OK, OK, so we are not yet at the end of Moore's Law, but we can SEE the end from here.
Moore's Law comes to us from Gordon Moore, one of the founders of Fairchild Semiconductor (FCS) and, later, Intel Corporation (NASDAQ:INTC). In a nutshell, Dr. Moore wrote a paper in 1965 describing his observation that the number of transistors on a given cost integrated circuit had doubled each year since the invention of the integrated circuit, and his belief that the number of transistors would continue to double, at the same cost, every two years for the foreseeable future. The subtle detail that is often missed in casual discussions of Moore's Law is that the number of transistors doubles AT THE SAME COST. So the size of a transistor has continually become smaller and less expensive.
This is almost unbelievable insight at a time where we were still four years away from landing on the moon, and we still didn't have the ubiquitous, and almost free, four function handheld electronic calculator.
Moore's Law (Moore's Observation) has been driving the progress in semiconductor technology for the past 50 years, and quite accurately, I might add. The feature size or "node" has been reduced from 10 um (micrometer) for the first commercial microprocessor (4004) to 14nm (nanometer) for the soon to be shipping Broadwell microprocessor. That is a 99.86% reduction in the size of a transistor. The 4004 had 2300 transistors and the Broadwell is expected to have about 1.6 billion transistors. That is almost 700,000 times more transistors than the first microprocessor!
All good things DO come to an end, and Moore's Law will come to an end. It might be at 7nm, or 5nm, but be sure that it will come to an end. The reasons for this are many and are matters of physical sciences. For example, if the insulating films on transistors become too thin, electrons will "leak" around and through the individual transistors. This will cause the switching speed to decrease. When the switching speed of a CMOS circuit slows down, it spends more time in a transition from, or to, a one or a zero. The power consumption of CMOS is highest when it is in these "in-between states." So, eventually making planar transistors smaller than about 28nm increases electrical leakage and actually reduces performance and increases power consumption; exactly the opposite of what we want today. Oh, and the smaller, worse transistors will cost more. All in all, a very bad deal.
About 10 years ago, at 130nm, the industry thought they were at the end of Moore's Law, but Intel invented "Strained Silicon" and moved smoothly on to 90nm. A second generation Strained Silicon allowed Intel to move to 65nm node. Everyone thought that 65nm was certainly the end of Moore's Law, but Intel invented the High K, Metal Gate (HKMG) process and brought us 45nm integrated circuits, a second generation of HKMG moved us down to 32nm. Certainly 32nm had to be the end of the line for Moore's Law, but Intel re-invented the basic structure of the transistor and brought us 22nm ICs based on a TriGate transistor. TriGate (or finfet for the generic name) got around the leakage problem mentioned above. A second generation TriGate process is about to be released at 14nm.
Intel says they have "line-of-sight" to 10 nm and "some ideas" for 7nm. At that point, even I have to jump on the Moore's Law end-of-the-line bandwagon.
There is another thing going on that brings on the end of Moore's Law and that is that as the processes shrink the parts of the IC that don't scale (shrink), such as the bonding pads and I/O functions become a larger percentage of the chip size and eventually these non-shrinking parts actually determine the chip size. At this point, further attempts at shrinks have rapidly diminishing value.
Ok, if you have hung in there for this long, it is time to ask the question, "What happens at the end of Moore's Law?"
Current production for Intel competitors is 28nm planar processing. Intel is at 14nm TriGate, arguably three nodes and six years ahead of the next best in class. To give the benefit of the doubt we'll view this as a two node, four year lead. For Intel, the two node, four year lead has cost about $10 billion each year in capital expenditures. Since Intel is two nodes ahead of the next best, which, in my opinion, is TSMC (NYSE:TSM), TSMC will have to continue to spend $40 billion in CapEx and work for four years in order to get to where Intel finds the "end-of-the-line" to be. Intel, meanwhile, will virtually stop CapEx spending for those four years, gross margins will spike almost unconscionably, and they will be operating fully depreciated factories that can sell products at a much lower price than any competitor, if there is such a thing.
It's even more interesting than that; any of the other perceived Intel competitors, such as GF and Samsung (OTC:SSNLF) will have to spend the same $40 billion as TSMC or admit defeat in the bleeding edge of semiconductor processing technology. That is a total of $120 billion in CapEx that has to be spent chasing Intel. I find this competitive conundrum absolutely fascinating. Before very much longer the weaker of the foundries will have to throw in the towel. Who will be first to give up? I'm betting GF. The other two will soldier on, spending $80 billion on a business that they are destined to lose.
If the best competitors, TSMC and Samsung, never get to the end of Moore's Law for technical reasons or, more likely, financial reasons, Intel will have the bleeding edge, end of Moore's Law business all to themselves at pretty much any reasonable price they wish to charge.
