The SK Telecom and Intel white paper details an AI pipeline that improves network quality prediction by processing massive 5G data flows with CPU-based architecture. Managing 400,000 cell towers and 120B daily records, SKT replaced GPU-dependent systems with a unified pipeline on Intel Xeon servers using FlashBase, Spark, TensorFlow, and Analytics Zoo. A memory-augmented model, combining recent and historical data with attention mechanisms, predicts sudden KPI shifts more accurately. This integrated in-memory pipeline cut delays, boosted training throughput 30–50%, and delivered inference speeds up to 6x faster than GPU setups, enabling proactive service quality management.

Join for free to read