The Evolution of CU-DU Split Architecture for AI-Native 6G RAN

CU-DU split has already proven flexible deployment, CU resource pooling, seamless mobility, RRC Inactive optimization, and multi-vendor interoperability in 5G.

Placing the DU near the RU while positioning the CU centrally or at the edge enables configurations tailored to regional needs, and splitting the CU's UP into multiple CU-UPs allows low-latency services and eMBB to be deployed in different locations. This also supports configurations like NTN, where the DU is placed on a satellite and the CU on the ground.

A centralized CU allows the L3 resources of multiple DUs to be shared as a pool, enabling elastic scaling of compute resources at edge data centers when traffic surges. As a result, it reduces the inefficiency of over-provisioning L3 functions such as RRC and PDCP at every base station, and helps address highly variable 5G-Advanced/6G workloads like XR, AI-inference, and sensing.

There are also significant benefits in terms of mobility and control. Within the same CU domain, the UE's mobility anchor is maintained, reducing the burden of context transfer and reconnection, and the CU can aggregate load, QoS, and mobility patterns from multiple DUs to perform more sophisticated RRM. Even in RRC Inactive mode, the need for repeated location updates whenever a UE crosses DU boundaries is reduced.

The standardized F1/E1 interfaces allow CUs and DUs from different vendors to be combined, increasing the feasibility of multi-vendor Open RAN. The article explains that Tier-1 operators in North America and Asia have already applied multi-vendor CU-DU combinations in commercial networks, and views this as validation of performance stability and operational maturity.

Moving into 6G, the CU evolves into an Intelligent CU, becoming the center for embedding AI/ML deeply into the RAN. The AI-L3 use cases presented in the article are as follows.

• Mobility Optimization: Predicting optimal handover targets and RLF using user mobility history and signal history
• Load Balancing: Predicting load across multiple DUs to proactively distribute traffic
• QoE Prediction and Control: Detecting experience degradation in advance to adjust resources or slices
• Energy-Aware Traffic Control: Linking RU/DU power states with traffic to improve energy efficiency

As AI functionality increases, the value of centralized, aggregated compute also grows. Rather than attaching separate AI servers and accelerators at each site, as in integrated base stations, expanding the compute pool at the CU is more efficient in terms of both investment and operations.

AI management also moves closer to the CU. Previously, model training, monitoring, and deployment were mainly handled within OAM/SMO, but by incorporating model onboarding, inference orchestration, training, and validation into the CU, AI models can be retrained and fine-tuned more quickly in response to changing conditions, based on real-time data coming in from multiple DUs.