PUBLISHER: ResearchInChina | PRODUCT CODE: 1777123
PUBLISHER: ResearchInChina | PRODUCT CODE: 1777123
High-level intelligent driving penetration continues to increase, with large-scale upgrading of intelligent driving SoC in 2025
In 2024, the total sales volume of domestic passenger cars in China was approximately 22.84 million, a year-on-year increase of 8.6%, which was about 3 percentage points higher than the 5.8% growth rate in 2023. Among them, the number of passenger cars equipped with L2 and above intelligent driving SoC exceeded 14.56 million, with installation rate of 63.8%, which was about 6 percentage points higher than that of in 2023.
The sales volume of China's domestic passenger cars has increased year by year, but the growth rate has slowed down. At the same time, intelligent driving SoC has been accelerating its installation in vehicles. From the perspective of intelligent driving levels:
L2.9: The most models in the 250,000-300,000 yuan price range are equipped with it. In 2024 and Q1 2025, the penetration rate has exceeded 50%; for the price range above 500,000 yuan, the penetration rate in 2024 and Q1 2025 exceeded 30%; currently, NVIDIA Orin-X, Huawei Ascend 610, Tesla FSD, etc. dominate the market. New products include NVIDIA Thor-X, Thor-U, Orin-Y, Qualcomm 8797/8397, Horizon J6P/J6M, Black Sesame Technologies A2000/C1236, RHINO Guangzhi R1, XPeng "Turing", NIO Shenji NX9031, etc., which will be put into mass production in vehicles one after another in 2025.
L2.5: The penetration rate of models above 200,000 yuan shows an upward trend. Among them, mainstream intelligent driving SoCs include Horizon J5/J6E, NVIDIA Orin-X/Orin-N, Black Sesame Technologies A1000, TI TDA4VH/VM, Mobileye EyeQ5H and other chips.
L2: It has been dropped to models within 100,000 yuan, most of which are equipped with low-level intelligent driving SoC chip products from manufacturers such as Horizon, Axera, and Mobileye.
Entering 2025, intelligent driving SoC has entered a replacement cycle. Some automakers have begun to introduce self-developed intelligent driving SoC, such as NIO and XPeng; most automakers have upgraded their products as in the past, such as Li Auto.
On August 23, 2024, the world's first AI chip, XPeng's "Turing" chip, successfully taped out; it was officially mass-produced and launched in June 2025, and was first equipped on the XPeng G7 Ultra, with 3 chips installed. Among them, 2 Turing chips run the VLA model (VLA-OL with autonomous reinforcement learning capability, responsible for motion control and real-time decision, with a frame rate requirement of >=20fps), and the other 1 Turing chip (runs VLM, handles environmental perception and semantic understanding, such as road sign recognition and instruction parsing, with a frame rate requirement of 1-2fps). The Turing chips communicate through PCIe, simulating the "human brain partition collaboration" mode (VLA is the motor brain region, VLM is the cognitive brain region).
Typical performance of "Turing" AI chip:
Single chip computing power 750TOPS, equivalent to three NVIDIA Orin-X;
AI large model customization, integrating 2 self-developed neural network processing brains (NPU), 2 independent image signal processors (ISP), and adopting DSA (Domain-Specific Architecture) for neural networks;
Equipped with 40-core processor, providing strong computing support for large models and supporting the operation of large models with high parameters on the local end;
An independent safety island is set up to carry out real-time full-vehicle blind-spot-free safety detection.
In 2025, the intelligent driving SoC of the Max and Ultra versions of Li Auto's L-series models will be upgraded to NVIDIA Thor-U chip, with a single-chip computing power of 730 TOPS, supporting VLA large models. The core capabilities of VLA include:
Language interactive driving: Support voice commands to perform complex operations (such as "avoid congestion" and "find a parking space to park"), realizing anthropomorphic decision.
Scene generalization: Optimize driving trajectories through diffusion models to solve long-tail scenes (such as unmarked roads and special-shaped intersections)
The Li Auto L6 Pro will add a lidar, and the intelligent driving chip will be upgraded from Horizon J5 (1 chip) to J6M (2 chips), with computing power upgraded from 128TOPS to 256TOPS, supporting up to 12 cameras, 6 radars and lidar. According to the official statement, after upgrading the hardware, the AD Pro intelligent driving system models will be comparable to AD Max in terms of active safety capabilities.
Intelligent driving SoC chip vendors move towards high computing power chips and accelerate the deployment of full-stack self-developed algorithm IP to reduce the technical application threshold for OEMs
Chip vendors continue to introduce new products to seize the high-level intelligent driving market. NVIDIA's latest Thor chip has multiple versions, with a maximum computing power of up to 2000TOPS, and will be mass-produced one after another starting from 2025. As of June 2025, models such as Li Auto MEGA, the intelligent driving upgraded version of L-series, Lynk & Co 900, Zeekr 9X, and Xiaomi YU7 have officially announced their installation. Potential partners also include BYD, GAC, Mercedes-Benz, Hyundai, Jaguar Land Rover and other automakers.
