In the field of intelligent driving, a technological revolution driven by computing power may be accelerating. It is said that the cooperation between Xpeng Motors and Huawei will go further, and perhaps Xpeng Motors will use Huawei's Ascend chip to strengthen its computing power construction, which not only marks the deep binding of the two companies in technology, but also indicates that the intelligent driving industry is entering a new era of "computing power is king". Behind this cooperation, there is not only the practical demand for complementary technologies, but also the far-reaching significance of reshaping the industry pattern.
1. The computing power arms race of intelligent driving
With the gradual maturity of autonomous driving technology, computing power has become a key factor in determining the performance of intelligent driving systems. From Tesla's FSD chip to NVIDIA's DRIVE Orin platform, major manufacturers are seizing the commanding heights of technology by improving chip computing power. According to the data, Tesla's FSD chip computing power reaches 144 TOPS, while NVIDIA's DRIVE Orin platform is up to 254 TOPS, and these chips provide powerful computing support for the autonomous driving system.
The rapid rise of Xpeng Motors in the field of intelligent driving in recent years is inseparable from its continuous investment in computing power. In 2024, based on the strong computing power support of Alibaba Cloud, Xpeng Motors has achieved more than 1 billion kilometers of driving video training, combined with 6.46 million kilometers of real vehicle tests and simulation tests, greatly improving the iteration speed of intelligent driving technology. However, with the introduction of end-to-end technology, Xpeng's demand for computing power has further increased. End-to-end technology simplifies the training process by directly generating driving instructions from input to output through neural networks, but also requires higher computing power. This technical architecture requires the chip to be able to process large amounts of data in a short period of time and make decisions in real time.
Pictured: An Influencer, ID"Thomas Electric Train", vice president of Xpeng Motors, shows the in-depth cooperation between Xpeng and Huawei
2. The rise of Huawei's Ascend chips
The rise of Huawei's Ascend chips has provided a new option for Xpeng Motors. With a half-precision floating-point computing capability of 256 TFLOPS, the Ascend 910 chip is one of the AI chips with the highest single-chip computing density. Compared to NVIDIA's H100 chip, the Ascend 910 performs better in terms of power efficiency, achieving 320 TFLOPS of computing power with 310W power consumption, while the H100 requires 400W power consumption. In addition, the Ascend 910 supports flexible expansion of multiple AI frameworks, including Huawei's self-developed MindSpore framework, which provides Xpeng with more possibilities for the development of intelligent driving algorithms.
Another significant advantage of Huawei's Ascend chips is their price-performance ratio. The Ascend 910B has a unit price of about $17,000, while the Nvidia H100 has a unit price of more than $30,000. Under the current international situation, the restrictions on the export of high-end chips by the United States have made Ascend chips an important choice for domestic enterprises. By cooperating with Huawei, Xpeng Motors can not only reduce hardware costs, but also accelerate the implementation of technologies with policy support.
3. The computing power requirements of end-to-end technology
The introduction of end-to-end technology is an important technological innovation in the field of intelligent driving. End-to-end technology improves the efficiency and accuracy of models by simplifying the training process. With the support of this technology, Xpeng's "triple play" architecture (XNet, XPlanner and XBrain) can better cope with the driving needs of complex road conditions.
However, end-to-end technology is extremely demanding on computing power. For example, it takes about 1 month for the NVIDIA H100 cluster to train a 175 billion parameter GPT-3 model, while the Ascend 910B cluster may need to be extended to 2 months. Despite the increase in training time, the Ascend 910B has reduced hardware costs by 30%, which is an important consideration for an automaker like Xpeng.
4. Reshaping the industry landscape
The cooperation between Xpeng and Huawei is not only a deep technical binding between the two companies, but also indicates the reshaping of the intelligent driving industry pattern. As computing power has become the core competitiveness of intelligent driving, more and more car companies have begun to cooperate with technology giants to jointly develop high-performance chips. For example, the cooperation between Xiaomi and BYD in battery technology, and NIO's use of NVIDIA's Orin chip, etc., are similar cases of ecological integration.
This cooperation model will give birth to more "technology giants + car companies" alliances to promote the rapid development of intelligent driving technology. In the future, the competition in the field of intelligent driving will no longer be limited to a single technology, but the collaborative ability of the entire ecosystem. The cooperation between Xpeng and Huawei marks the transformation of the intelligent driving industry from "single-point technological breakthrough" to "ecosystem competition".
5. Challenges and opportunities for the future
Although the cooperation between Xpeng and Huawei has provided a strong impetus for the development of intelligent driving technology, it also faces some challenges. The first is the performance bottleneck of Ascend chips in large-scale training. The interconnection bandwidth of the Ascend 910B is only 13% of that of the NVIDIA H100, which may affect training efficiency. Second, Huawei's Ascend ecosystem is still in its early stages, and compared with NVIDIA's mature CUDA ecosystem, there is still a gap between the developer toolchain and the efficiency of multi-card collaboration.
However, challenges often mean opportunities. By optimizing algorithms and compressing models, Xpeng can achieve higher inference efficiency on Ascend chips. For example, the inference efficiency of DeepSeek-R1's small 7B model on the Ascend 910B is close to that of the NVIDIA H800. In addition, with the continuous improvement of Huawei's Ascend ecosystem, Xpeng Motors will be able to make better use of this platform to further enhance the competitiveness of intelligent driving technology.
6. Conclusion
The cooperation between Xpeng Motors and Huawei's Ascend chips is not only a strong combination of the two companies in terms of technology, but also an important step for the intelligent driving industry to reach new heights. With the deepening of the cooperation between the two sides, we have reason to believe that intelligent driving technology will usher in more breakthroughs and innovations, bringing consumers a safer, more efficient and more convenient travel experience. In the future, with the acceleration of intelligent transformation and the continuous progress of technology, the cooperation between Xpeng Motors and Huawei will undoubtedly bring more innovation and change to the entire automotive industry.
