Li Auto is developing an AI inference chip for intelligent driving and a SiC power chip for motor controllers. An informed source told Chinese media outlet LatePost that Li Auto is currently assembling a team in Singapore for the research and development of SiC power chips. On LinkedIn, Li Auto recently posted five job vacancies in Singapore, including: General Manager, SiC Power Module Failure Analysis Expert, SiC Power Module Design Expert, SiC Power Module Process Expert, and SiC Power Module Electrical Design Expert.
As of late October, Li Auto‘s intelligent driving Chip R&D team has assembled in Shanghai, preparing for a trial production. In the chip industry, “trial production” refers to producing a small quantity of chips for testing purposes after the circuit design is completed by the R&D team. Currently, the overall personnel scale of Li Auto‘s chip team is over 160 people, distributed in Beijing, Shanghai, Silicon Valley, and Singapore.
According to LatePost, regarding the division of labor in Li Auto‘s chip development, Li Auto‘s internal team mainly focuses on front-end design, and the back-end physical design is currently outsourced, although Li Auto is also building its own back-end design team at the same time.
In July, Taiwan Times cited a Morgan Stanley research report stating that Li Auto intends to outsource the backend design of its intelligent driving chip to Taiwan AIChip Technologies. It is understood that once the backend design of the chip is finalized, it will be manufactured by TSMC. As of the time of writing, Li Auto has not officially confirmed this piece of information.
In June 2022, Chen Fei, the head of Li Auto‘s NPU R&D team, proposed at a recruitment event that future intelligent driving chips should have seven major features: high computing power, high scalability, easy software programming, high flexibility, low power consumption, high reliability, and low cost. In the field of intelligent driving chips, the industry is expected to achieve a performance improvement of 5-10 times compared to the existing mass-produced chips. He believes that current chips in the market cannot maximize performance yet. One important reason is that chip manufacturers have to consider generalization, designing and manufacturing chips for the entire industry, and each OEM (Original Equipment Manufacturer) may use different models, which can lead to some waste and cannot compare to chips that are tailormade. “There’s no way around it, the best performing chips available on the market are just those from Nvidia,” said Chen Fei.
Li Auto‘s CEO Li Xiang once explained at Li Auto‘s Spring 2023 Media Communication Meeting that, in terms of system development, automakers need to focus on how to reduce the cost. The cost of hardware for intelligent driving (including sensors like radar and cameras, and the computing platform) is $1,500 for Tesla, and $4,000 for Li Auto. “If Li Auto makes its own inference chips, it can achieve the same cost as Tesla, because the algorithm is in our own hands, and this includes the entire training platform and being able to make our own training chips.” A person close to Li Auto told LatePost that “If Li Auto does not promote the self R&D of intelligent driving chips in a short period of time, but continues to use Nvidia’s chips, it will not necessarily encounter a performance bottleneck any time soon, but Nvidia’s chips are expensive and there is a risk of supply interruption.”
At the moment, the shortage of cloud training chips for intelligent driving has become a real risk faced by many Chinese automakers and autonomous driving developers. On October 18, the U.S. Department of Commerce announced that it would expand the scope of restriction over the export of advanced chips to China. The restricted chip products include but are not limited to A100, A800, H100, H800, L40, L40S, and RTX4090. Among them, chips such as A100, A800, H100, and H800 are all to be used for cloud training of intelligent driving.
Developing proprietary chips for intelligent driving has its benefits and risks. Firstly, the technology is still evolving, and designs may quickly become obsolete. Secondly, developing high-performance chips like those for intelligent driving involves long development cycles, large investments, and high uncertainty. This can pose a significant financial burden on companies. Chips developed in-house are typically only used internally and may only be suitable for a few generations of models, challenging the sustainability of this business model.
Furthermore, the ultimate aim of automakers’ developing their own intelligent driving chips is to maximize commercial benefits. However, most automakers can’t compete with leading chip manufacturers in terms of chip shipments. While Tesla sets a good example where self-development can save certain costs and bring technical advantages, if the industry matures to the point where general solutions meet automakers’ needs, self-developed chips may lose their advantage.