Nowadays, new energy vehicles have become the first choice for many Chinese people to buy a car, especially the intelligent control experience is incomparable to traditional fuel vehicles. However, whether it is autonomous driving or intelligent control, these technologies are inseparable from the support of powerful computing power, especially the large amount of data generated by vehicle sensors will need to be processed and analyzed in real time.
In January 2024, the expert group of Digital China Miles and Computing Economy China Tour came to the Geely Star Smart Computing Center located in Changxing County, Huzhou, Zhejiang Province, to not only visit and inspect the computing infrastructure, but also gain an in-depth understanding of how computing power empowers various businesses of car companies, such as artificial intelligence, large models, intelligent driving, intelligent cockpits and other hot topics.
Zhang Yunquan, member of the 14th National Committee of the Chinese People's Political Consultative Conference and researcher of the Institute of Computing Technology of the Chinese Academy of Sciences
As Dr. Zhang Yunquan, member of the 14th National Committee of the Chinese People's Political Consultative Conference, deputy director of the Central Science and Technology Committee of the Jiusan Society, deputy director of the Central Science Popularization Working Committee of the Jiusan Society, leader of the Computing Standards Working Group of the National Beacon Committee, and researcher of the Institute of Computing Technology of the Chinese Academy of Sciences, said: "In the era of digital economy, computing power has become a new productive force, and computing power economy is a new economic form derived from the digital economy, especially in the automobile industry, where data and computing power are extended to the whole link." As an important direction for the upgrading and development of the automobile industry, new energy vehicles are the most direct manifestation of the application of computing power - before each car leaves the factory, engineers have carried out perfect intelligent control and application tests for the whole vehicle to ensure safer intelligent driving and more intelligent cockpit for users.
Dr. Chen Yong, Director of the Data Intelligence Development Center of Geely Automobile Central Research Institute
In order to deeply explore the value of data applications, Geely has built a full-scenario digital intelligent service through "1 center + 3 platforms + N full-scenario data intelligent services". Dr. Chen Yong, director of the Data Intelligence Development Center of Geely Automobile Central Research Institute, introduced that one center is the intelligent computing center. The three platforms are big data application platform, artificial intelligence application platform, intelligent driving data closed-loop and virtual reality simulation platform, and the data intelligence services provided include AIGC artificial intelligence content generation, virtual reality simulation platform, simulation computing application platform, remote diagnosis RVDC, after-sales intelligent services, etc.
Data-driven
Intelligent computing empowerment
The first phase of construction of Geely Star Intelligent Computing Center was completed in July 2022 and officially unveiled on February 27, 2023. With a total investment of 1 billion yuan, the data center covers an area of 52.12 acres, with 5,000 planned cabinets, and 2,520 standard 5kW cabinets have been built in the first phase At present (January 2024), the data center is running with more than 60,000 CPU resources, more than 2,000 GPU cards, and a storage capacity of more than 60PB.
The idea of building a traditional data center is to build up the infrastructure first, and then gradually move the business in. In the face of the wave of intelligence, Geely needs top-down planning, and urgently needs to support digital and intelligent business and drive the construction of a platform through business.
Xie Dong, Director of the Intelligent Computing Platform Development Department of Geely Automobile Central Research Institute
In July 2022, when the Geely Star Smart Computing Center was initially completed, a large number of businesses poured into it in a short period of time. Xie Dong, Director of the Intelligent Computing Platform Development Department of Geely Automobile Central Research Institute, said that as the first "cloud, data, and intelligence" integrated super cloud computing platform among global car companies, Geely has a unified intelligent architecture, unified R&D, unified operation, and full-stack tenant operation.
Geely's overall R&D system is basically based on data, including data collection, data processing, data application, and data compliance processing.
The computing facilities of Xingrui Intelligent Computing Center are mainly divided into three areas: intelligent driving area, general area, and simulation area. The business of the intelligent driving area includes data collection management, data annotation management, intelligent driving simulation management, data management platform, model research and development platform, etc. The general area mainly serves the R&D of intelligent cockpit and cloud vehicle machine, as well as the data lake with a unified R&D system. The main business of the simulation area includes collision, strength, electromagnetic, aerodynamic, thermal management and other aspects of vehicle operation, which requires a large amount of high-precision computing power.
