timely. When Intel bought Infineon more than three years ago, it pitched the acquisition as part of a plan to accelerate its deployment of LTE technology. That didn't happen -- the company's first 28nm fully compatible LTE modem, the XMM 7160, just started shipping to customers this past month. Intel's recent Fujitsu Wireless acquisition is part of a further effort to bring competitive LTE radios to market.
Intel's decision to keep Merrifield a dual-core part may have also been a strategic mistake. Hyper-Threading helped Medfield and Clover Trail+ keep up with dual- and quad-core devices, but Merrifield is a conventional two-core solution. Hyper-Threading gave OEMs some cover for claiming the chips were "like" quad-cores at a very small penalty to die size. Merrifield doesn't. Since consumers have picked up on the marketing message that quad core = high-end phone, this limits Intel's ability to sell Merrifield as a high-end solution. Benchmarks, meanwhile, are often coded to take advantage of four cores, even when general applications don't utilize them often, or particularly well.
Alongside the technical issues, there are a host of other potential roadblocks. Intel has always had certain margin targets it wants to maintain on its phone business, and while it was willing to reduce those rates to gain market share, phone manufacturers in the US are leery of being trapped and commoditized by Intel marketing. Historically, Intel's business practices have done an extraordinary job of keeping Intel's margins high. Dell, HP, and Lenovo's margins? Not so much.
The large vendors that could pay Intel's higher margin costs and focus mainly on high-end smartphones, like Samsung and Apple, are determined to create their own custom silicon and retain a greater degree of control over their own IP. Intel has enjoyed some success with smaller OEMs like Motorola and Lenovo, but can't apparently offer the kind of breakaway performance improvement that would drive serious OEM engagement.
Next page: The road ahead...
SoFIA: Of silk purses and sow ears
Last week, Intel announced SoFIA(Opens in a new window) -- a joint project with TSMC in which Intel will build both Atom and radios at the Taiwanese foundry. At best, this was an admission that the company is no longer prioritizing mobile growth and will try to enter the high-end market at a later date. Building a budget and value smartphone at TSMC might make economic sense since the foundry already handles Intel's wireless radios, but it cuts against the grain of everything Intel has always claimed about its phone business.
From the beginning, Intel sold its smartphone business as a long-term triumph of foundry technology. True, 45nm Atoms couldn't really hack it in the mobile environment, but Medfield could, thanks to a superior 32nm process. Skip ahead, and 22nm FinFETs were going to offer an advantage that TSMC couldn't match on 20nm planar. ARM cores might be smaller and use a bit less power on the same process, but Intel's technology leadership would close that gap -- then create a second, this time with Chipzilla in the pole position.
It may make short-term economic sense to keep radio manufacturing at TSMC, but it flies in the face of Intel's historic willingness to pay higher costs in the short term to reap greater economic benefit over time. Without an integrated radio, it's going to be much harder for the company to win designs with OEMs. Building at TSMC may solve that problem, but kills Intel's entire narrative regarding the superiority of its foundry process.The road ahead
Intel has shifted its efforts to tablets, where higher TDPs and non-integrated radios will be less of an issue. The company still plans to aggressively ramp its products in this space, with 14nm Airmont/Cherry Trail designs available next year. Cherry Trail will increase the number of GPU cores from 4 on the present Bay Trail all the way to 16, giving it Ivy Bridge-equivalent graphics in a 2-5W TDP envelope. That's a huge gain overall, and it comes alongside further refinements to the Atom core. By 2015, Intel's Broxton platform will supposedly be ready to drive Atom-powered systems all the way into the Performance segment.
Building high-performance Atom platforms is a complete change from Intel's old fears that doing so would cripple the Core family, and it's a sign of how much priorities at the company have shifted. And Intel hasn't completely given up on phones -- the company is still talking about a "Hero" platform, codenamed Moorefield, that will supposedly launch in 2015. By that time, if the company's foundry predictions are accurate, Intel will have an established 14nm line by 2015, while TSMC will still be ramping early 16nm FinFET volume by the end of the year. Intel is also far more confident on delivering significant advantages on its next-generation process.
If Intel had announced a simultaneous plan to bring radio manufacturing over to the 14nm node, with the SoFIA platform clearly positioned as a stopgap budget alternative, the strategy would seem like a genuine tactical decision, not a face-saving maneuver. Right now, this feels like Intel is effectively hitting "Reset" on its own metrics for phone market penetration. The only difference is that now we know Intel can build competitive phone processors -- it's just not doing so, for a variety of reasons.
Now read: Intel unveils 72-core x86 Knights Landing CPU for exascale supercomputing
