en.Wedoany.com Reported - China's Li Auto has publicly released the Die Shot of its self-developed automotive-grade AI chip, the Mach M100, showcasing for the first time the internal transistor layout and module partitioning of its 5nm dataflow architecture chip.

From the die shot analysis, the Mach M100 die area is approximately 400 square millimeters. The left side of the chip integrates a 24-core A78AE CPU cluster, an 8-channel LPDDR5X memory controller, and multiple I/O modules. The right side, occupying nearly half the area, features a 56-core NPU computing array composed of 14 clusters and 56 independent Tile units, each equipped with 2MB of high-speed SRAM cache. The chip adopts a proprietary dynamic dataflow architecture, allowing data to be directly transmitted to the computing units to reduce data movement overhead.

The Mach M100 is the world's first automotive-grade AI chip to employ a dynamic dataflow architecture. Based on a 5nm automotive process, the single-chip AI computing power reaches 1280 TOPS with a computing power utilization rate of 82%. When combined in a dual-chip configuration, the total computing power can reach 2560 TOPS, supporting full-scenario inference for in-vehicle intelligent driving, in-car large models, and smart agents. The chip is now in mass production and has been integrated into the new Li Auto L9, L8, and L6 models. The release of this die shot also confirms Li Auto's self-developed chip design philosophy of "large NPU, reduced traditional cache," leveraging architectural innovation to enhance the operational efficiency of in-vehicle AI, rather than following the industry mainstream approach of modifying GPUs.

Industry insiders point out that the public release of the complete die shot indicates an increased level of technical transparency in Li Auto's self-developed chip efforts. Its dataflow architecture has been validated at academic venues such as ISCA and the CCF Chip Conference, providing further evidence for domestic automakers in advancing the autonomy of high-computing-power in-vehicle chips.