Over the past year of operation, from the perspective of products (new energy vehicles), the intelligent computing center has also achieved many phased results, such as:
The full access to NOA mass production has brought more intelligent driving capabilities to mainstream price range models (such as Boyue L) and realized the automatic closed-loop of intelligent driving data;
A new generation of cockpit operating system, Galaxy NOS, has been launched, covering the needs of a variety of models, allowing users to have the same car-machine interaction experience as a mobile phone;
Full-scene AI services, including full-scene AI models, WOW wallpapers, AI music rhythms, children's picture books, etc., have been applied to large models, AIGC and other technologies.
Industry-leading big data application services, such as data lakes and governance capabilities, support the real-time data return data platform of more than 3.5 million vehicles, and the governance level has reached 0988, improving product user services and caring for users in real time through big data applications.
It can provide 7×24 hours of cloud protection for new energy vehicles, such as the Aegis battery safety system of Galaxy E8, which includes intelligent security control in the cloud, which can achieve early warning and alarm during the event.
At present, it is an intelligent computing center with the largest scale of self-built equipment, leading comprehensive computing power, the widest business coverage and high security level among domestic car companies. Geely is also working with the intelligent computing center with the strongest intelligent support capability of server manufacturers such as Ningchang, and based on the ultra-large-scale scientific computing cluster of the fourth-generation Intel ® Xeon Scalable processor and the application of the landing, the measured double-precision floating-point computing power reaches 3.54 PFlops, and the "super brain" of car manufacturing based on the fully intelligent business has significantly improved Geely's ® overall R&D efficiency.
Closed data loops
The second half of the decisive battle of intelligent driving
At present, the autonomous driving entering the market has undergone the evolution of sensors, algorithms, and computing power, and requires a large amount of data labeling, behavior warning, and road points. But these are only the tip of the iceberg of intelligence, or rather, the "first half" of autonomous driving. With the penetration rate of L2 assisted driving functions in new cars exceeding 30%, people have put forward higher requirements for intelligent driving, and will gradually begin to pay attention to the part under the iceberg. However, with the increase in the number of mass-produced intelligent vehicles and the sharp increase in data volume, there are still many challenges in how to obtain data in compliance, how to effectively use data, and how to transform data into product competitiveness.
Tao Shijun, Director of the Data Platform Department of the Intelligent Driving Center of Geely Automobile Central Research Institute
According to Tao Shijun, director of the data platform department of the intelligent driving center of Geely Automobile Central Research Institute, Geely is the only one with compliance collection qualifications for the main engine factory, and has accumulated a large amount of data by taking advantage of the number of vehicles. The amount of compliance data collected by a vehicle per month is on the order of several gigabytes, and the daily growth of data collected by the Star Smart Computing Center has reached the terabyte level. With such a huge amount of data, if it is not governed, the amount of effective data will be relatively low, and there is no way to use it directly. Moreover, if there is no good base platform, the flow of the entire data will also be very problematic. On the basis of the massive data accumulated by the Xingrui Intelligent Computing Center, it is necessary to continuously optimize in various scenarios and build data closed-loop capabilities: perception data closed-loop, map data closed-loop, driving data closed-loop and user data closed-loop.
Geely has accumulated a large amount of data based on user fleets and operating vehicle resources within the group to achieve the goal of data-driven perception improvement. In particular, there are more opportunities to discover corner cases (corner scenes or extreme scenes) and do some optimization iterations for regulation and algorithms.
Relying on Geely's intelligent computing center and data closed-loop technology capabilities, the goal is to use 0.1 times the computing power of the vehicle terminal, use a set of standard data formats, and use 10 times the computing power of the cloud to achieve 100 times the collection efficiency.
Data Synthesis:
A booster for the era of large models
The data closed loop involves a large amount of data, which faces two problems: first, long-term data collection is required in the field of autonomous driving, and it is necessary to focus on exploring corner case and long tail scenarios, which are widely distributed, have low probability, and high collection costs. Second, before any algorithm or system can be put on the road, it must undergo rigorous road testing, which needs to cover many scenarios and security risks.