Now, let's examine the potential competition to Intel:
Global Foundries is the sold off fab of AMD (NASDAQ:AMD). Even AMD paid to get out of the residual partial ownership of, and commitment to GF to take wafers, because the yield on those wafers was so bad.
Third parties talk about GF breathing down Intel's neck, but if you go to the GF website you can't find any native information on finfet or 20nm processes. If GF had a position in the technology, they would be shouting it from the rooftops… or at least whispering about it on their website.
Gotta remember, GF went gate first, with the IBM gang. It is virtually impossible to make finfets with a gate first process.
TSMC is working diligently to move their nodes of manufacture down as aggressively as possible. TSMC claims to be shipping gobs of 20nm planar (non-finfet) ICs. Nobody has seen any of these critters in the wild. Maybe they are like Big Foot; if you believe hard enough maybe it comes true. More likely TSMC is having a terrible time with 20nm, just like they had with 40nm and 28nm. Remember, Intel never even bothered with 20nm planar, because as I explained above, Mother Nature and physics get very hard to deal with in this particular arena. If TSMC is making 20nm planar ICs, I would bet they are no better than their 28 nm products and cost more. Two years ago Nvidia (NASDAQ:NVDA) called the TSMC 20nm process virtually worthless due to physics problems and physics problems don't change in two years.
TSMC also is working on a 16nm finfet process that they would hope to ship in real products in 2015-2016 depending on who you believe. This 16nm finfet process is 16nm in name only since parts of the 20nm planar process will be used.
All of that aside, shooting for a 16nm finfet from 28nm planar is what is called "node-hopping;" moving two nodes in one change. That is an incredibly risky undertaking. But, what choice do they have? AMD, Qualcomm (NASDAQ:QCOM), Nvidia and many other mobile chip suppliers are depending on TSMC for leading edge processes and their very survival.
For two years we have all heard various rumors that TSMC… had the Apple business, was going to get the Apple business, or that they had turned down the Apple business. I pretty much reject all TSMC/Apple rumors now. If any of them had substance TSMC shares would be selling at $40 a share.
IBM might have bailed some folks out, except for the fact that they are trying to sell their gate first semiconductor operation. Enough said about IBM.
Samsung has the money and they certainly have good technology. The latest rumor is that Apple (NASDAQ:AAPL) deemed Samsung's attempt at 20nm planar of not sufficient yield to even consider for the A8 chip. And, of course, Samsung is grinding away of some kind of a finfet process.
I have no idea what is going on at Samsung, but let's assume for a minute that Samsung is successful in getting to the end of Moore's Law with Intel (after spending $40 billion more money and four years more time), and let's say that TSMC doesn't make it. That would mean that the mobile chip suppliers mentioned above would have no choice but to buy their foundry services from their customers' strongest competitor, Samsung. Isn't that the same situation that Apple has been trying desperately to get out of for the past three years?
So, we have Intel who has been shipping 22nm TriGate devices for two years and is about to release the Broadwell processor built on second generation TriGate at 14nm. Who has the credibility? It seems to me that Intel has four years to get their stuff together on a compelling mobile SoC design. It will only take a year for that to happen.
The other thing that happens at the end of Moore's Law is that the fabless business model at the bleeding edge technology takes a hit. Imagine; if catching Intel was possible it would cost a ton of money, but in the fabless model the margins are split between the foundry and the fabless customer. This means that the foundry can't match the spending that would be required to compete with Intel.
The Intel $40 billion PC CPU business at 65% gross margin and the $10 billion server chip business at 85% gross margin just keeps clicking along on auto-pilot generating the cash required to defeat all comers in the bleeding edge high technology semiconductor business. I wonder what would happen if the PC WASN'T dead?
There's more. Vertical, otherwise known as 3D.
When the topic of 3D ICs comes up, the conversation turns to 3D NAND memory. There are other 3D semiconductor processes. My personal favorite is TriGate. Intel has years of experience in the etching and deposition, and plating technologies to make these vertical TriGate structures. The same kind of skill set is required for TSVs (Through Silicon Vias) as used in the Micron (NASDAQ:MU) Hybrid Memory Cube (HMC). It is no coincidence that the HMC was a joint development between Intel and Micron. You can also take it to the bank that Intel will be deeply involved in the Micron version of 3D NAND memory through the Intel/Micron emerging memory technology joint venture. So, 3D structures will keep density and performance improving, but the accompanying cost reductions that we have come to expect from Moore's Law will stop.
The semiconductor industry is in uncharted waters. There will be one undisputed technology leader. I've placed my bet that the ultimate winner will be Intel.
Intel is a buy and 2014 is the year.
Disclosure: I am long INTC, MU. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.