--- title: "End-to-end autonomous driving implementation requires 'tailored' cloud accelerators." type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/24156479.md" description: "At new car launch events, automakers announcing autonomous driving mileage has become a standard practice to showcase their capabilities. For example, in August, Aito announced that its autonomous driving mileage had reached 200 million kilometers. The reason automakers highlight these mileage figures is that the real-world data constitutes their digital assets for autonomous driving training, helping them continuously iterate and upgrade their models. This process occurs in the cloud. In fact, a continuous data loop operates 24/7 between the vehicles and the cloud—where data from the vehicles is fed back to the cloud for centralized model training and simulation..." datetime: "2024-09-27T09:56:22.000Z" locales: - [en](https://longbridge.com/en/topics/24156479.md) - [zh-CN](https://longbridge.com/zh-CN/topics/24156479.md) - [zh-HK](https://longbridge.com/zh-HK/topics/24156479.md) author: "[汽车之心](https://longbridge.com/en/profiles/3726156.md)" --- > Supported Languages: [简体中文](https://longbridge.com/zh-CN/topics/24156479.md) | [繁體中文](https://longbridge.com/zh-HK/topics/24156479.md) # End-to-end autonomous driving implementation requires 'tailored' cloud accelerators. At new car launch events, automakers announcing autonomous driving mileage has become a standard practice to showcase their capabilities. For example, in August, HarmonyOS Intelligent Driving announced reaching 200 million kilometers in autonomous driving mileage. The reason automakers highlight autonomous driving mileage is that these real-world mileage data form their digital assets for autonomous driving training, helping them continuously iterate and upgrade their models. This process occurs in the **cloud**. In fact, a continuous data loop operates 24/7 between vehicles and the cloud. Data from vehicles is fed back to the cloud, where centralized model training and simulation occur, and the updated model data is then sent back to the vehicles, forming a closed loop for OTA deployment and updates. Hence, automakers often say, "The more you drive, the better it gets," because the underlying logic supports this. This "end-to-end" approach is inherent to the training methodology of autonomous driving models. However, this differs from the industry's current hot topic of **end-to-end autonomous driving**, which refers to a technical path where previously separate modules like perception, prediction, planning, and control are integrated into a massive AI neural network, enabling rapid decision-making akin to the human brain. Interestingly, these two "end-to-end" approaches are now colliding, raising higher demands for model training: - Collecting and storing hundreds of petabytes of data; - Efficiently processing and training high-quality data; - Completing simulation tests integrating perception and control; - Ensuring compliance and security throughout the data lifecycle; ... Clearly, this requires automakers/autonomous driving suppliers to invest significant time and resources to build a mature, compliant, and stable data toolchain, supporting efficient AI model iteration. The faster the model iterates, the quicker autonomous driving products can be deployed, and the better their performance. For automakers/autonomous driving suppliers, building this foundation from scratch is a major challenge. Cloud service providers like **Baidu** see this as a huge opportunity to empower automakers, offering a comprehensive data toolchain covering data collection, labeling, management, simulation, and testing, supporting the development of autonomous driving services for many industry players. At the latest Baidu Cloud Intelligence Conference, Baidu Auto Cloud upgraded to version 3.0, optimizing the toolchain for end-to-end autonomous driving to help players build efficient data loops and overcome deployment challenges. _**01、**_**Using Generative AI to Solve the "Data Dilemma"** The end-to-end technical paradigm is widely recognized as the optimal solution for high-level autonomous driving. This AI-driven approach, akin to building a skyscraper, places greater demands on the foundational data toolchain. On one hand, **data volume is exploding, and processing complexity is rising**. A rough benchmark: - L2/L3 autonomous driving demo models require millions of images; - L2/L3 mass production requires billions of images, exceeding 100TB; - L4 demo models require petabytes of data; - L4 mass production exceeds 50PB. Clearly, as complexity increases, data processing becomes harder to manage. The autonomous driving data pipeline includes filtering, cleaning, and labeling to make raw data useful. For end-to-end autonomous driving, "useful data" means training systems to generalize like experienced drivers, handling complex scenarios like corner cases(extreme situations), such as pedestrians suddenly crossing, multi-intersection roundabouts, or sharp turns. Metaphorically, it’s about training a smart brain that not only