BYD Unveils In-House 4nm Autonomous Driving Chip "XuanJi A3", Matching NVIDIA Thor's Advanced Process

The Groundbreaking 4nm Automotive AI Chip

China’s leading electric vehicle manufacturer, BYD, has officially unveiled its self-developed automotive autonomous driving AI chip, XuanJi A3. The chip adopts the 4nm automotive-grade process, with design and testing fully completed in-house. The chip’s three-chip configuration achieves over 2100 TOPS of computational performance while keeping power consumption significantly lower compared to general-purpose GPU architectures. The debut of XuanJi A3 marks a new milestone in China’s autonomous driving chip market, positioning it in the same tier as NVIDIA’s Thor on the global stage for automotive AI computing chips. BYD, which has historically focused on battery technology and electrification, has now made a significant foray into AI chip design.

Technological Features of XuanJi A3 Here are

the core technical specifications of XuanJi A3: The chip employs a 4nm automotive-grade process. Unlike consumer electronics, automotive chips must withstand extreme temperature fluctuations, vibrations, and long-term reliability demands. BYD’s Wang Chuanfu explained that “the technical difficulty of a 4nm automotive-grade process is roughly equivalent to a 2nm process in the consumer electronics sector.” This is due to stringent standards, including redundant design and error-tolerant circuits during the design phase, the use of high-cost materials, and additional manufacturing processes. The CPU component of the chip features a 16-core configuration, achieving a computational capacity of 420K DMIPS. This enables it to handle not only autonomous driving tasks but also complex operations related to the cockpit and vehicle body control simultaneously. The memory bandwidth is 273GB/s, and its integration with a proprietary in-house bus achieves nanosecond-level low latency at the hardware level.

Differences Between NPU and GPU Architectures

The most distinctive feature of XuanJi A3 lies in its choice of an NPU (Neural Processing Unit) architecture specifically designed for AI inference tasks, rather than relying on general-purpose GPU-derived architectures like NVIDIA’s Orin and Thor. GPUs were originally developed for graphic rendering, excelling at handling large amounts of similar tasks in parallel. NVIDIA has evolved GPUs into general-purpose parallel computing platforms, making them applicable for scientific calculations and AI training. NPUs, on the other hand, are designed specifically for AI-related tasks. They implement AI-specific operations like matrix multiplication, convolution, and activation functions directly at the hardware level. To draw an analogy, a GPU is like a skilled craftsman capable of performing a variety of tasks, while an NPU is akin to a specialized worker who is exceptionally efficient at specific tasks. General-purpose GPUs need to maintain a certain degree of flexibility in hardware resources to cater to various automotive manufacturers, algorithms, and models, which leads to overhead in both chip area and power consumption. XuanJi A3’s NPU is deeply integrated with BYD’s proprietary algorithms. While general-purpose chips cater to diverse customer needs and include design trade-offs, XuanJi A3 is customized to BYD’s unique autonomous driving algorithms, enabling higher effective performance for the same theoretical computational power. Consequently, it achieves lower power consumption per unit of computational power and significantly improves computational efficiency.

Reducing Latency to Enhance Safety One of the

most direct benefits of the specialized NPU architecture is its ability to reduce latency. In urban navigation scenarios, the process from sensor data collection to execution involves several steps, including perception, prediction, planning, and control. Each step involves intensive calculations, and insufficient computational power or architectural inefficiencies manifest as “hesitations” in vehicle movement. Examples include delays in responding to interruptions, hesitation at complex intersections, or sluggish maneuvers in scenarios requiring detours. The NPU core of XuanJi A3 natively supports large-scale Transformer models and achieves nanosecond-level data scheduling in combination with its proprietary bus system. A demonstration video shown at the launch event featured testing in Shenzhen’s Pingshan district, showcasing the vehicle’s smooth and non-robotic maneuvers in scenarios such as sudden encounters with electric bicycles, parked vehicles on the roadside, and U-turns on narrow roads. While human reaction time is approximately 300–500 milliseconds, current autonomous driving systems can compress this to about 100 milliseconds. Dedicated NPUs can further shorten this window, providing additional safety margins. Even a difference of a few tens of milliseconds can be the deciding factor between stopping in time and a collision.

BYD to Fully Compensate for Autonomous

Driving Accidents At the launch event, Wang Chuanfu made a bold declaration to demonstrate confidence in vehicles equipped with XuanJi A3. “In the event of a traffic accident caused by autonomous driving during urban navigation, BYD will fully compensate for the economic losses incurred by the vehicle, with no cap on liability.” This is not merely a marketing statement. Such an assurance is only possible because BYD has complete control over the full stack, from hardware to software, combining its self-developed chips with its proprietary algorithms. BYD has already implemented a multi-redundancy architecture and numerous sensors, and with the integration of XuanJi A3, it is prepared to unlock hardware capabilities directly for L3/L4 autonomous driving as soon as regulations permit.

The Significance of Full-Stack In-House

Development With the introduction of XuanJi A3, BYD joins the exclusive group of companies capable of developing the entire process in-house, from batteries and electronic control systems to vehicle architecture and autonomous driving chips. This group includes notable names like Tesla, Huawei, NIO, XPeng, and Li Auto, and BYD’s entry into this space is highly significant. BYD’s strength lies not only in developing autonomous driving chips but also in its ability to mass-produce them and integrate them into its extensive range of vehicles. Its capacity to iterate quickly by designing, manufacturing, installing chips, and collecting real-world driving data gives it a significant competitive edge. As AI technology extends from vehicles into the broader domain of physical AI, XuanJi A3 serves as a foundational computing solution for BYD. Rather than competing on theoretical computational power with general-purpose GPUs, XuanJi A3 focuses on executing more efficient computations with fewer resources. Its AI-specific hardware supports more complex models, faster responses, and higher safety thresholds — embodying the core of BYD’s technological vision.