Welcome. This is Muly Yudvir Singh and sitting with me is a special man, Lars Reger, who is global CTO for NXP. Lars, welcome to Autocar Professional. Very nice time for being here in India. The AI summit is going on but we've not spoken a lot about AI and cars for today. Is it because something or the other has not been done on the chip level as of yet?
Well, there is a key reason for all of that because there is a big difference between having tools like ChatGPT in the cloud. So I fire a question that gets loaded into a huge data center where you have a nuclear power plant running next to it to provide all the energy and that answer comes back. And from there, this big jump into building a robot, a trusted machine, a rolling robot in that case, a car. For the robot, it is much more important to be absolutely trustworthy, to be safe and secure because the moment where this machine is doing something that you don't want, it gets dangerous. Secondly, it is all about energy efficiency and only then it is about intelligence. So we are shifting now, and this is what you also seeing, your minister Ashwini Vaishnaw very nicely pitched it in Davos. He says we go now from these big models, big data center fixed installations into physical AI robotics building in my language. And for that you need much smaller models but you need to be able to 150% trust them. And this is what companies like NXP are doing. A lot of the others have done this big data center stuff and you see that shift in the industry now. You're coming to the rolling robot and there of course it gets into the different architectures. How do you do intelligence systems in this rolling skateboard? How do you do intelligence systems in the driver replacement? So take glass out of the seat and put a robot in. How do you do intelligence systems for all the other vehicle functions? And that is only starting now. Not because we are only now coming to that idea but it is so much more difficult to make this trustworthy and bring safety and security into the game.
Interesting. I mean, referring cars to rolling, I think autonomous cars is what you're referring to as rolling robots.
And you're bringing in the aspect of cyber security but not a lot of OEMs, even tier ones, are talking about cyber security especially here in India.
Because it's in a normal car. It's not a function that you are seeing. You only see cyber security when it's getting wrong, right? When your credit card is getting hacked or when someone is posting on your behalf on social media and you don't want to have it and so on. So this trust, both on functional safety, so your braking or your ABS systems never ever ever fail and if they fail, they don't create a fatality, but they slow down the car and park it, and cyber security. Your car never ever ever gets hacked and if it gets hacked it's shutting down the data transmission. It's recognizing it but it's not leaking all data. These type of functions are hard to realize. You need to have decades of deep know-how and that is what we are trying to lift over. And that is a comparatively un-researched area how you are combining these worlds. So NXP has invented 25 years ago these contactless NFC banking cards that are in your wallet. We are taking this know-how on crypto security, make post-quantum crypto systems out of its own systems that cannot be hacked even by quantum computers and bring that into vehicles. We do the same, 60% of our NXP electronics is in automotive since decades and we know how to build ABS and braking systems. So with these ALD functional safety systems, how do we map this and how do we add the right size of the brain to it? But what I mean by that is the misunderstanding in autonomous driving is like I'm sitting there and at the 18th birthday I'm handing my car keys to my son and say, 'Jonas, you are an intelligent system. You have seen now for 18 years how that is driving. Enjoy. Go driving.' This is not working. We are today sending the kids to a driving school, say, 'Okay, first you learn all the rules theoretically, then you learn to apply them practically on the road.' This has nothing to do with creativity or AI, it's just rule-based, deterministic, comply to the rules and only inside these corridors the creativity is allowed to happen. Rules, I mean in Germany you are driving on the right side of the road, in front of a red traffic light you are stopping. If there is a 50 sign you don't drive 80. And by the way in India you are not driving on the right side of the road. Yeah. So rule-based, in between I can accelerate, I can overtake, I can do all these type of things. That's my own creativity. But we are boxing in the brain. Boxing in the brain into a functional, safe and secure environment makes the robot. And it is not about having a super sophisticated 'paint me a picture' or 'compose me a song' type of question to ChatGPT. There is a huge difference.
Interesting. But then you know if I park cyber security at the moment and ask you, what you referring to is data. If in simple language if I put it in...