solves routine problems but also tackles novel challenges through reasoning. Thus, training strategies must emphasize breadth and depth, using diverse and challenging data. This requires quickly mining and labeling high-quality data from vast databases to create robust training sets. On the other hand, **simulation training logic is changing, posing new challenges**. Simulation is the final defense in autonomous driving R&D, acting as an evaluation system. High-scoring models proceed to deployment, while low-scoring ones are debugged and retrained. Current evaluations fall into two categories: open-loop and closed-loop. - Open-loop assesses individual tasks like perception or planning against ground truth; - Closed-loop creates feedback loops in virtual worlds, aligning with the industry’s "**world model**" concept. End-to-end autonomous driving integrates perception and planning, necessitating **closed-loop evaluation**, which demands toolchains supporting this unified approach. Another challenge is the need for massive data to build realistic virtual worlds simulating real-world scenarios (e.g., puddles, floating bags). These must cover long-tail cases comprehensively, as flawed evaluations render training futile. To address these demands, Baidu Auto Cloud upgraded its data toolchain in two key areas. First, **adding intelligent data search (text-to-image, image-to-image)**. This helps quickly filter and retrieve high-value data (e.g., flooded roads) for model training. Second, **using generative AI to repurpose real-world data for simulation**. Normally, high-quality data is used once, but generative AI can modify scenes (e.g., removing obstacles, adding vehicles) to create new scenarios, reducing costs for data-scarce companies. NVIDIA’s global VP and automotive head Jensen Huang believes end-to-end models will redefine cars with limitless AI rules. Generative AI and autonomous driving are reshaping intelligent experiences, advancing high-level autonomy. To win in this era, players must **leverage AI’s transformative power**, as Baidu is doing. Baidu, as a top autonomous driving player, leverages its decade of research and vast data (e.g., Baidu Maps’ city-wide coverage, millions of test kilometers) to build superior simulation platforms. _**02、**_**Full-Pipeline Training Optimization: Maximizing Compute** One supplier noted data consumes over 80% of end-to-end autonomous driving R&D costs, including compute. Massive data requires powerful compute, especially for concurrent simulation tasks stressing CPU/GPU resources. Compute has become a battleground, driving two trends. First, **building AI data centers**. Tesla invested billions in supercomputers, aiming for 100 exaFLOPS by year-end. Chinese players, while less funded, are also pushing limits. Second, **developing world models**. End-to-end autonomy’s validation demands make world models/simulation platforms critical, especially with generative AI enhancing realism—at high compute costs. End-to-end autonomy starts at ~1000P compute, scaling up with heavier costs. Low efficiency worsens this, making **maximizing limited compute** key. Baidu’s **Baidu AI Heterogeneous Computing Platform 4.0** offers cost-efficient solutions. Baidu AI Heterogeneous Computing Platform 4.0 The platform emphasizes multi-GPU efficiency, supporting diverse chips (A100, A800, domestic) and 10,000-card clusters to boost performance. Its end-to-end acceleration adapts to various frameworks, avoiding inefficiencies from poor framework-model coupling via auto-tuning for parallelism, memory, and operators. Key metrics: - GPU sharing halves auto-labeling costs; - Training throughput rises 138% (up to 400%), cutting time by 80%; - Simulation supports 1M km/day. One automaker achieved 170% faster training and 2.5x GPU efficiency with Baidu. As robust AI infrastructure, the platform maximizes compute across scenarios, boosting R&D efficiency. _**03、**_**Cloud-AI Integration: Accelerating the Intelligent Future** The next phase hinges on automotive cloud capabilities, defining performance ceilings. Two trends emerge in this ecosystem. First, **vehicle-road-cloud integration**. Baidu’s partnerships with traffic agencies enable dynamic road/weather alerts via AI, improving route planning without navigation—enhancing smart cabin experiences. Second, **smarter cabins**. AI models shift cabins from multimodal interaction to proactive AI, predicting needs (e.g., voice control, automatic navigation, "sentry mode" for incident recording). As autonomy and cabins advance, cloud providers like Baidu compete on data toolchains. Pony.ai’s CTO notes **data chain maturity determines model quality**, crediting Baidu’s L2-L4-ready Apollo platform. Baidu leads China’s AI cloud market (26.4% share, #1 for 5 years), empowering automakers through open collaboration, turning smart driving visions into reality. $Baidu(BIDU.US) ### Related Stocks - [Baidu, Inc. (BIDU.US)](https://longbridge.com/en/quote/BIDU.US.md) - [BIDU-SWR (89888.HK)](https://longbridge.com/en/quote/89888.HK.md) - [BIDU-SW (09888.HK)](https://longbridge.com/en/quote/09888.HK.md)