Once upon a time, the average mobile phone had chips from a few a dozen vendors inside. Tear open a phone today, and you are likely to see just a handful of suppliers for the active components. And someday we may see a phone where all the components (except for memory) come from a single vendor. We think it is worth examining how that happened and what it means for emerging chip categories like those for AI and autonomous vehicles (AV).
Below is a photo of the internals of iPhone 12, courtesy of the wonderful people at iFixit. By our rough count, there are about 50 active semiconductors in this phone, and 100 passive components. Outside of Apple, no one really knows what all those chips are, but the folks at iFixit have helpfully highlighted the key chips in primary colors (Note: there is another board not shown). These include the applications processor (from Apple), several modem related parts from Qualcomm, the RF front end module, and a host of power, display and camera parts. All together there are about a dozen vendors for these chips, with Apple itself supplying almost half the key parts. All squeezed together on tiny boards.
By contrast, here is the main board for the Nokia 3310, a classic phone launched 20 years ago. (This is actually a modern refresh of that device.) A few things stand out for us here. First, the board is fairly large and has many more passive components. And while this phone has far fewer active components it is also an incredibly basic device – no touch screen, no video playback, no camera, no AI, no operating system, no application processor. Yet still there are a dozen vendors represented on this board.
Modern phones are in an advanced stage of evolution. The basic telephony functions of the device, as represented by the 3110, occupy only a tiny portion of the board today. These elements have been shrunk down to fit all the things we need for a smartphone.
Aside from the addition of all those ‘smart’ functions, during the intervening years the active components of the phone have largely been subsumed into a much smaller number of chips, from a much smaller pool of vendors. True, both devices still have many chips for RF functions, but the modern device supports at least triple the number of frequency bands. Those old phones had components we no longer even think about anymore – transceivers and Intermediate Frequency (IF) to name two. Instead, the functionality those chips provided has been integrated (subsumed) into the chips we see today – notably the modem and the application processor. We still have transceivers in phones (which incorporate the IF functionality), but a smaller amount of packages which are much more performative.
There is a common pattern in semis. Complex boards end up consolidated into a smaller number of chips, with added functionality integrated into those chips that remain. This is one of the ways in which Moore’s Law works its magic. A single chip can contain the functionality of multiple chips produced on an earlier step of Moore’s Law. So the key question becomes who survives? Who wins the designs for that smaller number of more integrated chips? Take any category of consumer electronics and you can see the same pattern repeated over and over again. The companies that survive end up having a head start for winning chips for new features. Qualcomm started with modems, added IF, transceivers and now RF. But along the way, they had pole position to introduce applications processors – a chip which does not exist in the 3110.
The factor that determines who survives can largely be attributed to whichever company holds the strategic ‘high ground’. Since mobile phones need a modem to be mobile, that chip has the ability to influence design choices for many of the other chips on the board. The same holds true for CPUs in PCs. Or we could look at this from a different angle, the chip that runs a device’s most differentiated software is the chip that holds that high ground. Again, in mobile the key software is the wireless standard (3G, 4G, 5G, etc.), and in PCs the operating system (OS).
Which brings us to the question of where will we see this integration process strike next?
First, it should be clear that this will continue in all the existing product categories out there. To name just one example, Qualcomm continues to borg up the RF chain and integrate it with its other chips. Similar things are happening in networking, consumer electronics and beyond. The big question in many of these markets is to what extent will the Internet and Hardware companies adopt the same strategies with their own silicon? Apple is clearly going down this path with its modem, but will it extend to Bluetooth and Wi-Fi connectivity or display drivers? (Probably). Will Amazon extend in the same way beyond its Alexa-powering processors? (Probably not)
Beyond that, this integration will become very important in new growth areas – notably AI and Autos. (IOT should fall into this group too, but it is too big a mess for us to address here.)
That being said, it seems to us that these two markets will go in different directions on this topic.
The big advance in AI is towards more purpose-built silicon (AI accelerators aka TPUs aka DPUs). The whole point of these chips is to be really good at one thing. At least in the data center, but in consumer devices we are already well down the integration path – take the latest Apple or Oppo apps processors which have dedicated AI blocks built-in already.
So for us, the really big question will be what happens to auto semis. This is going to be the big fight of the next decade. Everyone agrees that autos of the future will have a lot more semiconductors in them. But beyond that, no one can quite agree exactly which semiconductors those will be. The Big Goal here is of course autonomy. Today’s autonomous vehicles (AV) are essentially data center racks with tens of thousands of dollars of CPUs and GPUs. But when these become commercial, a key sticking point will be reducing the size and critically the power consumption of autonomous systems. The only way to square that circle will be integrating down to a much smaller number of chips. This will play out across other auto systems as well – Lidar for one is already struggling through this process, but ADAS, infotainment and all the rest will as well.
And no one can say today whose chips will end up with the strategic high ground. Going by history, it will be the company closest to that critical software. If OEMs had to pick that solution today it would be Nvidia or possibly Intel’s MobilEye, but the fight is still very far from over.