An ongoing IoT debate centers on the notion that just because we can do something does not mean we should. From discussions at the recent MEMS Executive Congress, looking at what TSMC and some others see as the endgame for a trillion sensors signals possible trouble ahead.
The TSMC story starts out OK, as told by George Liu, director of business development. “The next big thing” is the IoT, bigger than the sum of all opportunities that came before it. Liu made the excellent point that for the first time, semiconductor growth will no longer be limited to growth of electronic devices. Instead, the IoT cuts across all industries, with sensors, processing, and communications seamlessly embedded in applications.
Liu put sensors into six general categories, shaped around smartphone needs: touch, vision, audio, motion, environmental, and biometric. He did a decent job of explaining existing gaps in IoT chip design: physical size, ultra low power consumption, always-on and power management, mixed signal and RF integration, packaging, capacity, and price. That would be the list of the usual suspects.
image courtesy TSMC
Then, he portrayed the case for TSMC as a MEMS sensor supplier, although I have to say it was vague at best. Most of Liu’s reasoning hung on one claim: TSMC will run more mixed signal wafers in 2014 than the #1 analog ODM. I’ve written in my blog about the pressing need for more microcontroller and SoC capacity to support even conservative IoT projections. We certainly need the scale TSMC can bring to the party.
However, that implies “all else being equal,” where things are not equal.
Fabricating a MEMS sensor is very different from fabricating a microcontroller with an ARM, MIPS, or ARC core and some mixed-signal peripherals. Most MEMS sensors are anything but monolithic CMOS. Bosch, ST, and others have made progress with package-level integration of MEMS sensors with Cortex-M cores. 3D-IC technology or stacked chip-on-package also work, but are not cheap.
Advanced nodes, like 16nm FinFET, are not the answer. Microcontrollers and MEMS sensors are on more mature nodes with good reasons. Many automotive parts are still on TSMC’s 180nm BCD process. The most aggressive MCU we can currently find is the Atmel implementation of the ARM Cortex-M7 on 40nm; ST has their Cortex-M7 on 65nm. Some of the technologies, such as embedded flash and A/D converters, are very unlikely to scale much past 28nm.
My lunch table hosts were from Merit Sensors. They build MEMS pressure sensors on a 4” fab, and I didn’t get a definitive answer but it sounded like either 350nm or 180nm. Merit bases their business on high precision designs, customer service, and support of longer lifecycle applications. There is absolutely no reason for them to move to more aggressive process nodes.
GE Healthcare had a fascinating presentation on developing MEMS RF switches surviving billions of cycles for their MRI systems, right down to proprietary non-ferrous metallurgy and packaging. Their problem is these devices use exotic materials in low volume, and they are looking for application partners to help scale costs.
mCube (who won the Innovation Award voting among six finalists at the conference, thanks to my vote) highlighted their MC7030 iGyro motion sensor. These are 9 axis sensors with a 3 axis gyro estimated from readings off a 3 axis accelerometer and 3 axis magnetometer. By replacing the physical gyro with software algorithms, they give up some sensor stability versus a full 6- or 9-axis IMU, but gain with package size and power consumption reduction.
VTT presented some of their research on MEMS in Fabry-Perot interferometry, essentially tunable optical filters for microspectrometers and extensible to hyperspectral imagers. With these sensors, they are putting a tool for skin cancer diagnosis in a handheld device, and fielding hyperspectral imaging on miniature quadcopters for agricultural applications.
At the beginning of any growth cycle, evangelism is essential. Finding more applications for MEMS sensors drives volumes, lowers costs, and builds momentum. All good. But, let’s not get carried away just yet. There are still the issues of behavior and macroeconomics.
Chris Wasden, formerly of PwC and now at the University of Utah, gave a solid presentation on inevitability and the hype cycle. He used an example of telehealth. It was 40 years from the first vision of a “radio doctor” donning the cover of a popular hobbyist magazine in 1924 until NASA created the first working implementation. Another 50 years have passed before CMS (the Medicare and Medicaid payer) has finally decided to pay providers for telemedicine starting in 2015. Technology didn’t make it happen without significant money flow.
Going back to TSMC and our point, the question for Liu at the end of his presentation: what has to happen to costs for a trillion sensors to happen? The answer he gave: “Well, it’s going to have to look like RFID … 2 cents per sensor.” You could almost hear the oxygen leave the room.
