Case Study | 400G Intelligent Computing Network Helps Leading Autonomous Driving Company Improve Computing Training Efficiency

According to Gartner's forecast, autonomous driving technology is developing rapidly and is expected to bring significant commercial benefits in the coming years, especially in the fields of decision intelligence and Edge artificial intelligence (AI). Currently, a leading company is actively embracing the path of digital transformation based on large models. As a leader in the autonomous driving field, this company is actively responding to this trend. The company focuses on intelligent cockpits, autonomous driving technology, and connected services, continuously developing highly integrated intelligent hardware and cutting-edge software algorithms to create intelligent and efficient integrated mobility solutions for consumers.

Currently, the company urgently needs to upgrade its intelligent computing center to meet the growing demand for large-scale model training and providing comprehensive autonomous driving model solutions for customers.

Intelligent computing networks need to meet the demand for improving the training efficiency

To ensure the smooth construction and efficient operation of the autonomous driving intelligent computing center, attention needs to be paid to the high reliability of the hardware, the cost-effectiveness of the networking scheme, and the scalability of the network bandwidth. The company identified three core requirements:

First, given the intense competition in the new energy vehicle market, the solution must minimize the construction time to ensure a swift and effective response to dynamic market changes.

Second, building an intelligent computing center is a long-term and high-cost investment; the solution must maximize cost-effectiveness and maintain stringent cost control.

Finally, considering the rapid iteration nature of intelligent computing centers, the scalability of the intelligent computing network must be considered to meet the company's continued business growth and development needs.

Specifically, in terms of hardware configuration, high-performance, high-reliability servers, storage devices, network equipment, and professional GPU clusters must be selected to meet the stringent computing requirements of autonomous driving algorithm training. Optimizing the existing InfiniBand (IB) network solution is critical, considering both business costs and training efficiency. This requires a comprehensive evaluation of the IB network solution compared to other network solutions to ensure that performance requirements are met while minimizing costs. Simultaneously, sufficient space and capacity must be reserved to accommodate future technology upgrades and expansion needs, ensuring that the intelligent computing center's technology does not quickly become obsolete, thus maintaining its long-term competitiveness and market adaptability.

How to Build a High-Speed ​​and Efficient Intelligent Computing Center for Autonomous Driving

After thorough verification and testing, H3C's RoCE (RDMA over Converged Ethernet) network solution can seamlessly integrate with customers' existing systems and, in most scenarios, achieves the performance, reliability, and scalability standards of IB networks.

The intelligent computing network solution adopts a dual-plane network architecture with separate storage and compute. The core components include:

A 400G lossless computing network, consisting of 42 S9825-64D data center switches;

A 100G lossless storage network consisting of 12 S9820-64H data center switches, supporting 17 nodes of UniStor CX5036G6 distributed high-performance parallel storage. The solution uses a RoCE Ethernet network architecture combined with an innovative Layer 2 box-box architecture design, effectively meeting the needs of 100 high-performance GPU servers in the initial network construction, while also reserving future expansion capabilities. The more mature and efficient RoCE technology architecture significantly reduces the deployment time, reduces training time, and lowers costs, providing customers with a higher return on investment.

S9825-64D Data Center Switch Throughput Test

In a dual-plane network architecture, the forwarding plane is dedicated to network data transmission, while the control plane handles network management and control information processing. This architecture significantly improves network reliability and security through mutual backup between the two planes. If one plane fails, the other can immediately take over its functions, ensuring continuous network operation and stability. In contrast, a single-plane network architecture concentrates all network functions on a single plane, failing to separate control and data forwarding, thus limiting its flexibility and security. Clearly, the dual-plane network architecture offers significant advantages in ensuring network stability and security.

Intelligent Computing Center Model Training Efficiency Improved by 11.1%, Accelerating Enterprise Large-Scale Model Development

H3C's RoCE intelligent computing network solution maintains computing performance comparable to IB networks, enabling enterprises to handle large data volumes. Its low latency and high throughput characteristics significantly shorten the training time for enterprise autonomous driving models, while improving model training efficiency and accelerating business processing speed.

The 400G RoCE network facilitates future bandwidth upgrades for enterprises. Based on Ethernet technology, RoCE boasts a mature and extensive ecosystem, which is conducive to future technology upgrades and solution evolution. Data processing capacity is expected to increase by 50% within the next two years without requiring large-scale network architecture changes. Furthermore, the investment in RoCE networks is expected t to achieve a full return within three years, driven by substantial cost savings and enhanced business efficiency. This transition is expected to deliver a 10% increase in overall Return on Investment (ROI).

From a deployment efficiency perspective, although RoCE networks still require manual configuration, the widespread adoption and maturity of Ethernet technology reduces deployment time by an average of 15%, minimizing downtime and labor costs, thereby indirectly saving overall costs. Compared to IB networks, RoCE typically has lower equipment and maintenance costs, which is highly advantageous in controlling overall business costs. Due to the ubiquity of Ethernet components, replacement and maintenance are more economical; from an overall project accounting perspective, the maintenance cost of a RoCE network is approximately 20% lower than that of an IB network.

From a deployment and cost perspective, RoCE networks offer significant advantages over InfiniBand (IB). Although both require manual configuration, RoCE's foundation in widespread Ethernet technology reduces deployment time by an average of 15% through saving overall costs by decreasing downtime and labor expenses. Furthermore, RoCE typically features lower direct equipment and maintenance costs. Given the ubiquity of Ethernet components, replacement and maintenance are more economical, leading to an overall project maintenance cost that is approximately 20% lower than that of IB networks—a highly advantageous factor in controlling long-term business costs.

Clearly, the introduction of the 400G intelligent computing network not only improves the efficiency of computing training for autonomous driving enterprises but also provides strong technical support and cost advantages for their future development. With continuous technological advancements and expanding market demand, this company is poised to occupy a more significant position in the global autonomous driving market, contributing to the future of intelligent mobility. This transformation not only marks a new era for autonomous driving technology but also points the way for the development of the entire automotive industry. Powered by 400G intelligent computing networks, the company will be able to develop and deploy advanced autonomous driving models more quickly, laying a solid foundation for a safer and smarter travel experience.