Overview
Vehicle autonomous driving features are rapidly penetrating the market, driving fast growth in the intelligent driving domain controller market. The intelligent driving domain controller is a key component in the driving decision chain. Its main functions are to process perception data and perform planning and decision making. The core component is the compute chip. Vendors such as NVIDIA and Horizon have prominent positions in this segment. As consumer demand for autonomous features increases, base L2 functionality costs decline and adoption of lower-compute solutions rises; leading OEMs continue to improve driving capabilities, expanding urban NOA coverage and increasing demand for high-compute domain controllers. The wide adoption of BEV+Transformer architectures also drives higher compute and structural changes for domain controllers.
Market Penetration and Drivers
Autonomous driving feature penetration is rising quickly but remains at a relatively low level, leaving significant room for growth. As technology matures, product prices fall and user demand for smarter experiences increases, autonomous driving features are moving down from luxury segments into mainstream models, accelerating penetration. Higher performance and increased fitment rates for driving features will drive demand for intelligent driving domain controllers.
The intelligent driving domain controller functions as the vehicle compute center. Domain controllers typically include the compute chip, safety MCU, memory chips and other passive components. The compute chip handles camera image processing, runs deep learning algorithms, outputs recognition results, performs sensor fusion and trajectory prediction. This chip determines domain controller performance. Intelligent driving chips have high technical barriers; current high-end chips are led by NVIDIA and Huawei, while mid- and low-end suppliers include Horizon, Mobileye, TI and Black Sesame.
Chip Vendors and Market Structure
Major chip vendors are expanding product portfolios and market share for China-based chips continues to rise. Domain controller suppliers that establish stable co-development relationships with key chip vendors are positioned to benefit.
Chip development and manufacturing capability remain core competitive advantages. With BEV+Transformer adoption, onboard compute requirements increase, especially for urban NOA and future L3 functionality, so high-compute chips remain the preferred option for high-end models from leading OEMs. Domain controllers are still mainly supplied by Tier 1 suppliers in the Chinese market; although some leading OEMs consider in-house development, barriers include R&D history, funding and sales scale, making in-house development difficult and costly. Currently L2-and-above domain controllers are dominated by suppliers with comprehensive product lines, strong chip design capability, mature mass-production capacity and high shipment volumes.
L2 Function Penetration and NOA
L2 penetration is accelerating
In August 2021 the Ministry of Industry and Information Technology proposed a recommended national standard for vehicle driving automation grading. The standard divides driving automation levels considering factors such as operating domain limitations. L2 automation key functions include ACC adaptive cruise, AEB automatic braking and LKA lane keeping, enabling a degree of active vehicle control. L3 systems should be able to sustainfully execute all dynamic driving tasks within their designed operating conditions. Because the gap between L2 and L3 is large, the industry often uses an intermediate L2+ category that includes NOA and other higher-assist functions that still require driver supervision.
Functional evolution: single-lane to multi-lane to NOA
Early ADAS focused on single-lane longitudinal and lateral control, achieving collision avoidance and lane maintenance. Typical functions included ACC, LCC, TJA, CCS and automatic parking. As technology advanced, systems began handling multi-lane scenarios and lane changes, adding features such as ALC, multi-lane TJA and multi-lane HWA. Recently, vendors have pushed toward point-to-point control driven by navigation, i.e., highway NOA and urban NOA.
Penetration still low, with significant upside
As technology matures, prices drop and user demand grows, autonomous driving features are shifting from luxury to mid- and entry-level models. According to industry data, in the first half of 2023 the China passenger vehicle market saw 3.244 million units with factory-standard L2 (including L2+) ADAS, a 37.7% year-on-year increase. Factory-standard fitment rate was 34.9%, up about 8 percentage points year-on-year. For new energy vehicles, factory-standard L2 (including L2+) fitment reached 1.471 million units, a 75.6% year-on-year increase, with a 50.4% fitment rate, up about 10 percentage points. Improvements in feature performance and fitment will increase domain controller demand. Most vehicles with intelligent driving features use a domain controller architecture; domain controllers handle perception processing and decision making, so their installed volumes should grow with feature penetration.
