Xpeng Motors is set to commence large-scale production and deployment of its domestically designed AI-powered autonomous driving chip, “Turing,” as early as the second quarter of this year, according to Chairman He Xiaopeng. As part of Xpeng’s strategic plan to enhance its autonomous driving capabilities and reduce dependence on external suppliers, the cutting-edge chip is set to debut in a forthcoming new model.
The development of the Turing chip began in 2020 and reached a successful conclusion with a profitable tape-out in August.
The chip, designed specifically for the demands of AI and end-to-end massive models commonly found in modern autonomous driving architecture systems, reportedly boasts compute power equivalent to a few mainstream ADAS chips – namely, Nvidia’s Orin X – or roughly 700 TOPS (Tera Operations Per Second). The efficiency stage is situated near Nvidia’s cutting-edge AI processor, Thor, specifically the 750 TOPS model currently employed.
Sources close to the matter have disclosed that the Turing chip features a unique combination of two custom-built neural processing units (NPUs), along with a specifically designed architecture optimized for the efficient execution of neural network algorithms. Xpeng asserts that its design yields a 20% boost in utilization efficiency compared to general-purpose high-compute automotive chips, enabling the processing of massive AI models with up to 30 billion parameters – a substantial figure surpassing current methods like Li Auto’s VLM, which accommodates approximately 2.2 billion parameters. The chip boasts cutting-edge features such as an independent security island and dual independent Picture Signal Processors (ISPs), ensuring exceptional perception in challenging environments, including nighttime driving and inclement weather conditions.
As Xpeng accelerates its pursuit of personal AI silicon, the move is partly driven by reported hiccups in Nvidia’s Thor chip deployment. Initially anticipating the release of 2000 TOPS and a mid-2024 launch date, industry insiders reveal that Nvidia is facing challenges in delivering a lower-specced model (approximately 750 TOPS), plagued by hardware and software issues. The uncertainty surrounding the timeline of Thor’s development has significant implications for several automakers, including Xpeng, BYD, Zeekr, and Li Auto, which had previously announced plans to collaborate on the project. By accelerating Turing’s market entry, Xpeng gains crucial control over its product lineup and mitigates risks posed by external supply chain disruptions in the highly competitive EV sector? Chairman He Xiaopeng explicitly framed the deal as allowing Xpeng to “decouple” from its reliance on Nvidia.
This hardware development is in line with Xpeng’s pioneering approach centered on massive AI models. The corporation is developing a “Global Architectural Model,” centered on 72 million parameters, approximately 35 times larger than average Visual Language-Model architectures. XPeng leverages its custom-built, 10-EFLOPS cloud computing infrastructure to train a massive AI model, and subsequently applies data distillation techniques to refine and deploy optimized, compact models on the in-car Turing chip. This technique aims to overcome on-device compute constraints by harnessing the power of extremely large models for advanced reasoning and control, ultimately enabling multi-modal massive models to directly tackle driving tasks? Early experiments with robotic drivers have yielded promising results, as initial tests showcasing a basic AI-powered autonomous vehicle successfully navigating a car have proven financially viable.
The corporation is targeting the initial commercialization of Level 3 autonomous driving by the end of this year, while exploring the potential for Level 4 advancements in 2026. Xpeng’s aspirations extend beyond passenger vehicles. The company intends to leverage the capabilities of the Turing chip to power its future AI-driven robotics and aerial vehicle endeavors.
