As the competition in autonomous driving intensifies, Tesla, XPeng, and BYD are pushing the boundaries with their latest intelligent driving architectures. This article compares their approaches in terms of hardware configuration, software strategies, and technical features.
Tesla: Vision-Only End-to-End Architecture
Hardware Architecture
Tesla’s current Full Self-Driving (FSD) platform, known as Hardware 4 (HW4), is built around an 8-camera vision system. It is powered by Tesla’s custom FSD chip, delivering hundreds of TOPS of processing power. Although Tesla once removed radar sensors entirely, some have been reintroduced in HW4 for redundancy and edge-case support.
Software Strategy
Tesla employs a pure vision-based, end-to-end neural network approach. Its key component is a Bird’s Eye View (BEV) perception model, which fuses multi-camera images into a semantic spatial map. The system then directly outputs driving trajectories without relying on high-definition (HD) maps.
Technical Highlights
End-to-End Design: Simplifies the traditional modular pipeline by training neural networks to handle perception, planning, and control jointly.
Data-Driven Iteration: Uses massive real-world data collected from the fleet to continuously improve the FSD software through over-the-air (OTA) updates.
Black-Box Tradeoff: The architecture sacrifices interpretability for performance and scalability.
XPeng: Multimodal Fusion and Large AI Models
Hardware Architecture
XPeng’s flagship G9 vehicle is equipped with a dual-NVIDIA(https://www.avaq.com/manufacturer/nvidia) Orin setup providing 508 TOPS, 12 cameras, 5 mmWave radars, 12 ultrasonic sensors, and 2 LiDARs. This creates a rich multi-sensor fusion environment to support high-level autonomy.
Software Strategy
XPeng has launched an end-to-end architecture supported by three core AI modules:
XNet: A unified deep vision network for 3D environmental reconstruction.
XPlanner: A neural planning network for generating smooth and human-like driving trajectories.
XBrain: A large language model (LLM) used for scenario understanding and behavioral prediction.
Technical Highlights
Map-Free Driving: The XNGP platform eliminates dependence on HD maps by relying solely on real-time sensor data and AI inference.
Large Model Efficiency: By leveraging massive datasets and large-scale training, XPeng claims the system performance will increase 30x in the next 18 months.
Multimodal Intelligence: Integrates visual, spatial, and language cues for robust and generalizable autonomous decision-making.
BYD: “Xuanji” Architecture with Smart-Electric Integration
Hardware Architecture
BYD’s “Xuanji” architecture represents a unified approach combining electrification with intelligent systems. It includes:
One Brain: A central AI controller.
Two Ends: Coordinated cloud and on-board intelligence.
Three Networks: Integration of in-vehicle network, 5G, and satellite connectivity.
Four Chains: Sensing, control, data, and mechanical systems.
Software Strategy
BYD’s intelligent driving system is structured in tiers:
DiPilot: A Level 2 ADAS platform (DiPilot 10/30).
“Eye of God”: An advanced L2+ system (DiPilot 100/300/600) evolving with computing power.
These systems are supported by BYD’s full-stack capabilities across electrification and vehicle intelligence.
Technical Highlights
Full-Stack In-House Development: Integrates platforms like e-Platform 3.0, DiLink, and DM-i to ensure tight software-hardware synergy.
Multi-Modal AI Models: Supports over 300 driving scenarios with self-developed neural networks and large AI models.
Real-Time Perception: Combines edge inference and low-latency communication for millisecond-level decision-making.
Comparison Table
Company Sensing Approach Architecture Map Dependency Key Strengths
Tesla Vision + radar End-to-End BEV Neural Net No Scalable via OTA, simple architecture
XPeng Multi-modal (LiDAR, Radar) XNet + XPlanner + XBrain No Robust generalization, LLM integration
BYD Camera + LiDAR + Radar “Xuanji” + DiPilot stack Partial Smart-electric fusion, full-stack autonomy
Conclusion
Tesla, XPeng, and BYD each embody distinct philosophies in autonomous driving:
Tesla focuses on pure vision and simplicity via neural networks.
XPeng leverages sensor fusion and large AI models for comprehensive understanding and adaptability.
BYD integrates intelligent driving within a broader vehicle system architecture, showcasing synergy between smart and electric technologies.
Each path reflects not only a technical strategy but also a business positioning: Tesla emphasizes global scalability, XPeng targets rapid AI iteration, and BYD anchors its approach in system-level integration.