NOA rollout is advancing and is a key focus area
Navigate on Autopilot (NOA) includes highway NOA and urban NOA. By combining ADAS with map navigation, NOA delivers point-to-point assisted driving. Industry data shows factory-fit NOA deliveries reached 263,000 units in Jan–Jul 2023, up 120.4% year-on-year. Highway NOA is a current priority for OEMs because highway scenarios are more structured and rollout is earlier. Early highway NOA relied on high-definition maps to enable automatic lane changes and ramp merging, but high-precision maps and positioning are costly.
To reduce costs, highway NOA may evolve to simplified versions that drop ramp merging, instead providing takeover reminders before ramp entry. Urban NOA is accelerating to cover primary driving scenarios and enable continuous point-to-point functionality from highways to urban areas. Early urban NOA often used high-definition maps for precise localization, but map collection cost, limited coverage and slow updates hinder large-scale rollout. As vehicle compute and sensor capabilities improve, many vendors adopt a "light map" plus onboard perception approach that strengthens map data only where navigation maps struggle, such as ramp mouths, to realize NOA functions.
Urban NOA Expansion
As many OEMs migrate to BEV+Transformer visual perception routes, dependency on maps decreases and urban NOA becomes more scalable. Several vendors announced aggressive city coverage plans. These efforts demonstrate rapid urban NOA expansion and increased city-level deployment.
Standards and L3 compliance progress
Regulatory frameworks for intelligent and connected vehicles are being implemented. For example, Shenzhen introduced detailed regulations covering definitions, testing and responsibilities for intelligent vehicles, and national agencies have issued guidelines for pilots and standardization. As the standard system and management policies become more detailed, industry development is expected to accelerate in the Chinese market.
Consumer Trends and Demand
Automotive consumer upgrades are increasing demand for intelligent features. Higher-priced segments grew their share in 2023, and the average transaction price for vehicles with factory-standard L2 (including L2+) rose to CNY 266,000 in H1 2023. Consumers increasingly view cars as lifestyle devices and prioritize entertainment, interaction, comfort and safety, driving interest in intelligent vehicles and accelerating penetration.
Policy, demand and supply together support rapid advancement of vehicle intelligence. New consumer cohorts emphasize intelligent cabins and driving features, and OEMs are improving user experience through vehicle design, services and autonomous functions. The industry consensus is moving toward a digital, software-defined vehicle that supports ongoing feature upgrades.
NOA Adoption Outlook
NOA is in a rapid growth phase. As software improves and perception shifts toward vision-heavy approaches, costs decrease and NOA can expand into lower-price segments and more scenarios. Current factory-fit NOA models are mostly priced above CNY 300,000, but some models under CNY 200,000 are beginning to adopt NOA. Projections suggest NOA will become standard on mid- to high-end models while being offered on lower-priced models as an option, progressively reaching broader market tiers.
Vision-First Trend
Tesla embraces vision-heavy perception
Since 2018 Tesla focused on 2D image processing with CNNs and manual labeling, later adopting partial automatic labeling to improve efficiency. In 2020 Tesla moved to a BEV approach and replaced CNNs with Transformer-based networks, initializing BEV features and fusing them with 2D image features via multi-layer Transformers to produce BEV features and reduce fusion errors. The BEV approach expanded from road network perception to full spatial reconstruction.
Vendors in the Chinese market shift to vision-first
Early China-based approaches used multi-sensor fusion with HD maps, millimeter-wave radar and lidar for perception, relying on maps and lidar for static obstacles and vision for dynamic obstacles. To reduce cost and improve generality, leading companies are transitioning to BEV+Transformer vision-heavy solutions.
