I spend a good part of my day (and screen real-estate) monitoring the news-feeds across the semiconductor industry (which is what I primarily spend my days researching). There's a lot of information (and BS) that constantly hits the wires and the news sites, and to stay truly current, I like to keep myself as plugged in as possible.
Today, this little gem of an article from Digitimes showed up on my radar titled, "Samsung to build 20nm foundry capacity, sources say." Now, as somebody who follows the semiconductor equipment vendors, it was pretty well known that Taiwan Semiconductor (TSM) - the world's leading foundry - has been leading the charge on its 20nm planar process, while the other players have been more focused on getting to their FinFET processes (as these will offer substantial performance improvements over the 20nm planar processes) and have been fairly non-committal about 20nm.
However, it looks as though TSMC will have some competition at 20nm as Samsung is allegedly stepping up to the plate as it supposedly plans to put in capacity for 30,000 wafers/month. However, what does this have to do with Intel (INTC)? Read this little gem from the above Digitimes piece,
The current yield rate of Samsung's 20nm is estimated to be as low as 10%, the sources noted. At TSMC, its 20nm facilities are operating at 20-30% yield rates, the sources said.
So, let me get this straight. Samsung and TSMC, here in September of 2013, can't get yields of its planar 20nm process to levels that anybody would consider manufacturable, but Intel not only has designs that are complete, taped out, and are being shown in live demonstrations, but will go into high volume production on the parts by Q4 2013.
Why Is This Important?
Whenever you see a PR from the likes of ARM (ARMH), TSMC, or Samsung about some "test chip" being "taped out," this is usually just sleight of hand designed to get people (usually the tech press and investors) excited who don't know any better. While there are certainly technological barriers to implementing, say, FinFETs, High-K Metal Gate, and so on, the barriers aren't just about creating one or a few of these devices. It's about being able to build these devices in scale with good yields. In short, it's about being able to do it economically and in high volume.
Let me divert your attention to this:
The chart illustrates what is known as defect density. That is, given a semiconductor wafer, how many defects do you have per square centimeter? The fewer defects you have, the more good chips you have, which means that more of that wafer can be converted into revenue (meaning, ultimately, better gross margins).
How it works at Intel is that the company continuously develops the process (think of this as the "recipe" used to manufacture the chips) until the defect density is good enough to make it worth running a substantial number of wafers through the factories in order to try to get chips ready for sale. According to CFO Stacy Smith on the most recent earnings call, Intel will be in high volume production of its first 14 nanometer product - "Broadwell,"
We should get some good news associated with startup cost, and then the offset in Q4 is going to be, we should be well into the build of Broadwell
Now, understand this. Intel has been building chips on its 22nm FinFET process since Q4 2011, and products have been in the hands of consumers since April 2012. Today, the vast majority of Intel's product stack is built on the 22nm FinFET process from tablets to server chips and everything in between. No competitor has even demonstrated a working FinFET product.
Now, do I believe that the foundries will eventually get there? Of course. The engineers at Samsung and TSMC are certainly capable and they'll continue to advance. What I don't see is Intel losing its process edge - in fact, I expect it to widen over time. With 14 nanometer products going into high volume production in Q4 2013, and with the first 20nm parts from TSMC only beginning to go into volume production in 1H 2014 (as per TSMC's own statements), the nearest competitor to Intel will not only not even have FinFETs (Intel will be on its second generation of the technology), but it will have a significant density disadvantage. This is pretty win-win for Intel.
Even more interestingly, I expect that by the time that TSMC has its 16 nanometer (FinFETs and a slight density improvement over 20nm - not a full generation) in full production and available in shipping products (likely by 2H 2015), Intel will be ramping its 10nm process beginning in Q4 2015, for production in mid-2016. This looks good for Intel. However, there's a snag.
A Caveat Applies
Anybody following the industry will notice something very interesting (and even contradictory) to this thesis. Here is Apple's (AAPL) A7 built on Samsung's 28nm HKMG process (image courtesy of Chipworks):
According to Apple, this puppy weighs in at "over 1B transistors" and sports a die size of 102mm^2. Here is Intel's "Bay Trail,"
This chip also weighs in at 102mm^2. We do not have a transistor count.
However, what has people fairly concerned is that Apple has designed a core that seems to trade blows with Intel's Silvermont in mobile power envelopes (the benchmarks here are single-threaded, results courtesy of AnandTech):
(click to enlarge)
In terms of GPU, Apple's A7 is markedly faster than Intel's "Bay Trail" (benchmarks courtesy of Anandtech):
Note that Apple currently licenses its GPU technology from Imagination Technologies (OTC:IGNMF) and that Intel will, for its smartphone platform, be doing the same. Intel's in-house GPU architecture should get significantly better in tablets and above with the upcoming "Gen 8" GPU architecture (the current Gen 7 is good, but not great), but for smartphones it looks like Imagination for at least a couple more generations.
Anyway, so the obvious question then is, "wow, if Apple can achieve comparable or even better performance in the same area on an older process node, then does Intel even have a process lead?"
The answer to this is "yes." From a raw transistor standpoint, the two main metrics that determine density are "contacted gate pitch" and "metal pitch." TSMC's 28nm process seems to have a contacted gate pitch of 118nm and a metal layer 1 pitch of 96nm. Intel's 22nm process sports a "contacted gate pitch" of 80nm and a metal layer 1 pitch of 90nm. So, given equal design teams, design goals, tools and methodologies, Intel's process should have a fair density advantage over TSMC (and Samsung) as well as a non-trivial transistor switching speed advantage. But there's the real snag, isn't it? These factors are not all equivalent, the designs and architectures aren't the same, and it's really apples to oranges.
The really important trend for investors to watch is how the delta from a product perspective against competition has evaporated. Two years ago, Intel didn't have any parts suitable for tablets. One year ago, "Clover Trail" got power and CPU performance under control but graphics performance wasn't spectacular; the chip was quickly out-classed by parts from Samsung, Apple, and Qualcomm (QCOM) several months after Clover Trail-based tablets shipped. But today, "Bay Trail" arrives right around the same time as NVIDIA's (NVDA) Tegra 4 and Qualcomm's Snapdragon 800, and features best-in-class CPU performance and GPU performance that's at the very least competitive all with what looks to be best-in-class power consumption.
Make no mistake, having the world's best transistors isn't enough, but it looks as though on the design side (and on timing) Intel has its stuff together in tablets. It seems that smartphones are more difficult, particularly as Qualcomm's dominance on the cellular baseband side of things (and its experience in actually integrating everything together) is formidable. Intel's demonstration of carrier aggregation on a Merrifield-based smartphone at IDF was encouraging, but I'll wait until I see what designs Intel is able to win on the smartphone side before I start getting too giddy. Over the longer term, I have no doubt that Intel will be the merchant vendor of choice in tablets and above, and I further believe that it will be a strong competitor (perhaps not the leader, though) in smartphones.
Process technology will help Intel gain an edge, but remember that it's the right designs coupled with the best process technology that ultimately wins the day (as we saw in the Intel vs. AMD battles long ago). Right now, Intel's got the best low power process technology, the best low power CPU core (outside of Apple's "Cyclone", anyway), a good (but not leadership) tablet-class graphics architecture, and a massive base of OEMs all looking to distinguish themselves in a very crowded mobile market.
Intel will be fine, and as soon as the money starts coming in from these new products, the money will start flowing into the stock, too.