Xpeng Motors soars: The era of "intelligent computing" for autonomous driving is coming fast

Under the wave of intelligent upgrading of the automobile industry, autonomous driving has become one of the main selling points of more and more car companies. It is also an important indicator to measure the degree of automobile intelligence.

On August 2, Xpeng Motors announced that it would cooperate with Alibaba Cloud to build China's largest autonomous driving intelligent computing center "Fuyao" in Ulanqab, which will enable autonomous driving model training at a lower cost and faster speed, and also let everyone see the importance of computing power in the autonomous driving track.

The large-scale implementation of autonomous driving faces great computing power challenges

"Autonomous driving" can be simply understood as the automatic identification of road conditions and precise control of the vehicle through the vehicle's intelligent system. Autonomous driving can be divided into five levels: L1 driver assistance, L2 partial automation, L3 conditional automation, L4 high automation, and L5 full automation.

However, at present, most companies remain at the L2 stage, and the advancement to L3 still requires a "qualitative change" in technology, which is extremely challenging. At present, there are two technical paths in the industry: one is the pure vision solution represented by Tesla; the other is the multi-sensor fusion solution represented by Xiaopeng. The pure vision solution is to obtain environmental information through the camera and analyze and process it, but the camera's perception effect will be affected in harsh environments (such as fog or at night). Multi-sensor fusion adopts the mode of camera + lidar + high-precision map to adapt to different scenarios such as parking lots, highways, and urban roads. However, data fusion and response speed between different devices are also difficult points.

Regardless of which solution is adopted, the autonomous driving algorithm model needs to be trained persistently through massive amounts of road condition data, which requires the support of powerful computing power, making the training of autonomous driving models a "computing-intensive" beast.

High-quality data makes autonomous driving “smarter”

Data-driven is the recognized direction of autonomous driving development. Autonomous driving algorithm models such as visual detection, trajectory prediction and driving planning rely on massive data sets for machine learning.

It can be said that data intelligence is the most fundamental driving force for the evolution of autonomous driving AI. Through further learning, mining, processing and training of feedback data, more advanced algorithms and service models are obtained and OTA is delivered to the vehicle end, which in turn brings better system performance to users. In this process, cost and speed have become the ideological stamp of data intelligence. For this reason, major autonomous driving car manufacturers are maximizing the input-output ratio of their data assets by obtaining high-quality training data.

For intelligent driving technology, data accuracy is particularly important: on the one hand, real traffic scenarios are complex and there are many security threats, so the efficiency and agility of data analysis are very important; on the other hand, the quality of labeled data will directly affect the safety of intelligent driving. For example, the accuracy of data labeling such as portraits, buildings, plants, roads, traffic signs, and vehicles directly determines the intelligent driving AI's judgment of road conditions, and the quality of training data directly affects driving safety.

The "Fuyao" intelligent computing center dedicated to autonomous driving built by Xiaopeng Motors in Ulanqab with a computing power of 600PFLOPS supports the training time of Xiaopeng's core autonomous driving model from 7 days to within 1 hour, which is nearly 170 times faster and greatly improves the efficiency of model training.

"Fuyao" landing: committed to achieving lower costs and stronger computing power for autonomous driving

The "East Data West Computing" project was officially launched, pressing the "accelerator" for the national computing power network. According to statistics from the Ministry of Industry and Information Technology, my country has built more than 20 intelligent computing centers.

Cai Yinghua, President of Global Sales of Alibaba Cloud, said that Alibaba Cloud is actively deploying intelligent computing centers in the national computing power hub nodes, including "Fuyao". The overall scale of Alibaba Cloud Ulanqab Intelligent Computing Center is 3 EFLOPS ("floating-point operations per second" and "peak speed per second"), which is the world's largest autonomous driving intelligent computing center. Massive data processing capabilities and efficient AI development capabilities are becoming the key infrastructure for automotive intelligence. The cooperation with Xiaopeng Motors is "the embodiment of Alibaba Cloud's basic computing power and business scenario applications."

In the era of L1 and L2 autonomous driving, since the amount of data is relatively small, many car companies can accept closed integrated solutions with strong coupling of chips and algorithms. However, in the era of L3 and L4, with the surge in data volume, algorithms have become more complex, and high-computing power chips are needed to meet the demand.

