INSPECT: Data analysis tool for Proactive Link Failure Detection on Meta's AI Infrastructure#
OCP Global Summit 2025 digest Program Core pick 13–16 October 2025
Speakers: Granthana Rangaswamy (Meta), Arushi Sharma (Meta)
- Why: At 100K-GPU scale, interconnect health is the availability problem, and this is a deployed ML-based early-warning system over SerDes telemetry with a concrete standardization angle — collecting SerDes parameters uniformly across vendors.
- What: INSPECT, a software daemon that collects and analyzes SerDes parameters to predict link failures before they cause unplanned downtime and job restarts, used for new-system provisioning, training-job deployment, and continuous monitoring across Meta's AI systems. (from abstract only)
- How: ML-based anomaly detection with pass/fail criteria over SerDes telemetry; standardizes collection and analysis across SerDes vendors by mapping common SerDes building blocks to a predefined schema, enabling a single trained model across heterogeneous AI systems.
- Where applicable: Meta AI infrastructure (100K-class clusters); the vendor-neutral SerDes schema idea is the generalizable piece for any operator running multi-vendor optics/SerDes fleets. Linked materials point to related NSDI work.