Before June, the most powerful assisted driving vehicle in China was Nio’s ET9, an executive sedan with over 2,000 TOPS (tera operations per second) of computing power and a starting price of RMB 788,000 (USD 110,320). But Xpeng Motors has now reshaped that leaderboard.
On June 10, Xpeng introduced its new SUV model, the G7, at a launch event in Guangzhou. But much of the spotlight wasn’t on the car itself. Instead, CEO He Xiaopeng dedicated over half of his presentation to Xpeng’s new Turing chip, which also dominated questions during the post-event media session.
The G7 is equipped with three Turing chips developed internally by Xpeng. CEO He claims this setup delivers more than 2,200 TOPS, roughly the equivalent of nine Nvidia Orin X chips. By comparison, the standard industry configuration of two Orin X chips offers around 508 TOPS, and even new models from Xiaomi and Li Auto using Nvidia’s ThorU chip top out at about 700 TOPS.
Xpeng intends for the G7 to set a new benchmark for in-car computing over the next five years. “Many analyst reports say we’ll need hundreds, or even 1,000–4,000 TOPS within the next two to three years, or by 2030 at the latest. Xpeng is getting ahead of that,” He said.
He argues that such computing power is essential for enabling Level 3 and eventually Level 4 autonomous driving. This is why Xpeng has centered its marketing for the G7 on its processing capabilities, rather than conventional vehicle specs, despite its starting price of RMB 235,800 (USD 33,012).
Still, full Level 3 designation depends not just on intelligence but also on hardware redundancy and regulatory approval. The Turing chip is not yet standard across the G7 lineup. Only the Ultra variant includes the full three-chip setup; the Max variant still uses the standard dual Orin X configuration.
Two of the Turing chips in the Ultra version handle assisted driving functions, running a locally deployed vision-language-action (VLA) model. The third supports the smart cockpit, running a vision-language model (VLM). According to He, Level 3-capable vehicles must be able to locally process both VLM and VLA models, each representing a significant step beyond existing setups.
VLMs are designed to interpret visual data in conjunction with text for perception and understanding. VLA models go further, generating control outputs to guide vehicle behavior. Xpeng said its Turing chips can support models with up to 30 billion parameters, giving the VLA setup substantial headroom.
According to 36Kr, a single Turing chip can support large models with up to 30 billion parameters. That gives the dual-chip VLA model setup ample room to scale.
“Higher compute capacity raises both the upper and lower bounds of artificial intelligence, especially the lower bounds tied to safety,” He said. On straight roads, he added, a low- and high-compute system may appear similar, but processing speeds diverge: lower-end chips may handle just 3–10 frames per second, while more advanced systems can hit 10, 20, or even 30 frames per second with less latency. “The ride may feel the same, but the safety margin is completely different.”
The G7’s VLA-OL model, short for VLA with online reinforcement learning capabilities, builds on this approach. He described it as incorporating a “motion cortex” powered by internal reinforcement learning. According to him, it allows for better handling of real-world driving scenarios, from detecting emergency vehicles to navigating potholes and collapsed roads. These features, however, are still being refined. The hardware is in place, but software optimization and user experience development are ongoing.
“We’re all crossing the river by feeling the stones,” said He during the press roundtable, acknowledging the uncertainties in adopting large model strategies for assisted driving.
Other players are exploring similar directions. Li Auto is also working on VLA systems, while Tesla continues to focus on E2E approaches, largely avoiding multimodal models. What both companies share is a need for high onboard computing power. Tesla’s forthcoming AI5 hardware is expected to deliver 3,000–7,200 TOPS, based on industry estimates.
Xpeng sees compiler engineering as a future lever to consolidate compute loads. “Eventually, we may be able to replace three chips with one. If one chip can match the power of four, that would be a huge leap,” He said.
But the assisted driving field remains in flux, with new architectures emerging almost every year. Xpeng only finalized the Turing chip’s core design in 2022, after scrapping earlier iterations. To create a product with a five-year performance buffer, the chip team had to anticipate eight years of future tech trends, particularly in memory bandwidth. Even Nvidia has faced challenges bringing its Thor chips to scale, from thermal design to yield issues. As a first-time chipmaker, Xpeng still has a long road ahead.
Aside from the Turing chip, the G7 offers several standout features: a single configuration with a 702-kilometer range, trunk space that fits 37 twenty-inch suitcases, an augmented reality heads-up display (HUD) co-developed with Huawei, and an AI-driven “eagle eye” vision system.
By narrowing its SKU lineup and leveraging both inherited features from existing models and its proprietary compute stack, Xpeng has clearly positioned the G7 as a cornerstone product. It fills the pricing gap between the G6 and G9, with the company hoping to drive both sales volume and a higher average selling price.
Still, competition is intensifying. In the latter half of the year, the premium electric SUV segment will see new entries like Xiaomi’s YU7 and Li Auto’s i6. The contest is heating up even before the G7 reaches customers.
Ultimately, Xpeng’s broader ambitions in assisted driving hinge on the success of the Turing chip. At the same time, the chip’s mass production and deployment are closely linked to the company’s roadmap for Level 3 capabilities and VLA model integration. Progress on either front depends on the other.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Li Anqi for 36Kr.