In He Xiaopeng's view, starting from 2025, some domestic cars will enter the era of true autonomous driving, commonly known as the driverless era. In this process, the demand for computing power for autonomous driving model training will also increase significantly. Regardless of the algorithm models such as visual detection, trajectory prediction or driving planning of autonomous driving, as the amount of data increases, the lack of computing power will cause the research and development speed to be far behind the speed of data growth. Based on this, in order to improve efficiency and reduce costs, Xiaopeng chose to build an intelligent computing center with Alibaba Cloud - "Fuyao", to achieve stronger computing power at a lower cost. First, fine-grained segmentation and scheduling of GPU resources will increase the virtualization utilization of GPU resources by 3 times, support more people to develop online at the same time, and improve efficiency by more than ten times. At the communication level, end-to-end communication delay is reduced by 80% to 2 microseconds. In terms of overall computing efficiency, linear expansion of computing power is achieved. Storage throughput is 40 times higher than the general level of 20GB/s in the industry. In addition, Alibaba Cloud's machine learning platform PAI provides AI engineering tools such as model training deployment and inference optimization.

Huo Jia, general manager of the sales department of Alibaba Cloud Intelligent Solutions, said that Fuyao uses a cluster model exclusive to public clouds. "On the one hand, it can achieve the tuning and optimization of Xiaopeng's computing power requirements, and on the other hand, it also retains the possibility of using the elasticity of large-scale cloud computing infrastructure in the future. This is also the original intention of our overall design."

"Fuyao" rises: Accelerating the promotion of a new pattern of autonomous driving

As we move towards the era of computing-driven cars, the value of intelligent car computing services is highlighted. From the perspective of business logic and technological evolution, data centers of technology companies also need to help car companies achieve better autonomous driving functions with lower costs, higher system concentration, and lower power consumption.

"At the beginning of last year, we estimated that the investment in computing power would exceed 1 billion yuan in 2025," said He Xiaopeng, founder of Xpeng Motors. "In terms of improving computing power, controlling costs, and even making corresponding technical preparations, Xpeng Motors discussed internally whether to build its own cloud computing power or to build it through cooperation?" He said that after the internal workload "decreased a lot" over time, the cooperation with Alibaba Cloud made efficiency higher and investment costs lower.

Based on the changes in Xiaopeng Motors' autonomous driving simulation training needs and the demand for powerful local + cloud computing support, Xiaopeng Motors finally teamed up with Alibaba Cloud's intelligent computing platform to build a dedicated intelligent computing center "Fuyao" for autonomous driving to further improve the efficiency of model training.

He Xiaopeng said, "In the past, it took about a week to train, but now the computing platform can achieve it in just one hour. I believe that more car companies will do this in the future. Strong coupling must be formed. This is the only way forward."

Currently, technology companies including Alibaba, Google, and Microsoft are building intelligent computing centers for machine learning around the world. This time, Alibaba Cloud and Xiaopeng Motors have joined hands to build the largest intelligent computing center in China, "Fuyao", which is hailed as "a historic moment for China's autonomous driving innovation and upgrade". He Xiaopeng said: "With the launch of Xiaopeng's 'Fuyao' intelligent computing center, the computing power of autonomous driving in simulation training will be greatly improved. I believe that in 2025, Xiaopeng Motors will begin to remove the word "assisted" from automatic assisted driving and truly enter the era of autonomous driving."

In He Xiaopeng's view, Ulanqab's location and climate advantages are not only conducive to computing, but the clean energy in the west it represents is also particularly important at present. In 2020, Alibaba put into operation a data center here. Now Xiaopeng Motors and Alibaba Cloud's "Fuyao" has also chosen to settle here. In addition to meeting the demands of "low-carbon and environmental protection", it is also responding to the country's "Eastern Data and Western Computing" strategy. It is understood that "Fuyao", as a greener and lower-carbon intelligent computing center, combines the natural climate advantages of Ulanqab, adopts green technologies such as air cooling, AI temperature control, and modular design, which can achieve fresh air operation for more than 80% of the time throughout the year, with an annual average PUE of less than 1.2.