ARM (ARMH) infrastructure partners, TSMC (TSM), GlobalFoundries, and others have played down the importance of Intel's (INTC) FinFETs and made aggressive claims about how they will reproduce the Intel advances and nullify their importance in 6 months to a year. This article will show what was involved in the development of FinFETs, how long it took, and how unlikely it is that TSMC and GlobalFoundries will be able to rescue ARM, and AMD (AMD), from the effects of the Intel developments. This is a significant long opportunity for Intel, and perhaps an even more significant short opportunity for ARM.
Ever since Intel announced the performance and power consumption figures for Silvermont:
"The Silvermont microarchitecture delivers industry-leading performance-per-watt efficiency. The highly balanced design brings increased support for a wider dynamic range and seamlessly scales up and down in performance and power efficiency. On a variety of standard metrics, Silvermont also enables ~3x peak performance or the same performance at ~5x lower power over the current-generation Intel® Atom™ processor core."
All of the ARM infrastructure suppliers and users have been making their own announcements to the effect that they would have 16/14 nm FinFet parts in production this year. And many Seeking Alpha comments - mostly in response to positive Intel articles - have noted this "fact" and deprecated the importance of the Intel announcements. This would be funny, if it weren't so sad.
Why this is sad is because it gives investors a false picture of what ARM and AMD can do in response to what Intel has done. Intel has been quite open in describing what they did and when they did it, as is evidenced by what follows. They describe in detail, how they achieved the power consumption improvements and why they did it. And they described some of the lithography and materials challenges they had to face to accomplish this breakthrough. Think about what you read below. Think about it in the context of how difficult it will be for the ARM camp to pull off the same breakthroughs, and when they might do so. Also, think about how they are going to line up super-modern fabs to build the next generation parts and those that follow.
And finally think about the financial impact of the superior Intel processes and components on ARM, TSMC, GlobalFoundries, and Samsung. Think about a persistent three year gap in technology and the competitive advantage that goes with it.
More than a year ago, on May 24, 2012, Intel announced upgrades to three existing fabs for 22nm production and three new or refitted fabs for 14 nm production.
What this means is that Intel's decision to commit the 22 nm FinFET process to production occurred sometime in 2007 - if not earlier. That was 5 years before the first salable chips left the fab! And it corresponds to what I told a fund manager who consulted me on June 7, 2013.
Why does it take so long for Intel, arguably the best semiconductor house on earth, to bring this process to market? And how long will it take TSMC and Global Foundries and Samsung to accomplish the same feat? Those are questions we will try to answer in this article.
In May 2011, Intel made two important and relevant announcements:
The first was by Mark Bohr, and he displayed a Powerpoint that detailed their accomplishments in the development of the 22nm FinFet process.
The second was also by Bohr and it summarized those accomplishments, below:
In fact, more than 2 years ago, Intel told us what was going to happen, and when it was going to happen - and Intel brought it off! To make these accurate predictions they had to have already made the breakthroughs necessary to achieve the required yields in production. These breakthroughs required significant advances in material quality and lithography. The presentation by Meagly details what needed to be done.
Briefly, he shows that with EUV lithography extended to 32nm and smaller line-widths, the roughness profile that results from using materials from a variety of sources far exceeds the requirements for manufacturing yield. Further, he shows that technology gaps were identified, materials research was basically done, and collaborations cemented four years before manufacturing. Below is a key slide from Meagly's presentation.
This presentation was published in September of 2004! About eight years before 22 nm production.
After Meagley became engaged in the required research work, he and other Intel managers worked for three years to solve the "research" problems, so that the second phase - "Path Finding" - could proceed. The decision to actually build P1270 probably occurred late in the research phase, and had to have been made 5 years - or more - before the first 22nm chips rolled off the manufacturing line. Even, after the Research was done, and after the Pathfinding and the process development and optimization for yield was done, Intel's Development work took another two years before production.
I have not brought into this discussion the time necessary for the development of design tools for defining transistors and other components, for model development, for design rules governing the ways in which connections are made to components, and so forth. This time can only add delay to the program, not accelerate it.
Now, we have a chronology that defines what was involved and how long it took to bring these chips to market. We also know that Intel is taking another two years after P1270 (22nm FinFET's) to bring P1272 (14nm FinFETs) to market.
I think we are starting to get an understanding of why Samsung is now buying Intel chips. They realize that they can't wait that long to develop this technology for their own semiconductor manufacturing lines. They need to buy from Intel to remain competitive with others who are or will be buying from Intel.
Now, tell me again how TSMC is going to pull this off in a year. Tell me how GlobalFoundries is going to do it in six months. Tell me how they are going to work out a viable manufacturing process for FinFETs, build and/or outfit the plants, bring up the plants, achieve manufacturing yields, develop design rules and models, and then build chips - just as Intel did.
Can't be done!
Long INTC. Short ARMH. Hold Samsung. TSM and AMD are toss-ups. GlobalFoundries, one of the big losers, is a private company - I would love to be able to short them.
Additional disclosure: I have no positions in any of the other stocks mentioned, and no plans to initiate any positions within the next 72 hours.