In our previous post, we walked through the current and past state of the market for automotive semis. Here, we are going to dig into some of the products and vendors, and then look forward a bit.
Part of the problem of thinking about semis in cars is that we have been reading too much science fiction. Everyone is fascinated by Autonomous Vehicles (AV) and have visions of AI grandeur. In reality, the typical car today has around $300-$500 of semis in each one, and those are not glamorous or even particularly expensive parts. Cars today do not have CPUs or GPUs let alone futuristic AI engines. Instead, think $5 microcontrollers for managing power windows and lots of $0.50 tiny components all over the engine and in the dash, doing very mundane tasks.
That seems likely to change. AVs will require some pretty powerful systems. The AV test systems in cars run by companies like Cruise blocking all the intersections in San Francisco have multiple GPUs and CPUs. This is not sustainable. As much as some people like to talk about cars becoming “data centers on wheels”, the amount of power required to run half a server rack in today’s test systems is a non-starter for electric vehicles, and ruinously expensive for any car. As we noted in our last post, the industry has no idea how much consumers will be willing to pay for AV, but adding $10,000 of compute to a $30,000 Honda Civic is not going to happen. Much of this will be solved by moving to purpose-built AV engines (i.e. AI ASICs). There is intense competition by the big chip companies to supply these, but the industry is still in very early stages of building the AI models to drive these cars, so we cannot even define the requirements for these ASICs, let alone build them in quantity.
Put simply, widespread AV adoption is still years away, and before we get there the computing needs of cars are going to continue to grow for a few reasons. First, as much as there are already lots of chips in cars, there are many more that can be used. The principal driver of this is consumers growing preference for more electronic features. Consumers, especially younger car buyers, want their cars’ controls to look a lot more like smartphones. There are also a lot of components that could do with a higher degree of coordination, think sensors and electrical power systems. In addition, there is a lot of interest in replacing wiring (which is heavy) with wireless systems. While critical systems will need to be hard-wired for safety reasons, there are a lot of other things that could live without those. And probably the biggest factor is the increasing number of electric vehicles on the road – designers of these are always looking for components that use less power, and every advanced chip helps.
If we take a step back, all those electronics could likely benefit from some form of powerful digital logic. This is one of those products that some will some day have a standard name like CPU or Mobile Applications Processor (AP). We have heard it called an Automotive Processing Unit (APU), but it does not seem like that name has been widely adopted. But what kind of chip will this be?
An APU will need to do a few key tasks: coordinate all (most of) the car’s electronic systems; drive the central console and probably the dashboard; run infotainment systems (another term that could use an upgraded name); process the increasing number of sensors appearing in cars (proximity radars now, possibly Lidar further down the line). This is a general purpose compute engine. So why not use a CPU? For most of these uses, a CPU is probably overkill – too much high-end math and too much power consumption. More importantly, APUs are going to need to do a few very specific tasks really well. This means the chip will need to be purpose-built, an ASIC, which will be comprised of multiple systems, so System on a Chip or SoC. This is an important design trend in semis today. Apple’s A Series phone processors is the best example – it has a CPU, a GPU, and a lot of fancy video processing all in one. The APU will likely look very similar, but with a different mix of systems.
When we think about it in this light, it seems very likely that future APUs will resemble mobile processors in many ways. CPUs will draw too much power and lack a lot the auto-specific features. The same is true for GPUs. So who is going to build these?
One likely candidate is the auto maker itself. This is the model Tesla is taking, building their own chip – the D1, although this seems much more geared to Tesla’s aspirational AV features. We believe some of the AV start-ups are also building their own chips, and some of the incumbent auto OEMs are working with various partners as well.
Last month reports emerged that two Chinese auto makers have (maybe) started to build their own chips. First came reports that electric vehicle maker Nio had hired a senior Xilinx engineer, and then a week later reports emerged that Geely was also considering building its own chips (and as we went to print this seems to have been confirmed). Both of these sources are framed in highly conditional terms, with a big ‘If’ in front of them and an unspoken question mark at the end. And as we noted last week, there is not a lot of clarity in China in terms of who is building a chip and what that exactly means. Nonetheless, there are a lot of reasons why it makes sense for these companies to build their own chips (not least as a sign of their patriotic participation in China’s semis initiative).
That being said, there is no standard model for this. There is no operating system for cars, and so no standard chip requirements. Every vehicle maker is essentially pursuing their own path. True, there are initiatives from companies like Blackberry and industry organizations like the Linux Foundation to change that, but there is still a very high degree of individual effort produced by each car maker.
And by now, it should be clear that not every company can (or should) build its own chip. So what will everyone else do? The industry is going to need a merchant chip maker. Someone who can build highly power-efficient SoCs, capable of running multiple compute engines, coordinating across multiple communications streams. Someone unburdened by producing legacy CPUs or GPUs. When viewed in these terms the best positioned company to capture the medium term demand for these kinds of chips is Qualcomm. While Nvidia and Intel both have big ambitions here, it is not clear that they have the right products. They could build them, but it would take time. Qualcomm already has everything required. Marvell and Broadcom could probably do this as well, but they do not seem to have exerted much effort down this path. And it is hard to think of anyone else who has a sufficient background in building processors.
We still have a long way to go, and a lot could change. The industry is in early stages, and it is clear there will be a lot of roll-your-own chip work done by the auto makers, but it is also likely that there will be room for a merchant vendor, and that may give the merchant vendor an important set of relationships for when the even larger AV opportunity finally arrives.