The BEV+Transformer acceptance changes domain controller architecture. Previously, camera-focused systems used fewer cameras and relied on radar for 360 coverage. Typical configurations included 5R1V, 3R1V, 1R1V and similar. HD maps provided road network information and lidar handled static obstacles. Under BEV, at least 6–8 cameras are required for full surround perception, increasing camera interface demand. Removing HD map dependency can simplify or reduce some localization modules. Large models deployed on the vehicle require higher compute; low-compute chips cannot support on-device deployment of BEV+Transformer models, so chipsets must evolve to support Transformer operators. Without native Transformer operator support, models trained on GPUs face difficulties being converted and deployed at the edge, often leading to suboptimal approximations. Leading chip vendors such as Horizon and NVIDIA have already added Transformer operator support, enabling BEV+Transformer on the vehicle side.
Perception emphasis is moving to cameras. After the BEV+Transformer shift, low-cost vision sensors gain importance. To ensure data quality, the number and resolution of cameras increase, with 8MP cameras becoming common and total camera counts moving toward 10 or more per vehicle. Millimeter-wave radar and lidar remain important for near-term redundancy. Because multi-frame reconstruction with techniques like NeRF is still developing, lidar or 4D mmWave radar may be required to obtain road network data and lane markings. Pure vision struggles in low-light and nonstandard object cases, so 4D mmWave radar is often used for static obstacle detection. Camera and 4D mmWave demand is expected to grow.
Compute Requirements
NOA demands higher perception accuracy and compute than basic L2
BEV perception can run on low-compute chips (for example 8 TOPS) but with lower perception accuracy compared to mid/high-compute platforms. Algorithm choices differ: high-compute platforms often use Transformer-based carrier solutions; low-compute platforms may use BEV Depth or BEV Det style 2D-to-3D methods. For map-less or light-map NOA, domain controllers must handle more perception data, increasing compute requirements. As driving functions advance, high-compute chips remain the preferred choice for premium models. The core of domain controller hardware is chip compute capability.
Higher-level autonomous driving and large vision models increase on-vehicle compute needs. L3 requires much more compute than urban NOA; currently mass-produced chips struggle to meet L3 demands, and chip development toward higher compute continues. OEMs need to provision sufficient compute headroom for future upgrades. For L3-class systems, redundancy designs are required to ensure safety, such as dual domain controllers or backup chips, which can increase overall chip demand.
E/E Architecture Evolution
E/E architectures evolve through four stages: distributed ECU architecture, domain-consolidated architecture, quasi-central compute architecture, and central compute architecture. OEMs are accelerating E/E evolution. Domain-concentrated stages include several vendor-specific architectures. Quasi-central architectures combine a central compute platform and regional controllers; final evolution aims for full central compute architectures that consolidate functions into a single central controller.
Distributed ECU architectures feature many independent ECUs connected via CAN and LIN, leading to increased wiring, higher installation cost and difficulty in maintaining and upgrading heterogeneous modules from multiple suppliers. Domain controllers (DCUs) integrate similar ECU functions, addressing many issues of distributed E/E architectures. Bosch-style domain classification divides the vehicle into powertrain, chassis, cockpit, automated driving and body domains, each managed by a domain controller. Domain controllers centralize processing and replace distributed ECUs for many functions.
Domain Controller Advantages
DCU benefits
1) Wiring reduction lowers installation cost. 2) Higher integration eases coordinated management, standardizes data interfaces and reduces development and manufacturing cost. 3) Higher compute makes OTA upgrades easier. With fewer modules and more centralized control, OTA can target the DCU and avoid compatibility and protocol issues across many ECUs, simplifying security verification and reducing anti-tamper overhead.
Intelligent Driving Domain Controller Role
The intelligent driving domain is a critical E/E subsystem and foundational for vehicle autonomy. The domain controller is the decision-making center that must process diverse perception streams—cameras, mmWave radar, lidar, inertial sensors—and complete compute and decision tasks in tight timeframes. It is central to enabling L3 and above.