Zhang Guangbin, founder of Yiqi Research Institute
In the view of the Yiqi Research Institute, the former problem is an "unknown danger", and the latter problem is a "known danger". As data accumulates, the former problem will gradually be discovered and transformed into scenarios that should be covered, but it will become more and more difficult to discover new extreme scenarios. The latter problem faces difficulties that are difficult to cover or reproduce, such as special climatic conditions such as freezing rain, and illegal traffic participants. Therefore, in recent years, the industry has begun to introduce synthetic data into visual problems to solve the problem that data is difficult to obtain and label.
Synthetic data can solve the problem of diversity in the first place. Corner Cases are difficult to obtain on real roads, and they can't even be reconstructed in 3D with known rules. However, synthetic data is programmatic and can generate a large number of scenarios by permutating and combining a large number of parameters, introducing randomization, etc. Increasing the amount of synthetic data in the dataset can make the trained model have better generalization capabilities.
Synthetic data is highly efficient. The annotation of real data requires a high cost, a picture needs 1~5 yuan, and the accuracy is also limited by human cognition, emotion, fatigue and other consistency problems. Generally, the accuracy of the pre-labeled model is insufficient, and a large number of manual corrections are required. The accuracy of automatic annotation of large models is higher than that of humans, which can reach 99.2%, and the efficiency is higher. Synthetic data has the natural advantage of being compliant and secure. The real-world data collected needs to be desensitized, such as license plate numbers, and is not easy to share among different enterprises.
Liang Zhenbao, Chief Engineer of Artificial Intelligence Model Development of the Prospective Innovation Department of Geely Automobile Central Research Institute
Liang Zhenbao, Chief Engineer of Artificial Intelligence Model Development of the Prospective Innovation Department of Geely Automobile Central Research Institute, introduced Geely's overall layout in synthetic data: relying on the powerful computing power of the intelligent computing center, the industry's largest asset data lake in the production center, and the large autonomous driving model, a virtual metaverse world is built for the collection of synthetic data, including some automatic annotation based on real data, similar aggregation, and some identified data governance. On this basis, it can provide some application service capabilities, including data synthesis, scenario synthesis, algorithm testing, regulatory certification, annotation and data governance, as well as digital twin smart cities. The entire system framework actually revolves around the simulation kernel, connecting the traditional design and physical sensor models, including the automated evaluation system and software and hardware test links.
Geely's synthetic data and simulation test platform has several key capabilities:
First of all, it relies on the powerful computing power of Xingrui Intelligent Computing Center.
Secondly, it has a set of high-fidelity vehicle models, whether it is your own car or another car, its characteristics are the same as the real car, including the sensor modeling in it. The traffic theory part is not only a simple straight drive or left and right lane change, but also uses AI to drive behaviors, including line pressing, malicious lane change and other behaviors. This ensures that the data is synthesized or tested with good fidelity.
Third, it's real enough. Through the artificial intelligence model, the synthesized data can be transferred to the real data style, and the style and fidelity of the migration will be improved.
Fourth, AI-based automatic modeling is more efficient than manual modeling. According to the real scene data and regulatory restrictions, thousands of kilometers of behavior scenes can be generated within an hour, and the light, weather, and lighting can be simulated. For example, if you do an overseas project, you can build it according to local rules, and you don't need to go abroad to collect it.
Finally, improve data governance. Using this platform, Geely has called back a large amount of data, which in the past may have just been lying in a data lake. After categorical governance, it is conducive to quickly find these data and analyze the distribution of data. The research center has made a set of operators based on image classification, with nine categories of strategies and more than 120 subcategories, including various weather, working conditions, obstacles, and some identification algorithms for identification, which can quickly label the information contained in the picture, and then count and govern all the mass production return data based on this label.
More than just smart driving
Electrification and intelligence complement each other, and intelligence gives new attributes to automobiles. In the era of intelligent networking and smart cities, the car is a mobile intelligent network node, receiving, generating and transmitting a large amount of data. Geely's "1 Center + 3 Platforms + N All-Scenario Data Intelligent Services" demonstrates Geely's industry-leading intergenerational advantages in intelligent computing in the new era of smart cars, ushering in Geely's "intelligent computing era".