Horizon Robotics (地平线)

Solution

Baidu AI Cloud became Horizon's largest infrastructure supplier and rebuilt the core training stack for mass-production autonomous-driving models. On storage, Baidu optimized the bottom architecture to deliver EB-scale BOS object storage plus high-performance PFS parallel file storage for large data ingest and training throughput, including millisecond cross-cloud link support under very high query loads. For compute, Horizon deployed a very large CCE container cluster that can manage more than 100,000 concurrent pods and unify management of over 10,000 GPU cards, improving both operational efficiency and GPU pooling. Baidu and Horizon also automated GPU fault discovery and recovery so failures could be handled in minutes instead of interrupting long-running training. Network upgrades added dedicated links and bandwidth expansion to keep the training environment highly available. The overall design created a scalable AI-infrastructure base for Horizon's growing autonomous-driving portfolio while also adding data-compliance support for R&D and production workflows.

Data flow

Autonomous-driving data is stored in BOS and PFS, large training jobs run on the CCE container cluster with pooled GPU resources, automated monitoring handles GPU failures in minutes, and the upgraded network keeps data and model workloads moving reliably across the high-concurrency training environment.