Domain Controller Structure
Complex architecture centered on compute chips
Hardware components include: 1) Compute chip: camera image processing, deep learning inference, sensor fusion and prediction. 2) Safety MCU: handles high-assurance functional safety tasks, external interfaces such as radar, vehicle control and communications. 3) Memory: eMMC, NOR Flash, DRAM. 4) Other components: resistors, capacitors, thermal solutions, sealed metal housing, PCB, interfaces, gateway and power management ICs. Software includes the underlying OS, middleware and applications. Domain controllers differ from ECUs mainly by higher chip compute, software/hardware decoupling and heavy dependence on the main SoC plus software integration.
Market estimates project rapid growth: domain controller prices are expected to decline while fitment rates rise. Lower-cost single-chip parking-and-driving solutions (for example single TDA4 or J3 chips) for basic L2 functions have significant market potential. A projection estimates the China passenger car domain controller market could reach CNY 47.94 billion by 2025, with 2021–2025 compound growth around 109.9%.
Compute Chip as Core Component
Compute chips determine domain controller performance. AI chips used in domain controllers — CPUs, GPUs, DSPs — remain mainstream, with FPGA and ASIC solutions also emerging. Domain controller vendors and chip vendors cooperate on software stacks and ecosystems; early access to advanced samples and prior integration experience provide advantages for rapid development.
Intelligent driving chips have high barriers and a relatively concentrated vendor base. High-end chips are led by NVIDIA and Huawei; mid/low-end vendors include Mobileye, TI, Horizon and Black Sesame. Qualcomm and others are expanding into the AD space. Domestic (China) chip replacement trends are evident as Horizon and other China-based vendors increase market share. According to one 2022 ranking for L2+ NOA domain controller chip solutions in the China market, Horizon and NVIDIA held the top two shares, together accounting for about 95% of the segment, with TI, Mobileye and Huawei occupying much smaller shares.
NVIDIA
NVIDIA leveraged its graphics and GPU compute experience to enter automotive SoCs in 2015, providing compute chips and development tools. NVIDIA targets high-end segments; its Orin SoC is among the highest compute mass-produced solutions available.
NVIDIA first released the Xavier SoC in 2016 and achieved volume production by 2020. Xavier used an ARM CPU, Volta GPU and ASICs on 12nm process with up to 30 TOPS. Orin, released in 2019 and mass-produced from 2022, uses a 7nm process, integrating 17 billion transistors to deliver up to 254 TOPS with DLA and PVA accelerators. Orin remains a top choice for advanced driving systems; high-performance platforms like NIO's Adam use multiple DRIVE Orin chips for very high aggregate TOPS. NVIDIA announced Thor in 2022 with targeted 2000 TOPS and multi-domain capabilities; Thor is expected to support intelligent driving, occupant sensing, infotainment and other functions, enabling domain convergence.
Horizon
Founded by machine learning experts, Horizon launched its first automotive chips in 2017 and has since developed series including Journey and Sunrise chips. Journey 2 was mass-produced in 2019. Journey 3 (J3) and subsequent chips have been widely adopted in low-compute platforms. Horizon's Journey 5 is a vehicle-grade chip offering up to 128 TOPS on a single chip, supporting many cameras and multiple sensors and enabling BEV algorithm deployment. Journey 5 has been chosen for production of several models and supports single-chip parking-and-driving solutions for high-performance functions.
Horizon provides hardware reference designs, algorithms, middleware, toolchains and development platforms. Its toolchain includes an "algorithm warehouse", quantization and compilation tools, middleware frameworks and a cloud-based AI development platform for data labeling, training and deployment. Horizon also offers an automotive application development kit supporting mass-production development processes and standardized interfaces to improve integration and verification efficiency.
Qualcomm
Qualcomm launched Snapdragon Ride in 2020 targeting L1/L2, L2+ and L4/L5 segments with scalable compute ranges. In 2023 Qualcomm announced Ride Flex with Mid/High/Premium tiers, aiming at cabin and driving convergence with up to 2000 TOPS. Qualcomm emphasizes an open, programmable platform and an Arriver visual stack to support from single-camera to multi-camera solutions and enable OEMs and Tier 1s to build driving strategies and software stacks.
