In August last year, Wu Xinzhou left Xpeng to join Nvidia, and by the end of the year, Wu Xinzhou returned to China as Nvidia's global vice president and head of the automotive division, and began a large-scale recruitment plan for autonomous driving talent in the Chinese market. This is seen as a new round of strategic expansion of NVIDIA's intelligent driving business in the Chinese market.
As one of the regions with the fastest development of intelligent driving in the world, the accumulation of related products, technologies and experience of Chinese car companies has been at the forefront of the world, including the recent introduction of Tesla's FSD into China, which also proves the commercial soil and imagination of autonomous driving in the Chinese market.
For NVIDIA, expanding its territory in China and the global market is of course an important task at the moment. However, NVIDIA also continued its usual advanced style, making more strategic deployments in advance from the perspective of future autonomous driving development, and finally making this route the consensus of the industry.
At the recent media communication conference of NVIDIA, in addition to Wu Xinzhou, Norm Marks, vice president of data center of NVIDIA's automotive division, and Liu Tong, global vice president of NVIDIA and head of China's automotive division, were also present to introduce the technology and commercialization progress of NVIDIA's intelligent driving in detail to Cheyun and other media.
Judging from the information conveyed by this communication meeting, the fierce competition and huge demand in the Chinese market have made NVIDIA see more opportunities and imagination in the automotive field;
On the other hand, NVIDIA also hopes to use its leading technology and experience in the field of AI, combined with product development and landing experience in the Chinese market, to provide car companies with a complete end-to-end intelligent driving full-stack solution, not limited to hardware, and help NVIDIA occupy a position in the global market in advance in the future-oriented intelligent driving competition.
In this process, in addition to "empowerment" and "competition", Nvidia hopes to play the role of "leading".
AI-defined cars are the trend of the future
In the past decade, "software-defined vehicles" have had a very far-reaching impact in the automotive field, especially through OTA technology, turning a fixed hardware product such as a car into a self-learning and ever-changing user product.
With the development of AI, software-defined vehicles are becoming more and more mature, and the number of corner cases to be solved is also increasing. Generative AI and LLMs will have the potential to solve these more complex problems, and autonomous driving will usher in a new development paradigm. In the future, when AI defines automobiles, the focus of enterprise R&D will shift to the cloud to create a truly human-like high-end autonomous driving technology stack.
In Wu Xinzhou's view, the two stages of software-defined vehicles and AI-defined vehicles are actually in the same line. From the perspective of the development process of autonomous driving, the first generation of autonomous driving systems is completely rule-based and has a large number of engineer characteristics; At this stage, the second generation of autonomous driving has begun to replace the original engineer characteristics with a large number of AI, and some predictions and planning can be done through models. And then there is the end-to-end big model approach, which is also what Nvidia is currently doing in advance.
The difference between the two is that software-defined vehicles require a large number of engineers. For example, the recognition of corner cases is relatively slow, and a large amount of data is required to be tested and operated, and then the data is sent back to engineers, which requires very efficient infrastructure to support the closed-loop data facilities, and then update them through OTA. In short, each step requires a lot of human involvement.
In the new stage of AI-defined vehicles, most models can be trained in the cloud, identified by data-driven vehicles, and then self-iterated by automatic data-driven, and most of the verification can be completed through simulation in the cloud, greatly reducing the dependence on large-scale equipment deployment and testing on the vehicle side, and greatly simplifying the process. Similarly, as generative AI capabilities become more advanced, the cockpit parts, including cloud conversations, task test planning, and editing, will also be much simpler.
As one of the hottest technical routes in the intelligent driving track, although many companies emphasize their end-to-end layout and investment, they also face many unsolved technical challenges, such as the problem of black box. Wu Xinzhou believes that these problems will be gradually resolved in the future. For example, the future large model and end-to-end model have peripheral output points that can be observed, which is equivalent to opening several windows on the black box to see what kind of signal the signal is.
Although there are still some challenges at this stage, Wu Xinzhou firmly believes that end-to-end autonomous driving will be realized within five years. But he also stressed that this is also a very difficult process, not from pixels to actions as everyone literally understands, but requires a lot of other technologies, so it is impossible to launch an end-to-end model from the beginning.
