Core Viewpoint - The collaboration between Li Auto and NVIDIA has led to significant advancements in GPU optimization, particularly in enabling the Orin chip to run large language models (LLMs) through a deep architectural alignment and the reconstruction of the underlying operators based on the PTX instruction set [1][2]. Group 1 - Li Auto's work in optimizing the Orin chip is recognized as a high-level achievement, demonstrating the company's capability to push the limits of chip performance [1]. - The collaboration with NVIDIA involved obtaining original manufacturer-level technical guidance at the microarchitecture level of the Orin chip, which allowed for a more granular control over instruction pipelines and data lifecycles [1]. - The transition from conventional CUDA C++ development to PTX/SASS development signifies a shift towards more direct hardware instruction, enhancing register reuse efficiency and mitigating register spilling issues on the Orin architecture [2]. Group 2 - Li Auto's team has shown a strong capability in analyzing SASS and extracting hardware potential at the instruction level, indicating a high barrier to entry in this field [2]. - NVIDIA's positioning as an acceleration computing company rather than just a chip manufacturer highlights the strategic importance of software and hardware integration in achieving optimal performance [2].