PUBLISHER: ResearchInChina | PRODUCT CODE: 1845357
PUBLISHER: ResearchInChina | PRODUCT CODE: 1845357
Research on Autonomous Driving Domain Controllers: Monthly Penetration Rate Exceeded 30% for the First Time, and 700T+ Ultrahigh-compute Domain Controller Products Are Rapidly Installed in Vehicles
L2.5 Gradually Replaces L2 as the Entry-Level Standard Feature for Passenger Cars, Giving A Big Boost to Demand for Autonomous Driving Domain Controllers
L2.5 is gradually replacing L2 as an entry-level standard feature for passenger cars. The penetration rate of high-level intelligent driving in passenger cars in China will surge, entering the era of "universal intelligent driving". Moreover, some OEMs have proposed to reduce or stop developing new vehicle models below L2.5, promoting highway NOA to gradually become an entry-level standard configuration for new vehicles. We predict that the penetration rate of L2.5 intelligent driving will hit 35.6% in 2030, which will be 10 times that of 2024.
Thanks to the surging penetration rate of L2.5 highway NOA in the OEM market, the demand for Autonomous Driving Domain Controllers has boomed. According to statistics from ResearchInChina, from January to June 2025, 2.526 million sets of Autonomous Driving Domain Controllers were installed in passenger cars (excluding imports and exports) as a pre-installed standard configuration in China. Since 2023, the penetration rate of Autonomous Driving Domain Controllers has increased month by month. In June 2025, the penetration rate of Autonomous Driving Domain Controllers exceeded 30% for the first time, reaching 30.81%, compared with only 17.43% in the same period last year.
OEMs Adopt the "Hardware Unification, Layered Software Development" Model to Rapidly Boost the Penetration Rate of Intelligent Driving Systems.
In terms of intelligent driving systems, OEMs cooperate with multiple Tier 1 suppliers and divide intelligent driving solutions into different types according to the configuration of hardware and software. Examples include BYD's "God's Eye", Geely's "G-Pilot", Chery's "Falcon Pilot", and GAC's "ADiGO". Through the model of hardware unification and layered software development, these solutions cover all their models of different price ranges, meeting the differentiated needs of different users.
SAIC IM has launched a L2/L3/L4 intelligent driving integrated technology route. Through "highly reusable hardware and software and incremental function upgrades", it reduces R&D costs and quickens the pace of implementation. With regard to Autonomous Driving Domain Controllers, it will cover high-performance high-compute chips and cost-effective domestic chips, and its platform strategy will meet the product needs from 2025 to 2027 and beyond:
L2: Currently, it has achieved the full coverage of 7V/11V + 1R (1L) + 12 USS + DMS; in 2025, it will officially enter the "Door-to-Door" intelligent driving era, with no frequent takeovers required throughout the journey from home to destination; from 2025 to 2027, the takeover rate will achieve a step-by-step increase from "10 times to 100 times".
L3: It plans to install 520-channel LiDARs, and upgrades to 12V + 5R + 1L + 12USS + 1DMS to enhance front perception performance and redundancy. In 2025, it will reach a mass production level of nearly 90%, and in 2026, it will officially become production-ready. The key safety indicator (XP event) will be controlled at "less than 1 time per 10,000 hours" to meet strict on-road use safety requirements.
L4: It will further strengthen perception redundancy, and adopt 12V + 5R + 4L + 12USS + 4DMS to achieve all-directional general obstacle perception and 360° redundancy coverage. IM Motors and SAIC AI LAB have developed the 3rd-generation Robotaxi based on IM's mass-produced models and architecture, and have obtained licenses for road test, demonstration application and demonstration operation at various stages; in 2025, the total test mileage of L4 intelligent driving will exceed 6 million kilometers, and 10 million kilometers in 2027, accelerating the commercial implementation of autonomous mobility.
L3/L4 Autonomous Driving Domain Controllers Develop Rapidly and Will Be Deployed in Vehicles in 2025.
The mass-produced solutions of OEM intelligent driving systems for passenger cars in the Chinese market (excluding imports and exports) have gradually shifted from L2 to L2.5 and L2.9. Meanwhile, the industry is promoting the mass production and implementation of higher-level intelligent driving assistance vehicles on a large scale. In September 2025, the Ministry of Industry and Information Technology of China announced the L3 intelligent connected vehicle access pilot list (first batch), with 7 automakers including Changan, BYD, GAC, SAIC, BAIC BluePark, FAW, and NIO on the list. This means that from the fourth quarter of 2025, passenger cars with L3 intelligent driving will be mass-produced and launched.
In addition to the pilot "consortium" list, the announcement also disclosed the specific information of the first batch of vehicle models that have obtained L3 intelligent driving access by July 2025, and major automakers have launched their "flagship products".
