Engineering resilient AI infrastructure: Cross-layer observability at scale

  • Date
    2 November 2026
    Timeframe
    16:00 - 17:00 Seoul, KST
    Duration
    1 hours

      As AI infrastructure becomes increasingly dense and interconnected, even minor failures can have far-reaching consequences across large-scale workloads. At the same time, performance bottlenecks are continuously shifting from compute to memory and networking, while increasingly complex interconnect architectures span front-end and back-end networks, as well as scale-up and scale-out domains. These challenges demand a new approach to infrastructure observability. Monitoring individual hardware components or software services in isolation is no longer sufficient. Reliable AI operations require vertically integrated telemetry and cross-layer correlation to connect hardware health, network behavior, workload execution, and orchestration decisions — enabling faster root-cause analysis, more effective recovery, and better resource utilization. In this webinar, we share engineering lessons from Korea’s sovereign AI foundation model project, drawing on our experience operating a large-scale GPU training cluster. Through real-world cases involving storage bottlenecks, GPU failures, and automated workload recovery, we illustrate the importance of connecting infrastructure-level signals with software-level operational context. We conclude by discussing how cross-layer observability and software-driven orchestration can form the foundation for resilient, efficient, and scalable AI infrastructure, with implications for the evolution of AI-native communication networks.

      Are you sure you want to remove this speaker?