NVIDIA Thor has four main versions: X, U, S, and Z. Among them, the Thor U with 730T computing power has attracted the most attention from passenger car manufacturers; the Thor X with 1000T computing power is mainly used in Robotaxi scenarios, and some manufacturers are considering using a single Thor X or a combination of dual Thor U to build an L3 autonomous driving system. The estimated price of Thor-X is between 600-800 US dollars.
In addition, there are two other versions. One is Thor-X-Super, which is the combination of two Thor-X chips, possibly using MCM, a process similar to Chiplet, with an estimated price between 1000-1300 US dollars. The other is Thor-Jetson, used in robotics and industrial fields, with 1000TOPS computing power, Ethernet interface bandwidth of 100GB, and an estimated price between 400-500 US dollars.
At the same time, chip vendors are accelerating the deployment of full-stack self-developed algorithm IP, and everything is ready for OEMs to use directly.
NVIDIA launched the L3 autonomous driving system NDAS (NVIDIA DRIVE AV Solution) at the GTC 2025 conference, codenamed Alpamayo. It is actually an end-to-end system. NVIDIA covers from model training, sensor simulation, traffic flow simulation, synthetic data, world model to model deployment, from VLM to VLA. The first version of NDAS was launched in April 2025, the dual Thor highway version L3 will be launched in the first quarter of 2027, and the dual Thor Urban version L3 will be launched at the end of 2027.
Qualcomm mainly promotes two product lines: Snapdragon Ride and Snapdragon Ride Flex. According to data released by Qualcomm, 2025-2026 will be a large-scale mass production window for its assisted driving projects: more than 20 ADAS/AD projects worldwide are planned to be launched within these two years, with cooperating automakers covering Chinese and foreign brands such as Leapmotor, FAW, Geely, BAIC, Chery, Volkswagen, GM, Honda, Mercedes-Benz, and BMW.
According to Qualcomm's official data, more than 10 Chinese automakers and Tier-1 partners are using the Snapdragon Automotive Platform Ultimate Edition (including Snapdragon 8797 and Snapdragon 8397) to build driving assistance and intelligent cockpit solutions:
On June 27, 2025, Leapmotor and Qualcomm Technologies announced that Leapmotor's flagship D series will be equipped with the Snapdragon Automotive Platform Ultimate Edition (QAM8797P), with a single-chip equivalent computing power of 640TOPS and dual-chip is up to 1280TOPS. One chip specializes in intelligent cockpit, realizing immersive voice interaction and high-definition audio-visual entertainment; the other focuses on assisted driving, supporting L3+ level intelligent driving functions, realizing in-depth integration of cockpit and driving;
BAIC confirmed that it will cooperate with Autolink and Zhuoyu Technology to promote intelligent landing based on Qualcomm's solutions.
Based on the Snapdragon Automotive Platform Ultimate Edition, Qualcomm has, through model lightweight and scheduling optimization, collaborated with Tier-1 manufacturers to achieve smooth operation of 14 billion parameter (14B) large models on the local end. Autolink, Bosch, and Visteon have demonstrated solutions integrating 7 billion parameter (7B) end-side models. From the perspective of performance indicators:
SAIC-GM plans to equip the Snapdragon 8775 chip to realize the "8-screen linkage" experience in Buick models.
14B model inference can achieve a frame rate of more than 40 FPS, which can be increased to 50-60 FPS after optimization;
The 7B model has a frame rate of 60-72 FPS, meeting real-time interaction needs, such as understanding user's voice and semantics, and building user portraits.
Domestically, Black Sesame Technologies has launched its high computing power chip platform designed for next-generation AI models - Huashan A2000 family.
Manufactured using 7nm process, integrating multi-functional units such as CPU, DSP, GPU, NPU, MCU, ISP and CV, with a built-in self-developed "Jiushao" NPU, featuring single-chip multi-task processing capability. When the intelligent driving function is operating normally, it can desensitize, compress, encode and store the entire vehicle's data;
Including three products: A2000 Lite (urban intelligent driving), A2000 (full-scenario general intelligent driving) and A2000 Pro (high-level full-scenario general intelligent driving), which are respectively aimed at different levels of autonomous driving needs, with computing power ranging from 256TOPS to 1000TOPS.
At the same time, Horizon officially launched Journey 6P (J6P). Compared with the previous generation Journey 5 (about 20 billion transistors), the transistor scale has increased by nearly 85%. Based on the newly developed BPU "Nash" architecture, it adopts a super-heterogeneous computing core, integrating five computing modules: CPU, GPU, NPU, MCU and VPU. Cooperating with a three-level storage architecture and data transformation engine, it significantly improves the chip's processing capability for complex intelligent driving algorithms.