Revolutionizing Automotive AI: Insights from Ideal’s CTO on New Computing Architectures for the Future

Xie Yan, the CTO of Li Auto, emphasized the necessity for a new computing architecture in the age of AI, particularly in the automotive sector. As the company approaches the mass production of its self-developed chip, the Mach M100, Xie highlighted that the current achievement rate of their business goals stands at only 60%. He stated, “True success is when the L9, equipped with the Mach M100, achieves the industry’s best autonomous driving capabilities and sells well. Only then can we confidently say we have met our business targets.”

Li Auto initiated its chip development project in 2022, investing billions to create a custom chip designed to enhance the intelligent driving experience. Xie Yan, who joined Li Auto after being introduced by Meituan’s founder Wang Xing, has been instrumental in this project since his promotion to CTO six months later.

Historically, American computer scientist Alan Kay asserted that to take software seriously, one must also create their own hardware. Following this principle, Li Auto has committed to developing its own chips, which has become increasingly important as competition in the electric vehicle market intensifies. As new players in the industry reach their tenth year, product homogeneity and pricing pressures complicate the landscape, making large investments in custom chips imperative.

Typically, inexperienced companies begin with smaller chips to minimize costs while testing designs; however, Xie argued against starting small. He insisted on tackling the most critical issues directly, aiming for the first self-developed chip to outperform the best Nvidia chip available at that time. Instead of following the mainstream GPGPU route to catch up with Nvidia, he opted for a dataflow architecture.

According to Li Auto, the Mach M100 achieves an effective computing power of 1280 TOPS, which is three times that of Nvidia’s Thor-U, thanks to optimizations made possible by the dataflow architecture developed by several professors at MIT in the 1970s. Xie first encountered this concept during his graduate studies at the University of Delaware and believes it is more aligned with the needs of large-scale AI computing compared to the von Neumann architecture.

Despite its advantages, dataflow architecture saw limited commercial application until now, primarily due to the required scale for its benefits to become evident. Xie noted that the efficiency losses associated with centralized scheduling and extensive data movement become more pronounced as system size increases, while dataflow architecture allows for data-driven computation, theoretically reducing waiting times.

However, this shift in architecture places additional demands on software, compilers, and hardware-software coordination. Chen Yiran, a professor at Duke University’s Department of Electrical and Computer Engineering, acknowledged the promising engineering attempt represented by the Mach M100 but cautioned against assuming that dataflow architecture is the future direction—it depends on specific designs and their synergy with application and software systems.

This collaboration is essential, especially given the rapid evolution of autonomous driving algorithms. The architecture must maintain sufficient flexibility across various computing paradigms, such as CNNs and Transformers. Xie acknowledged the complexity of the software involved in the Mach M100’s development, which has increased due to the more complex hardware design.

As the team continues to refine their processes, they aim to fully leverage the chip’s hardware performance using software tools like compilers. The Li Auto chip team, which has grown from a small group to about 200 people, remains focused on core developments without being swayed by fluctuating monthly sales.

In a recent conversation, Xie discussed various aspects of the Mach M100’s development, including its testing phases and the implications of achieving its performance goals. He emphasized the importance of comprehensive testing before declaring success, contrasting with the common practice of announcing success immediately after initial tests.

While addressing the challenges of integrating the chip into vehicles, Xie expressed his belief that achieving the best autonomous driving capabilities is paramount. He noted that maintaining strong performance and cost competitiveness is essential for their self-developed chips to be justified.

As they look to the future, Xie is optimistic that the Mach M100 will serve as a foundation for Li Auto’s broader ambitions, including the development of intelligent robots and AI systems, highlighting the significance of integrating AI capabilities into their vehicles.

In conclusion, Xie Yan’s vision for Li Auto is clear: by harnessing new computing architectures and maintaining a focus on innovative design, the company aims to lead the charge in the evolving landscape of automotive technology.