SlotScope: Slot-Aligned Multi-Layer Observability for CPU-Resident AI-RAN Workload Assessment

Abstract

Artificial intelligence (AI) support is becoming a critical design requirement for next-generation radio access networks (RAN), with AI-RAN emerging as a leading architecture to integrate AI workloads alongside RAN functionalities. While accelerator-based platforms are widely pursued for high-demand AI-RAN applications, CPU-based Distributed Unit (DU) nodes remain a practical and cost-effective deployment point. This raises a central question: Can AI workloads share a CPU-based DU host while preserving real-time RAN operation? To answer this, we develop SlotScope, a slot-aligned, multi-layer monitoring framework that jointly captures perslot RAN events, processing time, CPU-pipeline counters, and platform state. We validate SlotScope on an over-the-air 5G NR testbed running an OpenAirInterface (OAI) DU with a commercial radio unit (RU), serving as a proof-of-concept CPU-resident AI-RAN node. Evaluating across AI-for-RAN, AI-on-RAN, and AI-and-RAN workloads, SlotScope reveals that CPU-resident feasibility is workload-dependent, with each failure traceable to a distinct limiting mechanism that no single monitoring layer alone could expose.