When Intel (INTC -0.45%) announced it would be launching a family of fully integrated apps processors with connectivity/communications processors for smartphones and tablets named "SoFIA," Intel finally looked on the right track. These parts would be built at TSMC (TSM -3.85%) on its 28-nanometer process, and with die-sizes that seemed to imply good -- but not Earth-shattering -- performance levels, this Fool expected they would essentially be "proof of concepts," with their 14-nanometer successors making a real splash. Bad guess!
SoFIA 3G is dual-core Silvermont, not a single-core
I had guessed the 3G version of SoFIA would be a single core with a PowerVR G6100 GPU block, and that the LTE version (due to a significantly expanded die size) could be a dual-core with a larger GPU block (PowerVR G6200) and a more advanced modem. Of course, this assumption hinged on the die sizes implied including connectivity into the main system-on-chip. This assumption turned out to be false, as you can see below.
Intel's SoFIA allows the company to be competitive in the highest volume smartphones. Source: Intel.
Indeed, with connectivity (Bluetooth, Wi-Fi, GPS, etc.) as part of the separate RF chip, the main SoC complex looks a lot like Intel's mobile SoCs today except for the inclusion of a cellular baseband. Further, given that Intel is likely being much less aggressive on the GPU for SoFIA 3G or LTE compared to Merrifield/Moorefield (Intel's current performance phone SoCs) -- which used up a good portion of the latters' area -- the die sizes given make a lot of sense.
SoFIA LTE is a quad-core
One of the biggest concerns surrounding Intel's current-generation Merrifield system-on-chip is the fact that it is a dual-Silvermont core implementation. Now, those cores are fast, and in a smartphone, a dual core is usually the best use of a given power budget, but marketing is marketing, and Intel's competitors have quad/octa-core implementations. Of course, those octa-core solutions are made up of extremely weak cores, so the quad-Silvermont is likely the right balance of performance and "marketing."
Intel's tablet market continues to get more aggressive. Source: Intel.
Further, given that the die size of this quad-Silvermont SoFIA is smaller than Merrifield, but appreciably larger than its 3G cousin, it's likely Intel went with the PowerVR G6200 GPU block, which should offer roughly half of the GPU performance of Merrifield, but in the value segment of the smartphone/tablet market -- so it'll be quite competitive.
28-nanometer isn't a hindrance
While some may look at Intel's choice of the TSMC 28-nanometer process as a "hindrance," particularly as competitors talk about moves to 20-nanometer or even 16-nanometer FinFET, do note that SoFIA is a value oriented chip. Even Qualcomm -- the mobile processor leader -- is still on 28-nanometer poly Si (this is the weaker of the two 28-nanometer mobile processes that TSMC) today for its value processors, with a transition to 28-nanometer high-K metal gate (i.e., faster and better) likely in 2015 alongside Intel's SoFIA.
A wafer inside of Taiwan Semiconductor's fabrication plant. Source: TSMC.
In a nutshell, the leading edge 20-nanometer and 16-nanometer FinFET foundry processes will only be usable by mobile players that can afford to pay a premium for higher performing transistors/processes. This implies high-end, hero-device focused products will be at the leading edge node while the fabless players in the low-end/mainstream will be on 28-nanometers (either poly Si or high-K metal gate) for a long time. Intel will be competing on even footing in this space.
14-nanometer could mean the end game, however
During 2015, Intel will compete at the low end with parts built on TSMC's 28-nanometer process (likely high-K metal gate), but at the high end, it will be on its 14-nanometer process with Broxton (mid-2015 silicon shipments) for phones and tablets. Intel should do quite well at the high end with Broxton on performance/power. SoFIA will be competitive during 2015, but Intel has indicated that during 2016, SoFIA will be moved to the 14-nanometer process.
By 2016, TSMC and likely Samsung will have a FinFET process, but remember: This process will be brand new for these players while Intel will have had a whole year to build the entire gamut of high-margin/high-value products at its 14-nanometer factories, which means depreciation per wafer should be significantly lower for Intel. Lower cost will enable 14-nanometer value/entry parts for Intel while others are on 28-nanometer, perhaps moving to 20-nanometer -- a clear long-term advantage for Intel.
