Non-Real-Time RIC (Non-RT RIC)

The defining temporal boundary of the Non-RT RIC is its control loop latency of greater than one second. Unlike the Near-RT RIC, which executes microsecond-to-millisecond logic, the Non-RT RIC focuses on long-term policy optimization and network-wide coordination. This timescale allows for computationally intensive AI/ML model training and inference that would be impossible at the edge.

• Training cadence: Models are trained on historical data aggregated over minutes, hours, or days
• Policy deployment: Updated policies are pushed to Near-RT RICs via the A1 interface
• Use case fit: Energy saving, coverage optimization, and traffic steering across multiple cells

The Non-RT RIC serves as the execution environment for rApps (RAN Applications), which are modular, microservice-based software components that implement specific optimization functions. Each rApp operates independently, consuming data from the SMO's data lake and producing policy recommendations.

• Modular architecture: rApps can be developed by third parties and deployed independently
• AI/ML integration: rApps embed trained models for prediction, classification, and optimization
• Policy output: rApps generate A1 policies that govern Near-RT RIC behavior
• Lifecycle management: The SMO framework handles rApp onboarding, versioning, and termination

The Non-RT RIC is the sole termination point for the A1 interface, the standardized O-RAN protocol that carries policy guidance from the centralized intelligence to the distributed Near-RT RICs. This interface is the critical north-south communication channel in the RIC hierarchy.

• Policy types: A1 carries declarative policies, enrichment information, and ML model updates
• Protocol: Uses RESTful APIs with JSON payloads for policy creation and management
• Feedback loop: Near-RT RICs report policy effectiveness back via A1, enabling closed-loop refinement
• Multi-vendor interoperability: Standardized by O-RAN Alliance Working Group 2

The Non-RT RIC is a logical function embedded within the Service Management and Orchestration (SMO) framework. It is not a standalone entity but rather a core subsystem of the broader management plane, sharing infrastructure with FCAPS functions and the O1 interface termination.

• Shared data lake: Accesses unified RAN telemetry collected via the O1 interface
• Co-located logic: Operates alongside traditional network management functions
• Orchestration coupling: Policy decisions can trigger automated provisioning workflows
• Single pane of glass: Contributes to the unified management view of the entire RAN

The Non-RT RIC hosts the end-to-end AI/ML workflow required to build, train, validate, and deploy models that optimize RAN behavior. This includes data ingestion, feature engineering, model selection, training, and policy packaging.

• Training data: Aggregated from millions of UEs and thousands of cells via O1
• Model registry: Maintains versioned, validated models ready for deployment
• Offline training: Leverages the >1 second loop to perform computationally expensive training jobs
• Continuous improvement: Models are retrained as new data arrives, adapting to changing network conditions

The RIC hierarchy cleanly separates concerns: the Non-RT RIC handles strategic, network-wide optimization while the Near-RT RIC executes tactical, per-UE decisions. This division prevents latency-sensitive functions from being bottlenecked by centralized processing.

• Non-RT RIC: >1s loop, hosts rApps, terminates A1, resides in SMO, network-wide scope
• Near-RT RIC: 10ms–1s loop, hosts xApps, terminates E2, resides at edge, per-cell scope
• Policy flow: Non-RT RIC sets guardrails; Near-RT RIC enforces them in real-time
• Complementary intelligence: Both layers run AI/ML, but at different granularities and timescales