Everyone will buy sensors if they are that cheap. The problem is nobody will be interested in designing, fabbing, and selling them if they are that cheap. Not even TSMC. I thought maybe this was a sales guy talking out loud, not their official position. There is opinion, then there is bad analysis, then there is bad strategy.
There are two parallels in recent tech history. One is Intel. If you look at each successive generation of Intel PC processors, after an introductory period they stabilize somewhere around a volume price of $230 at maturity. That model is intentional, baked around costs, volumes, and capex. It stalled in-class competition for 20 years until disrupted by mobile SoCs, IP, and foundries.
A better comparison may be the $1, 32-bit MCU. It leverages IP and fabless to the hilt, at least in theory. At some point, price elasticity is counterproductive. There would be zero suppliers left if 32-bit MCUs suddenly dropped to 2 cents. Companies would be unable to cover R&D or COGS at those levels, fabless or not. That would leave TSMC with a lot of empty capacity, right? It hasn’t happened, and probably won’t, until some other technology comes along to replace the MCU.
Yet, Wasden says free hardware is inevitable. His argument is value is no longer in physical things, but in the data that comes off them. My brain cued up Mark Knopfler.
Now look at them yo-yos, that’s the way you do it
You put your sensors on the IoT
That ain’t workin’, that’s the way you do it
Money for data and your MEMS for free …
I have no doubt that on a micro scale, speaking of particular companies or specific application segments, monetizing IoT data will be a big thing. On a macro scale, nobody has shown me a convincing model where $29B of revenue per month from semiconductors transitions without major disruption. A 100x price drop can’t be made up in volume by wrapping a dollar bill around a 2.6 cent chip – no matter what this model Wasden showed says (and I’m not sure he believes it):
As Tom Kenny of Stanford put it in a later presentation, “winning can look a lot like losing.” Apple has no loyalty. They have thrashed Bosch, Invensense, and ST against each other in awarding design wins to gain a price advantage. Eventually, pricing for MEMS accelerometers in mobile will stabilize, just as MCU pricing has – it has to for suppliers to stay interested. Look at suppliers losing interest in mobile SoCs at a much higher price point to see what happens.
Lifecycles come into play as well. In mobile, devices are heavily subsidized against a forward stream of revenue and subscriber retention. Smartphones are replaced every 2 years or sooner, which has driven most of the MEMS accelerometer growth we see. In RFID, which is approaching the 2 cent frontier for passive tags, hardware is disposable. Passive tags are simply enablers riding along with the flow of goods. Did tags end up on every item? Pallets yes, high value items yes, but even the ballyhooed self-replenishing IoT refrigerator likely won’t drive item level tags. Why? The retail barcode infrastructure was good enough to work and expensive to replace.
On the actual IoT not in fantasyland, devices will last 5 or maybe even 10 years. Once light switches, door locks, and thermostats enter a home or building, they won’t be replaced quickly. Connected cars certainly will not be free, and their service life will be dictated by extended financing terms. Industrial IoT applications will resemble the IT maxim “if it ain’t broke, don’t fix it” once they produce ROI. Applications will scale out, not up.
What will happen on a 10 year horizon? First, as design wins accelerate with the IoT, MEMS sensors will mature around appropriate price points. Those might be 2 cents for passive RFID tags, 50 cents for accelerometers, and $5 for RF switches and spectrometers. Not everything will drive to the 2 cent mark, and averaging these classes together is pointless.
We definitely need more MEMS capacity for the IoT, the same argument as needing more MCU capacity, and TSMC might be part of that. I don’t think they will put niche MEMS foundries out of business, at least anytime soon.
Data will be monetized, but not on free hardware, and not on a gigantic scale that dwarfs the current semiconductor industry. The data plan subsidy model that benefits mobile won’t get enough wide-scale traction due to use case diversity and fragmentation. There are no free Teslas, or MRI machines, or Predator drones, and the HP printer model with consumables does not apply everywhere. Privacy and security will also rear their heads and make corner-cutting on hardware and software really dangerous.
We will likely grow MEMS sensors the old-fashioned way, with 15% to 25% annual growth rates (analysts don’t see it quite that high) and 2x or 3x price improvement over a decade. Anything beyond that will cause a structural shift leading to supply chain instability. While a few players may benefit from the asymmetry short term, the industry as a whole will experience a collapse if we are not careful. On a longer horizon, when new technologies emerge, more is possible.
As TSMC’s Liu concluded at the conference, the IoT isn’t about just growing the electronics industry; this is about benefitting mankind. Any effort driving MEMS companies out of the market early, unable to compete on an economic basis, will have the opposite effect.