Mobileye
Mobileye has been a leader in vision-based ADAS since the EyeQ1 in 2004. Mobileye provides a "black-box" integrated chip-plus-algorithm solution that reduces development time for OEMs with limited in-house capability. Mobileye also increased openness with EyeQ Kit to lower development cost. EyeQ chips have progressed from sub-1 TOPS to 176 TOPS in the upcoming EyeQ Ultra, with EyeQ5 and EyeQ6 series supporting various camera configurations and ADAS/AD capabilities.
Industry Competitive Dynamics
Chip development and manufacturing capabilities remain central. Domain controller design and production models vary: OEMs may outsource production, Tier 1 suppliers provide hardware and integration, Tier1.5 vendors focus on basic software platforms and Tier0.5 suppliers pursue full-stack participation. Common models include: 1) OEM-contract manufacturing; 2) Tier 1 production and integration with chip vendors and OEMs; 3) Tier1.5 focusing on domain controller middleware; 4) Tier0.5 deep OEM-supplier integration.
Supplier Landscape
Participants fall into four groups: global Tier 1 integrators with customer scale; China-based Tier 1 suppliers building full-stack solutions and deep OEM partnerships; software platform vendors providing modular platforms; and OEMs pursuing in-house domain controller or chip development. Domain controller supply in the Chinese market remains dominated by Tier 1 suppliers, as OEM in-house efforts face R&D, funding and volume constraints. Main China-based Tier 1 vendors include Huawei, Desay SV, and Jingwei Hirain, with others such as Joyson Electronics, Thundersoft, DJI and Neusoft Reach advancing rapidly. International Tier 1 vendors including Continental, Bosch and ZF are also active, but after L2+ they have been challenged by local Tier 1 progress.
Representative Vendors and Solutions
Huawei
Huawei offers a full-stack intelligent driving portfolio with MDC platforms covering multiple scenarios. The MDC product family includes MDC300F for mining/port/plant/highway logistics, MDC210 for mid/low-end fitment, MDC610 for high-end branding uplift and MDC810 for Robotaxi or high-level AD using Ascend 620 with compute over 400 TOPS. Huawei emphasizes functional safety and provides a toolchain, SDKs and cloud services including large scenario and dataset resources to accelerate development.
Desay SV
Desay SV has long automotive electronics experience and supplies infotainment, ADAS and connected services. It collaborates closely with NVIDIA and has iterated IPU01-04 domain controller products. IPU01 targeted surround and parking controllers; IPU02 is a lightweight driving domain controller for mid/low price bands; IPU03/04 are higher-performance platforms. IPU03 based on NVIDIA Xavier has been mass-produced on several models. IPU04 is a high-compute platform using dual Orin chips with total compute up to 508 TOPS and expandable further, supporting up to 16 high-definition cameras, multiple radar and lidar units. Desay SV leverages chip partnerships, mass-production capability and software-hardware integration to support rapid customer adaptation and high-volume supply.
Desay SV combines domain controller design with sensors and algorithms, providing a product matrix including cameras, radar, smart antennas and software for integrated solutions. Their long-term cooperation with NVIDIA and experience enable strong product-market fit and supply advantages, especially during component shortages.
Jingwei Hirain
Jingwei Hirain focuses on cost-effective driving solutions and has developed an integrated ADAS + ADCU + HPC product suite supporting different levels. Its ADCU integrates Mobileye EyeQ4 and Infineon AURIX platforms and supports radar, lidar and HD map inputs. Jingwei Hirain also developed a central compute platform based on NXP heterogeneous SoC architecture for real-time control and adaptive AUTOSAR, supporting vehicle data acquisition, OTA, SOA services and ASIL-D functional safety, and enabling integration of proprietary perception, planning and control algorithms.