Data, data processing, data handling, yeah, absolutely.
How do we generate, we generate a lot of, I mean a lot of data in India but how do we process and how does, how would you have the access to it especially to enable advance in India. How have you answered that question?
There is this nice saying, 'data is the new oil.' It is not because oil if you put it into your cellar is not rotting. Oil is not outdating. Oil is not getting irrelevant. Your eyes are taking every day petabytes of data, raw data. Your brain gets only very very little data, the delta pictures from your eyes and only then processes data like is the guy on the other side nodding or shaking his head. So a few bytes that are coming out of this raw data. So having all data of entire India doesn't get me anything. Having the right and the relevant data, that's science and that is art. Having the right data for training AI systems is super important because if you have the wrong data, you can make the system react in the wrong way. I can very easily take all my European engineering CVs, dump it into an AI tool and say generate me or hire for me the next best engineer. As I have the majority male Caucasians in the early 30s, that tool very easily has a chance to get let's say gender specific and racist. I don't want that, I just want to have the best electronics engineer to be hired but if I feed these tools with the wrong data and that the same happens in traffic and so on, paying attention to artifacts and then it gets so dangerous because I don't know which artifacts the system has been on and is driving me off the road. I have to box it in. You stay on that lane whatever you do and then I need to work on getting the artifacts out of the picture. So having the right data, training in the right way, certifying models and then bringing it back into these tools which is a comparatively small task. So for example this NXP solution from Hailo, it's an AI accelerator, it's carrying 10 billion parameters like a large language model for all sorts of unconnected systems, cars, computer tomographs, washing machines. Together with these microprocessors from our side, functional safe, you bring these together, 7 watts energy consumption, very small, on the right set of data doing the magic. For example for in-cabin management, chassis AI, it's not yet on the level of autonomous driving type of capabilities that you need a little bit more horsepower but for a lot of the other robot building tasks on your rolling robot, that is already sufficient. And it is about condensing the right data into the right AI models and processing then raw data again. So the answer, it's all about data, it's correct but not enough.
Interesting. You know, we, you just showed us one of the solutions, right? On the other hand, the industry is today talking about moving to a central architecture for ECUs. And this would add one more layer of an ECU or an SOC if I say that right, wouldn't that complicate things?
So these architectures, these software-defined architecture, software-defined vehicles, software-defined robots and so on are nothing else but a cleanup of functions. So you are thinking normally in domains. You have a powertrain domain. You have a connectivity domain. You have a driver replacement or ADAS domain. You have an infotainment domain in your car. So functional blocks. These blocks are usually anchored in certain domain processors and they are connected via big networking central compute devices. By the way, this is a central compute device that we have developed here in Noida. 13 billion transistors, 5 nanometer technology. So this is one of these high functional safe automotive networking device. It's the spine or the cerebellum of the car. Now these devices are sitting central with a lot of horsepower and of course they need sensing at the corners of the car and they need processing for these various domain architectures. Some of the domain architectures can maybe even be taken over by such a device. But by and large you have this hub and spoke type of architecture. That's a huge simplification compared to having 150 control units in a luxury car from 5 years ago. Due to that cleanup, you can go to a couple of big ECUs on simpler software environments and with the right operating system middleware on top, you can also easily add functions or move a function from the central device with a middleware to a zonal device in the periphery on the same middleware. It's like your tablet and my desktop are running the same operating system and we have the same tools programs running on top. So we can shift that backward and forward. That architecture cleanup is absolutely needed. And then of course the key point is can you bring these companions and for example can I slot them via the PCI Express, can I slot them into these central units just like co-processors, just like add-ons. So you don't need to have a different control unit. But how can I enrich the existing software-defined architecture by a large language model capability in the system, not in the cloud. And that gives you of course huge advantages in data security because nothing has to leave the car. You can compute here and not sending every data to the cloud and you have a very fast turnaround time and you have very safe and reliable processing of your data as well.
