The China Mobile–Alibaba Cloud Intelligent Computing Center in Hami, Xinjiang, is part of the Xinjiang-Chongqing computing power project. (Photo/Hami Media Center)
Chongqing - The race to build smarter cars is also becoming a race for computing power. On September 9, the Chongqing Computing Power Base in Hami, Xinjiang, began supplying Changan Automobile with computing resources to develop the one-stage end-to-end large model powering its SDA Pilot driver-assistance system.
This cross-regional supply is part of China’s national computing network, which allows companies to access remote data-center capacity through high-speed links. The shared infrastructure is not limited to automakers, with applications spanning more than 20 areas, including low-altitude aviation, urban governance, mining and agriculture.
Five days earlier, Changan unveiled SDA Pilot, a driver-assistance system powered by a one-stage end-to-end large model. The system was first equipped on the Changan Nevo Q06. As cars move deeper into the AI era, computing power is becoming a critical foundation for intelligent-driving R&D.
Unlike earlier approaches that relied heavily on manually defined driving rules, end-to-end models learn more directly from real-world driving data and continuously improve through training and iteration. SDA Pilot was trained on more than 20 million high-quality data segments. As data sets and models grow, so does demand for computing power.
That demand had become a challenge for Changan. The automaker has more than 700 petabytes, or 700 million gigabytes, of raw intelligent-driving data. But building its own computing infrastructure requires heavy investment and long construction cycles, while conventional long-distance networks face bandwidth constraints and unstable latency. A single round of large-model training once routinely took 30 days.
According to First Sight News, Jia Lishan, Deputy General Manager of China Changan Automobile Group, said access to computing resources in Xinjiang directly addresses three major bottlenecks in driver-assistance model training: long training cycles, high computing costs and unstable supply.
Xinjiang’s advantage lies in its energy resources and capacity to support large-scale computing. Large-scale computing depends heavily on power supply. Hami’s computing innovation demonstration zone can support more than 60,000P of computing capacity.
Once a 200-megawatt direct green-power project is completed, green electricity will account for more than 60% of direct power supply, while electricity costs for companies could fall below 0.3 yuan (0.04 U.S. dollars) per kilowatt-hour.
But putting computing resources more than 2,000 kilometers away from Chongqing also creates a transmission challenge. The Xinjiang-to-Chongqing project addresses this by linking computing centers in Xinjiang with industrial users in Chongqing through high-speed fiber-optic and all-optical networks.
One-way data transmission latency between Chongqing and Hami has now been cut to 16.2 milliseconds. When Chongqing sends an intelligent-driving model training instruction, Xinjiang’s computing cluster responds almost immediately, with the data making a round trip in about 32 milliseconds. This allows Xinjiang’s computing resources to be used directly by Changan for large-model training, massive data processing and simulation testing.
As green computing resources and low-latency networks become integrated into Changan’s R&D system, the company’s computing bottlenecks are beginning to ease. Jia said the cross-regional model of “R&D led by Chongqing and computing power provided by Xinjiang” has cut Changan’s AI algorithm iteration cycle by 30% and computing costs by 30%, while making large-scale computing supply more stable.
Computing power alone does not translate directly into intelligent capabilities. The key is connecting data collection, model training, algorithm iteration and vehicle testing into a continuous development loop.
Jia said Changan plans to expand its use of domestically developed computing resources at the Chongqing (Xinjiang) Computing Power Base. The additional capacity will support road-scenario data training, end-to-end model optimization and intelligent-cockpit interaction model iteration, forming a closed loop from “data collection to model training, algorithm iteration and vehicle testing.”
The cooperation will also work in the other direction. Changan plans to deploy intelligent connected vehicles in Xinjiang and feed data collected from real-world operations back into SDA Pilot. The data will support targeted training for extreme weather and complex road conditions, helping improve the system’s ability to generalize across different driving environments.