Therefore, before the end-to-end model goes live, there must be a "guardrail", because this model needs to be constantly optimized and grown. Just like the process of a person growing from a primary school student to a doctoral student, the end-to-end model also needs a very good second-generation or even first-generation autonomous driving stack that can have time to grow and become more powerful.
"The trend in the next few years is that the end-to-end model and the original model complement each other, and in some cases, such as dealing with more difficult intersections, it can show more anthropomorphic things, and ensure safety through the original model and method, which is the process of making the real large-scale deployment of the end-to-end model mainstream." Wu Xinzhou concluded.
The role that NVIDIA wants to play: enabler, competitor, and leader
At the Beijing Auto Show just past, although Nvidia did not open a booth, Nvidia was completely unable to ignore the existence of AI and autonomous driving as the key words in this auto show. In the highly homogeneous competition, whether high-end intelligent driving is equipped with NVIDIA's latest products has even become one of the selling points of the differentiated competition of car companies.
According to NVIDIA's official information, including Wei Xiaoli, BYD, Yangwang, Lotus GT Emeya, smart Elf #5, Zhiji L6, Haobo, ZEEKR 007, Xiaomi SU7, Wei brand Blue Mountain and other car company models, have begun to carry DRIVE Orin to create mass production cars.
Compared with Orin, which has already been widely implemented, NVIDIA's next-generation DRIVE Thor centralized in-vehicle computing platform has shown greater commercial momentum. The platform integrates the new NVIDIA Blackwell GPU architecture, which is designed for Transformers, large language models, and generative AI workloads.
Although DRIVE Thor will only start the first generation of SOP next year, since last year, there have been Geely, Xpeng, Li, Chery, Jiyue, GAC, BYD and other car companies have announced the use of NVIDIA's next-generation DRIVE Thor for the development and design of future L3 and above intelligent driving.
Wu Xinzhou said that DRIVE Thor is still growing rapidly, which is an unstoppable trend because it represents not only the next generation of chips with the highest computing power, but also the next generation of chips with the highest level of security, while being able to give the best support for generative AI and LLMs.
Among so many partners, including Wei Xiaoli, BYD, Great Wall, Geely and many other powerful car companies, they are also deploying self-developed intelligent driving chips. There are strategic considerations and trade-offs at the practical level. The competition and cooperation relationship between car companies and suppliers is also the focus of the industry.
Unlike many domestic suppliers of intelligent driving chips, Nvidia will have more problems to consider in the face of the Chinese market. As Huang mentioned earlier, Nvidia chips are not irreplaceable in the Chinese market, and if they can't be bought, then China will develop and produce them themselves. This is not without precedent, and of course, this is also a concern for Nvidia.
Wu Xinzhou positioned NVIDIA as an AI ecological enabler, rather than a company focusing on making cars, he said that most of NVIDIA's chips are based on general-purpose GPUs, but NVIDIA is also constantly adjusting and iterating, optimizing the overall hardware architecture, and then applying the most cutting-edge technology in the AI field to the vehicle-end chip to better empower the automotive and robotics industries, which is NVIDIA's advantage.
Wu Xinzhou deliberately emphasized security, which requires a lot of investment and practical experience, especially end-to-end security, including the security of chip and operating system modules, the security of all training tools in the cloud, and truly deploying software to vehicle-end chips through an efficient and secure network. Nvidia hopes to empower car companies to achieve safety with less investment, which is also Nvidia's advantage.
In addition to the well-known SoC, NVIDIA has also deployed a number of core basic capabilities, including data centers, tool training, automotive security platforms, and end-to-end full-stack solutions. For example, NVIDIA's Omniverse can help car companies achieve comprehensive empowerment from vehicle design, manufacturing, testing and verification, and even sales experience. These software capabilities are also the focus of NVIDIA's hope to be widely used in the future.