L4 Autonomous Driving Domain Controllers and Robotaxis installed are also being deployed. Currently, Lenovo Vehicle Computing has taken the lead in completing the mass production and implementation of Autonomous Driving Domain Controller projects based on the NVIDIA DRIVE AGX Thor platform. Among them, the L4 Autonomous Driving Domain Controller AD1 is based on the dual DRIVE Thor-X platform, with AI compute of 2000 TOPS. Its inference speed is 9 times faster than that of the previous generation, and its high bandwidth accelerates the processing of large language models. It also adopts a high-density blade array design, with an 8-fold increase in cooling area, providing a strong and solid computing foundation for complex intelligent driving scenarios.
WeRide has built the HPC 3.0 platform based on Lenovo AD1. It is trained on the NVIDIA DGX accelerated computing platform, adopts a dual-core NVIDIA DRIVE AGX Thor configuration and runs the safety-certified DriveOS, providing 2000 TOPS of AI computing power, making it a world's leading automotive-grade computing platform.
The mass production price is only a quarter of the previous generation, helping to reduce the cost of new-generation intelligent driving kits by nearly 50%.
It highly integrates functional modules such as Ethernet gateway, CAN gateway, inertial navigation positioning, and collision detection, lowering mass production costs and after-sales maintenance costs simultaneously. The total cost of ownership (TCO) of the platform is reduced by 84% compared with the previous generation.
It has passed multiple automotive certifications such as AEC-Q100, ISO 26262, and IATF 16949, and adopts a multi-redundancy design to meet ASIL-D level safety requirements. Its failure rate is lower than 50 FIT, MTBF is as high as 120,000-180,000 hours, and the design life is 10 years or 300,000 kilometers.
It is suitable for extreme environments (-40°C to 85°C), has passed full-scenario tests such as high-temperature aging, mechanical shock, and salt spray corrosion, and complies with global VOCs environmental protection standards, making it widely deployed in international markets such as the Middle East, Southeast Asia, and Europe.
AD1 has been installed in WeRide's flagship Robotaxi GXR, realizing the global debut and mass production of NVIDIA Drive Thor.
Pony.ai's L4 autonomous driving domain controller is independently developed based on 4 NVIDIA DRIVE Orin-X chips, with a computing power of 1016 TOPS and a design life of 10 years or 600,000 kilometers, providing a high-performance, automotive-grade, and high-safety computing base for Robotaxis.
700T+ Ultrahigh-compute Domain Controllers Have Begun to Be Installed in Vehicles and Adapt to VLA/World Models.
In terms of selection of main control SoCs for Autonomous Driving Domain Controllers, ultrahigh-compute domain controllers have started to be installed in vehicles. Typical chips include NVIDIA DRIVE Thor Series (700~1000 TOPS), Horizon Journey 6P (560 TOPS), and Huixi Technology's Guangzhi R1 (500 TOPS).
Desay SV's high-performance Autonomous Driving Domain Controller IPU14 is equipped with NVIDIA Thor, a chip with the highest computing power in the world, supporting L3 conditional intelligent driving and L4 intelligent driving in some scenarios. It is also Desay SV's high-end product for overseas markets. It will be mass-produced and installed in GAC Hyptec L4 models within 2025. In addition, this product has been designated by multiple automakers.
In August 2025, EXEED, a brand under Chery, launched EXEED ET5, a model equipped with the Falcon 700 driving assistance system. It is the world's first to be equipped with Horizon's J6P flagship chip and Horizon SuperDrive (HSD) system (built based on Horizon's high-compute flagship chip Journey 6P and a one-model end-to-end architecture). It will be mass-produced and launched on market in November 2025.
In addition, NIO's self-developed Shenji NX9031 (1000 TOPS) was officially mass-produced and installed with the delivery of the NIO ET9 in April 2025. It is also equipped with the self-developed vehicle full-domain operating system SkyOS, and will be available to the new models ET5, ET5T, ES6, and EC6 later. Xpeng' Turing AI chip (750 TOPS) has also been installed in the XPeng G7 Ultra. The car is equipped with 3 Turing AI chips, with computing power of over 2200 TOPS and 216GB RAM (each Turing AI chip with 64GB, totaling 192GB, and the 8295 chip with 24GB), bringing the industry's first L3 computing platform.
The XPeng G7 Ultra is equipped with 3 Turing AI chips, of which 2 run a VLA model (VLA-OL with independent reinforcement learning capabilities, responsible for motion control and real-time decision, requiring a frame rate of >=20fps), and the other runs a VLM (processing environmental perception and semantic understanding, e.g., road sign recognition and instruction parsing, requiring a frame rate of 1-2fps). The Turing chips communicate via PCIe, and simulate the "human brain partitioned collaboration" model (VLA is the brain motor area, and VLM is the brain cognitive area).
Typical Performance of Turing AI Chip:
Single-chip computing power of 750 TOPS, equivalent to three NVIDIA Orin-X chips;
Customized for AI foundation models, integrate 2 self-developed neural processing units (NPU) and 2 independent image signal processors (ISP), and adopt DSA (Domain-Specific Architecture) for neural networks;
Equipped with a 40-core processor, provide strong computing support for large models and support the operation of large models with relatively high parameters at the local end;
An independent safety island is set up for real-time blind-spot-free safety detection for vehicles.