ThunderSoft / ThunderAuto
ThunderSoft focuses on intelligent operating systems and scenario-based solutions. It has developed end-to-end low-speed autonomous driving products from surround view to AVP and launched domain compute platforms via its subsidiary. ThunderSoft is building ThunderAuto OS and SmartDrive ADAS, offering hardware abstraction, software openness, cloud services and scenario solutions to reduce development complexity. Collaborations with Qualcomm and other partners have produced RazorDCX domain controllers based on Snapdragon Ride, offering scalable compute and multi-camera connectivity for a range of use cases from parking to urban driving.
DJI
DJI leverages strong vision expertise to offer NOA solutions. In April 2023 DJI announced a solution achieving L2+ capabilities with as low as 32 TOPS in 7V/9V pure-vision configurations without reliance on HD maps or lidar. The system supports optional sensors for redundancy and can scale compute up to 200 TOPS for advanced functions. DJI is progressing toward production with OEM partners.
Joyson Electronics
Joyson is developing high-compute domain platforms based on NVIDIA Orin and Qualcomm Snapdragon Ride Flex. Projects include dual-Orin high-compute domain controllers for advanced highway and urban NOA and AVP, and Snapdragon Ride-based controllers offering up to 200 TOPS with redundant dual-chip architectures and advanced cooling solutions. Joyson engages with multiple chip vendors including Horizon and Black Sesame for collaborative development.
Neusoft Reach
Neusoft Reach has developed an all-domestic automatic driving domain controller platform using Horizon Journey 5 and other domestic chips, supporting L2++ functionality via a high-performance parking-and-driving integrated controller. The platform supports high-resolution cameras, 4D radar, lidar and ultrasonic sensors and provides multi-domain fusion solutions and central compute controllers designed with SOA principles to reduce development complexity and support cross-domain functions.
Many other companies offer parking-and-driving integrated domain controllers. Horizon, Black Sesame, Horizon Matrix, and other vendors provide various ADU solutions. International Tier 1s also invest in domain controllers, but integration complexity, development cost and cross-supply-chain barriers make rapid market entry difficult for newcomers. Leading China-market Tier 1s such as Desay SV and Huawei are expected to maintain significant market share due to R&D strength, production scale and deep OEM collaboration.
OEM Strategies
Few OEMs choose in-house R&D plus contract manufacturing
OEM decisions to develop in-house domain controllers depend on factors like cost, core technology control, supply stability and product differentiation. Hardware accounts for an estimated 60–80% of domain controller cost, with software at 20–40%. In-house development requires sustained R&D investment and time. Some head OEMs choose self-development to control costs and differentiation, while most OEMs opt to purchase Tier 1 solutions to avoid high upfront costs and long development cycles. The long-term landscape is expected to include a small number of high-end OEMs developing in-house domain controllers combined with contract manufacturing, while mainstream supply remains dominated by strong Tier 1 vendors such as Huawei and Desay SV.
Tesla
Tesla is a leader in full-stack in-house development. Its FSD hardware and software began in-house, with HW3.0 and later HW4.0 containing Tesla-designed SoCs and neural accelerators. Tesla pioneered techniques such as HydraNet, BEV+Transformer perception, occupancy networks and temporal-feature fusion for robust perception and planning. Recent HW4.0 revisions use advanced process nodes and multiple NPUs with increased CPU core counts and added 4D mmWave radar for redundancy.
Xpeng
Xpeng initially used third-party domain controllers but moved to in-house domain controller development for tighter software-hardware integration. Xpeng's XNGP driving system combines dual Orin chips for high compute and multi-sensor input, and its in-house-developed XNet vision network performs multi-frame temporal fusion to output 4D dynamic and 3D static information to enhance urban scenario performance. Xpeng also benefits from large-scale compute resources for training, significantly reducing training time.
Conclusion
Intelligent driving domain controllers are a crucial link in vehicle autonomy. The market is expanding as L2+ and NOA features diffuse from premium to mainstream models, and architecture shifts toward BEV+Transformer increase compute demands. Compute chips remain the core determinant of domain controller capabilities, and chip development plus manufacturing strength are key competitive factors. Leading Tier 1 suppliers and a small number of OEMs pursuing in-house solutions will shape the supply landscape as adoption grows and standards evolve in the Chinese market.
