Rejecting Transitional Silos: Defining a Unified L3/L4 High-Performance Computing Architecture Through the End-game
Release time：2026-07-31
On July 22, the 2026 CHINA AUTO FORUM was held in Shanghai. Peng Xueming, Senior Chief Engineer at Desay SV, delivered a speech titled Keeping Pace with the Era, Heading Towards the Ultimate Goal of Intelligent Driving. In the speech, he shared Desay SV’s strategic insights into the developmental trends of the autonomous driving industry.
Currently, China’s intelligent driving sector is moving past the pilot testing phase and entering the critical stage of large-scale commercial deployment. Following the first batch of admission approvals for Level 3 (L3) conditional autonomous vehicles issued by the Ministry of Industry and Information Technology of the People's Republic of China last year, the market penetration rate of new passenger vehicles equipped with ADAS functions in China has reached 70% this year. As automobiles accelerate their transformation into mobile intelligent terminals, establishing a clear framework of rights and responsibilities across the entire automotive value chain has become a core priority for OEMs to ensure the sustainable development of intelligent driving.
Strategic Foundation: High-compute Platforms
With the gradual implementation of L3 autonomous driving regulations and safety liability frameworks, approaches that strictly pursue low initial hardware costs are rapidly becoming obsolete. Peng emphasized that rather than rejecting reasonable cost reduction, the industry must ensure that cost-cutting does not come at the expense of future capabilities, safety liabilities, or redundant development. Looking toward L3 deployment and the long-term commercialization of Level 4 (L4) autonomy, stakeholders should move beyond short-term cost-accounting mindsets. Instead, they should focus on the entire lifecycle and liability chain, recognizing the enduring value of unified high-compute platforms.
Peng outlined three medium-to-long-term industry forecasts: First, autonomous driving is entering a decisive five-year “end-game” cycle, where algorithms, data, and chips, the three foundational prerequisites for scalable high-level autonomous driving, are simultaneously crossing a critical threshold. Second, a unified, homogeneous “central brain” will emerge as the long-term optimal solution for adapting to both L3 and L4 requirements. Third, there is an industry consensus that development is shifting from a rule-driven approach to an AI-native paradigm.
The Long-Term Optimal Solution: A Unified Central “Brain”
A company’s choice of technological pathway directly dictates its long-term core competitiveness. A common L3 technical scheme today is the asymmetric high/low compute architecture, where a high-compute primary domain manages core functions while a low-compute domain controller handles redundancy. While offering single-vehicle BOM cost advantages, it may become an expensive technical “debt” over a long lifecycle. Computing architecture upgrades and L3 compliance declarations would require a complete reconstruction of the development process. Furthermore, multi-hardware architectures struggle to meet the rigid requirements for unified risk management and coordinated fault degradation, leading to higher hidden compliance costs.
In contrast, a unified and homogeneous central “brain” will serve as the optimal long-term solution. This architecture utilizes a universal software stack: the primary computing unit is responsible for full-domain perception and core decision-making, while the secondary computing unit independently provides safety redundancy. Both share the same data closed-loop, simulation verification, and OTA (Over-The-Air) frameworks, achieving fault isolation purely through hardware partitioning. This approach not only meets L3 multi-redundancy safety standards but also enables smooth iteration to L4, eliminating redundant development and overlapping investments.
The Computing Power Threshold: 2000+ TOPS
Intelligent driving is evolving from manual, rule-based programming to AI-native iteration. Future-proof intelligent driving “brains” must support full-scale perception fusion, AI-native model inference, substantial headroom for long-term OTA evolution, and L3/L4 safety redundancy. Consequently, building a full-domain, high-compute foundation oriented around full-vehicle AI agents is an inevitable trend.
Peng predicted that a 2000 TOPS dual high-compute architecture represents the baseline computing threshold for AI-native L3 systems, while L4 platforms will require upwards of 6000 TOPS. He noted that true cost-effectiveness means deploying a solution that remains compliant, evolvable, and structurally sound five years down the line without needing a complete overhaul. Desay SV’s dual high-compute architecture relies on comprehensive systemic capabilities rather than merely stacking high-compute chips. Backed by high-compute hardware design, full-stack software, rigorous safety systems, and mass-production expertise, Desay SV translates high computing power into usable, reliable, scalable, and continuously evolving intelligent driving capabilities.
As the implementation of L3-related national standards approaches, the intelligent driving industry stands at a critical watershed. Amidst this profound transformation, Desay SV is concurrently celebrating its 40th anniversary. The convergence of an industry paradigm shift and a new chapter in corporate development reinforces Desay SV’s commitment to long-termism.