This can also be seen from NVIDIA's intelligent driving plan: "The first step is to hope that our software will reach the market-leading level or the first echelon level on the existing L2 and L2+ systems as soon as possible; The second step is to hope to make some new breakthroughs in the field of L2++, completely connect the upstream and downstream modules, and also do end-to-end training, so as to truly achieve the industry-leading level; The third step is to completely remove people from the system with L3, which is expected to be mass-produced in 2026, which is where the real value of autonomous driving lies. ”
Wu Xinzhou believes that these plans are based on NVIDIA's huge advantages and are completely feasible, and these three-step plans are the process of NVIDIA's market competitiveness and business imagination to improve step by step, and it is also the process of NVIDIA's realization of its goals.
"Our goal is to allow everyone to play with their mobile phones instead of driving in the car, which is what everyone needs. Driving is not a rigid need, going from point A to point B is a rigid need, and using a mobile phone is also a rigid need. ”
Adjustments to NVIDIA's business model
Although end-to-end has almost become the label of Tesla's intelligent driving, if you look at the timeline, the time when the NVIDIA autonomous driving team proposed end-to-end intelligent driving and made a demo is not later than Tesla. It's just that it was too advanced at that time, and the technology and market were not mature, so it was not able to successfully produce it commercially.
Of course, there is another very important factor, Nvidia is known for "selling cards", and the autonomous driving business many years ago was only a very inconspicuous part of Nvidia's automotive business, so the end-to-end problem at that time was at best just one of the countless episodes on its way forward.
But Huang is very serious about the auto business, and at the 2022 GTC, Huang said that according to his plan, when Nvidia's revenue reaches a trillion US dollars, the automotive business should account for 30% of the revenue, or $300 billion. However, judging from Nvidia's financial report, until Q4 last year, the automotive business was only 1.3% of Nvidia's overall revenue.
In the past two years, with the explosion of the intelligent driving business, in China's high-end intelligent driving market, NVIDIA's intelligent driving chip share once exceeded eighty. However, NVIDIA's market share in software related to intelligent driving does not match the strong position of its chips at all.
Last year, Wu Xinzhou's joining was interpreted by the outside world as Nvidia to focus on intelligent driving software capabilities and overall solutions in addition to "selling cards", and its primary task was to help Nvidia and Mercedes-Benz cooperate in the project.
In June 2020, NVIDIA and Mercedes-Benz officially announced their cooperation, and NVIDIA will provide AI software architecture for Mercedes-Benz's next-generation models, including autonomous driving software solutions, intelligent cockpits, etc. The reason why the project caused a sensation in the industry at the time was that the two parties chose a new business model, that is, a share of the sales of the whole vehicle based on the joint model, rather than the traditional supplier procurement model. If this model can work, this is not only NVIDIA, but also a benchmark case for the entire automotive industry.
However, this much-anticipated project has encountered many crises, and there is no mass production delivery timetable so far. Previously, it was said that Mercedes-Benz introduced Momenta to fight fires, or Wu Xinzhou's team will deliver solutions in the middle of this year, etc., but there is little official news amid the disturbance.
However, it is certain that from the project, it can be seen that NVIDIA is trying to seek to provide car companies with a complete end-to-end intelligent driving full-stack solution in addition to hardware, which can obviously help Nvidia achieve greater business imagination and go one step closer to 300 billion.
Today, in the Chinese market, this effort has also made great progress. At the communication meeting, Liu Tong introduced the end-to-end full-stack cooperation with BYD in the Chinese market, including Orin and Thor intelligent driving chips, data center solutions, NVIDIA AI Enterprise, a software series product for AI development and autonomous driving algorithm development, as well as a full range of cooperation in smart factory robots and Omniverse simulation software, which is also the most complete cooperation representative case of NVIDIA's intelligent driving full-stack solution.
In this process, the decline in software and hardware costs, the continuous iteration of technology, the decoupling of software and hardware, and the integration of cabin and driving, etc., will become the key points of competition for car companies.
This also means that the scale and commercialization speed of intelligent driving suppliers at this stage will largely determine the market pattern in the next few years, which is also an important reason for NVIDIA and more intelligent driving suppliers and car companies to accelerate the layout and competition.