Xpeng Motors is poised to start mass manufacturing and deployment of its internally developed AI autonomous driving chip, dubbed “Turing,” beginning within the second quarter of this 12 months, in keeping with firm chairman He Xiaopeng. The chip, set to debut in a forthcoming new Xpeng mannequin, represents a major step within the firm’s technique to bolster its superior driver-assistance techniques (ADAS) capabilities and cut back its reliance on third-party semiconductor giants like Nvidia.
Improvement of the Turing chip started in 2020, culminating in a profitable tape-out final August. Designed particularly for the calls for of AI and end-to-end giant fashions prevalent in trendy ADAS architectures, the chip reportedly boasts compute energy roughly equal to a few mainstream ADAS chips (implied to be Nvidia’s Orin X) or roughly 700 TOPS (Tera Operations Per Second). This efficiency stage positions it close to Nvidia’s latest-generation AI chip, Thor, significantly the ~750 TOPS model presently being provided.
Inner sources revealed the Turing chip integrates two proprietary neural processing items (NPUs) and incorporates a specialised structure optimised for neural networks. Xpeng claims this design achieves a 20% larger utilisation fee in comparison with general-purpose high-compute automotive chips and might domestically course of giant AI fashions with as much as 30 billion parameters – a considerable determine in comparison with present techniques like Li Auto’s VLM (Imaginative and prescient-Language Mannequin) with round 2.2 billion parameters. The chip additionally incorporates options like an impartial security island and twin impartial Picture Sign Processors (ISPs) for enhanced notion in difficult circumstances like evening driving or rain.
Xpeng’s accelerated push for its personal silicon comes amid reported challenges with Nvidia’s Thor chip rollout. Initially anticipated with as much as 2000 TOPS and a mid-2024 launch, trade sources point out Nvidia has confronted difficulties, primarily supplying a lower-spec model (~750 TOPS) with potential {hardware} and software program points. This uncertainty and potential delay affect a number of automakers, together with Xpeng itself, BYD, Zeekr, and Li Auto, who had beforehand introduced plans to undertake Thor. By bringing Turing to market forward of schedule, Xpeng features essential management over its product roadmap and hedges in opposition to exterior provide chain disruptions within the fiercely aggressive EV market. Chairman He Xiaopeng explicitly framed the transfer as enabling Xpeng to “decouple” from Nvidia.
This {hardware} development aligns with Xpeng’s bold software program technique centred round giant AI fashions. The corporate is growing a “World Foundational Mannequin,” focusing on 72 million parameters – roughly 35 instances bigger than typical VLA (Imaginative and prescient-Language-Motion) fashions. Xpeng’s strategy includes coaching this colossal mannequin within the cloud utilizing its devoted 10 EFLOPS computing cluster after which utilizing information distillation strategies to deploy optimised, smaller fashions onto the automobile’s Turing chip. This technique goals to beat on-device compute limitations whereas leveraging the facility of extraordinarily giant fashions for complicated reasoning and management, finally enabling multi-modal giant fashions to immediately handle driving duties. In response to the corporate, early assessments demonstrating a foundational mannequin controlling a automobile have already confirmed profitable.
The corporate is focusing on the preliminary commercialisation of L3 autonomous driving within the second half of this 12 months, with explorations into L4 deliberate for 2026. Xpeng’s ambitions lengthen past passenger automobiles. It plans to make use of the Turing chip in future AI-powered robots and flying automobiles.
