Li Auto CTO Xie Yan Speaks at TSMC OIP Forum, Explaining the Full Journey of Mach M100 from Technology to Application
On November 17, 2026, the TSMC OIP (Open Innovation Platform) Ecosystem Forum was held in Beijing. As one of the most influential annual technology events in the global semiconductor design, EDA, IP, and packaging and testing fields, this year's forum brought together top companies and experts in chip architecture, AI computing, and system innovation. Li Auto Chief Technology Officer Xie Yan was invited to deliver a keynote speech, making him the only specially invited speaker at this forum, and Li Auto became the first mainland Chinese company to give a specially invited speech since the inception of the TSMC OIP Ecosystem Forum.

Xie Yan's speech was titled "Design Thinking and Practice for Efficient AI Inference Systems," with the core content centered on Li Auto's self-developed Mach M100 chip. From architecture design to mass production and vehicle installation, this chip has completed the full journey from original technology to large-scale application.
The Mach M100 is the world's first mass-produced AI chip with a dynamic dataflow architecture, using a 5nm automotive-grade process and delivering 1280 TOPS of computing power per chip. Unlike most dedicated chips on the market that serve only autonomous driving, it is positioned as an edge inference chip built for the AI era—it has already been deployed across all intelligent scenarios in the vehicle, capable of running large language models, supporting agent applications, driving intelligent assisted driving, and reserving space for more AI scenarios in the future. The Mach M100 has now entered mass production and delivery, equipping models including the all-new Li L9, all-new Li L8, new-generation Li L6, new-generation Li MEGA, all-new Li i9, and the 2026 Li i6.

The core breakthrough of the Mach M100 lies in its dynamic dataflow architecture. This is a computer architecture fundamentally different from the von Neumann architecture, with the core idea of being "data-driven" rather than "instruction-driven."
When Li Auto started chip design in 2022, it already judged that merely building a faster chip would not solve the industry's fundamental contradiction. Moore's Law and Dennard scaling have both slowed, while demand for AI computing power has continued to surge from the deep learning revolution to the Agent era. The von Neumann architecture has dominated the computing industry for seventy years, with a large number of transistors used for management overhead such as caching, scheduling, and branch prediction, making it difficult to improve efficiency in the face of AI's massive parallel computing. AI computing is inherently parallel: data is tensors, relationships are deterministic, and flow paths are clear. Based on this judgment, Li Auto chose to let the flow of data directly drive computation, designing the architecture around the computational form of AI.

In AI inference scenarios, this architecture demonstrates three core advantages. Direct data access reduces data movement waste, with data passed directly between compute units, eliminating frequent round trips to memory; pipelined efficient execution, with multiple compute units operating simultaneously, greatly shortening inference time; and global broadcast capability, where the on-chip high-speed bus can deliver the same piece of data to multiple compute units at once, improving parallel efficiency.
Together, these three allow the Mach M100 to devote the vast majority of hardware resources to actual computation rather than scheduling overhead. The Mach M100 chose the dynamic dataflow route, maintaining high efficiency while offering full programmability—when AI algorithms iterate, only recompilation and adaptation are needed, with no hardware replacement required.
Integrated software-hardware co-design is also a core competitive strength of the Mach M100. The chip and algorithms worked closely together from the very beginning, rather than being developed separately and pieced together at the end. Li Auto's internal team fully understands the demand direction and long-term plans of its own products, enabling the most precise technical judgments at the very start of chip definition.
This year, Li Auto's foundation model team also jointly developed with the National Institute of Innovation for Decision Intelligence a software-hardware co-design law for edge large language models, using mathematical methods to make chip-algorithm collaboration more efficient and rigorous.
Of course, technological leadership also needs to be translated into user experience. On November 6, 2026, Li Auto pushed OTA9.2, officially launching Mach VLA 2.5, with the most notable feature being the all-new 3D ViT perception model. 3D ViT, or three-dimensional vision Transformer encoder, is regarded as the "eyes" in Li Auto's embodied intelligence technology system. In the past, visual perception was like looking at the world through a 2D screen, requiring inference of three-dimensional space from flat images; 3D ViT fuses 3D spatial geometry and semantic information at the encoding stage, allowing the model to simultaneously understand the position, category, and spatial relationships of objects. Its forward perception range reaches 300 meters, with equivalent 1080-line resolution and a 50% improvement in dynamic vision.

At the user level, Mach VLA 2.5 brings a reverse yielding feature, and also shows more human-like behavior in scenarios such as construction detours, animal avoidance, blind-spot passage, and defensive driving. The large-scale deployment of 3D ViT on the vehicle is backed by the Mach M100 chip. Li Auto's previous-generation computing platform was limited by memory bandwidth and data movement efficiency and could not natively support 3D ViT. Relying on the Mach M100's dynamic dataflow architecture, this model was finally able to be brought onto the vehicle.
The technical value of the Mach M100 has been validated across multiple dimensions. In early 2026, Li Auto's Mach M100 chip paper was accepted into the ISCA industry track, making Li Auto the first automotive company in the world to be accepted since the track was established. The capital market has also responded positively: in October 2026, Li Auto's chip subsidiary Machxin completed its first round of financing, raising 1.625 billion yuan, with a maximum post-financing valuation of 17.1 billion yuan.
Overall, the Mach M100 is undergoing a complete cycle from technological breakthrough to user value to industry recognition. For Li Auto, the significance of the Mach M100 goes beyond a computing power chip—it represents an automotive company's deep commitment in the field of underlying computing technology, and offers a brand-new possibility for the computing foundation of intelligent vehicles.
Related Articles
- Li Auto CEO Li Xiang: Will Gradually Open External Supply of Mach M100 Chips
- Reports Say TSMC and Other Foundries Are Raising Prices, with Increases Ranging from 10% to Over 20%
- Li Auto Legal Department: Rumors of "Traction Battery Recall" and "4,000 Yuan Compensation" Are Fabricated; Attached URLs Are Scam Websites
- AI Circular Financing Doubts Intensify; TSMC Plans to Supply Capacity Based on Each Company's Credit Line
- TSMC Outsources CoWoS Advanced Packaging oS Capacity to Xintec, Which Will Build Production Lines at TSMC Facilities
- Seeking "the Next 英伟达": TSMC Plans to Establish a $500 Million Venture Capital Fund