ZF Builds Its Computational-Power Assets
Martin Fischer, ZF board member and head of North American electronics, ADAS and safety initiatives, discusses the future of onboard computing.
Even by industry norms, Martin Fischer (right) has a lengthy job title. He is a member of the board of management for ZF and his responsibilities include electronics and ADAS, passive safety systems, active safety systems North America Region and South America Region corporate quality. In more plain terms, Fischer is at the tip of the spear for the industry’s fifth-largest supplier’s strategy to address the explosive transformation to electronic- and software-defined passenger and commercial vehicles. He recently spoke with SAE editorial director Bill Visnic about ZF’s new-generation vehicle supercomputer, and the auto sector’s vexing struggle to secure crucial semiconductor supplies.
What’s the takeaway regarding the new generation of ProAI, your central supercomputing platform?
We're talking about serious level-definition of the product. And what drove us there really is two things: we see the software-defined vehicle, we see the architectures changing more to domain and zone computers.
And at the same time, we ourselves have high computational demands when we think about our ADAS [automated driver-assistance systems] and AD [automated driving] systems. That made us conclude we’d better define a computer for that – and one of the attributes that we laid out early in the work was that it should be a scalable computer. Because if you do a Level 2-plus ADAS system and you go all the way up to a Level 4 system, there is quite some level of complexity being added.
And then if you look at the market, quite a few players need these computers. From our new automotive customer friends and startups to the OEMs, the well-established ones, so we thought if we make it modular and scalable, that's going to help drive a business case across the space.
It's all about really aggregating the computational power rate; that's the trend on these new architectures. Fewer ECUs, more bundling of functionalities, and then you start seeing that a better-performing computer makes sense. Let's say where it’s really starting from is from the top end, ADAS and AD – that's where we need heavy computation. But we have started with a modular design where we basically can exchange the ‘core.’ So, we started off with Nvidia [chips] and have really powerful SoCs [System on a Chip]. But we can take that out and plug in less-performing chips as well. That will then fulfill the lower needs of some zone-controller applications, or the main domain applications that need less computational efforts.
The global computer chip shortage probably has been the auto industry’s seminal story this year. The newest ProAI isn’t due to be in production until 2024, but with the rollout of such a compute-power-centric product, is the current chip-supply situation a concern for you?
Yes – for this product and the general trend. You can easily in the next five years probably expect to double the electronic contents in the vehicles. That's an issue. We are small fish in the whole semiconductor world, 5-10 percent or so of TSMC [Taiwan Semiconductor Manufacturing Company] but we are growing. So, it is a concern in how sustainably we can generate, create enough capacity – and also secure our share in that. If not, the next-generation Playstations set us back and, oh well, there are a couple hundred thousand vehicles [that can’t be built].
You are a significant semiconductor customer at this point. How do you secure that? Is it just a matter of saying to a chip supplier, “Listen, we promise you a certain amount of business so we expect a certain amount of service for that,” – or is it more complicated?
I think that is one of the key and the basic elements we have to go through. And I would say, this is short-term securing, really. The good news is that we are at the table for that and there is a good common understanding of that approach between the OEMs, the Tier 1s and the semiconductor makers. All of us are very busy and now figuring out 2022, first of all. And that's quite something, right? That means it's an outlook of more than 18 months; there has to be certain reliability and binding commitments in that.
And there are still also the bigger questions. How are we going to, in longer terms, also secure the situation geopolitically. Governments are now aware. Chips are important for so many industries and we will have to see how we deal with it from Europe and from the U.S.
You are the customer, it's really their supply chain to work out. But at the same time, you have work as something of a collaborator with them on this, correct?
We are going to be much more aware and conscious of where we source from. And that's not only one tier below us – it's two and three tiers below us. At minimum, we have to be really aware of what risks are from certain constellations and supply chains across the world. Impacting that is obviously the next step. And that's not easy for a single Tier 1 company like ZF to do. There are a lot of political things that have to happen – and maybe that have to happen also across the automotive industry.
There’s ongoing discussion from various levels of the supply chain about whether it's more sensible and efficient to relieve your perception-system sensors of computational processing and centralize that job. Is it better to use sensors almost strictly for the perception task and take image processing to a central computer?
In general, that's the trend that we are seeing. In particular, when you have Level 2, Level 2-plus systems where a little bit of computational effort is required. If we go to the few basic minimum front camera applications with just the mandatory safety functions – and of course still be a device where you have a camera that's smart enough to compute and give some simple signals back. But as soon as we start integrating the radar sensors with it, I think it's a smart concept to have camera ‘hats’ and radar ‘hats’ and do most of the computation in the central ECU.
And fusion gets a new character itself. The good old way is doing an object-based fusion. You get objects recognized from your smart camera and you have your radar objects basically, and then you fuse those. If it's all centralized then you can work from the direct data and you have point clouds, for instance, also from lidar sensors. And you can fuse that raw data and then apply the intelligence to it. That allows for new opportunities.
What's the advantage of installing a high-capacity central-processing platform like ProAI for vehicle architectures with comparatively low ADAS/AD functionality? Is it helping you to prepare for future electrification of subsystems or as everything kind of moves towards battery electric vehicle (BEV) architectures?
It’s scaling. I mean, if we were in a compact car with a 250 TOPS [Trillion Operations Per Second] ProAI, that car gets an extreme price increase. So that's not going to happen. But general trends of merging the systems, consolidating electronics, that's going to take place.
One example: we are very active in vehicle motion-control and chassis applications. In our active-safety divisions, we do steering, we do braking. In the next technology waves, there is going to be steer-by-wire and brake-by-wire. And we are now to a point that five years down the road, those systems are going to be price-competitive with the conventional ones. So now we have enough [compute] power on board to do these things electrically. That brings us to a cost equilibrium.
When you now start thinking through the systems, you can start integrating their electronics as well. When you go to these safety applications, you always need redundancy. If you now merge systems – where the primary brake controller also can be the backup control of the steering, and vice versa – you get really nice architectural savings. So, these kinds of architectural optimizations, that's something that drives us right now.