Centralized RAN Architectures

• Centralized RAN architectures are network designs that centralize baseband processing from multiple radio access points, enabling statistical multiplexing and improved inter-cell coordination.
• They employ techniques like dynamic functional splits and Coordinated MultiPoint to enhance throughput, reduce interference, and adapt to variable network demands.
• These systems balance pooling gains with strict fronthaul bandwidth, latency, and energy trade-offs, making optimization of resource allocation and processing critical.

A centralized radio access network (centralized RAN, or C-RAN) refers to a cellular network architecture in which the baseband processing from multiple radio access points is aggregated and virtualized at a central data center, leveraging high-speed, low-latency fronthaul transport. This design entails shifting functions traditionally performed at site-specific base stations to a pooled cluster of baseband processing resources, allowing for statistical multiplexing, enhanced inter-cell coordination, elastic compute provisioning, and advanced interference management. Centralized RAN is particularly advantageous in dense deployments (e.g., enterprise and urban hotspot scenarios) where underutilization of local resources and strong intra-cluster interference are prevalent (Rost et al., 2015).

1. Architectural Principles and Functional Decomposition

In centralized RAN, a set of $N_\mathrm{c}$ radio access points (RAPs), each equipped with a remote radio head (RRH), connect via high-capacity, low-latency fronthaul—typically dark fiber or equivalent—to a central data center. The RRH performs only the radio frequency (RF) front-end, down/up-conversion, sampling, and optionally some low-order PHY operations (e.g., FFT/IFFT), before digitizing and packetizing the signal. This digitized baseband data is transported to a central pool of virtual baseband units (vBBUs), implemented as virtual machines or software containers, where all joint baseband processing—including FEC decoding, demodulation, and interference coordination—is executed on a cloud-native server cluster (Rost et al., 2015, Simeone et al., 2015).

Uplink processing typically flows as follows:

1. The UE transmits over the air to its serving RAP.
2. The RAP digitizes and packetizes the received waveform (optionally performing FFT).
3. The fronthaul carries these packets to the data center.
4. A centralized scheduler in the vBBU pool assigns modulation/coding schemes (MCS) and scheduling blocks to decoding engines.
5. Advanced decoders (e.g., turbo-decoders) recover the transmitted bits, which are then conveyed to higher layers.

Key to C-RAN is the functional split, with a spectrum of options defined in 3GPP TR 38.801 (Options 1–8), determining which processing blocks are centralized (at the CU), distributed (DU), or remain at the RU (Simeone et al., 2015, Rihan et al., 2023). Fully centralized (Option 8) centralizes all PHY and above, with only RF left at the RRH, while intermediate splits (e.g., 7.x) localize selective PHY functions at the edge to mitigate fronthaul demands.

2. Computational Load, Pooling, and Complexity Metrics

Quantifying the computational requirements of a centralized RAN is central to dimensioning the BBU pool and understanding architectural scalability. The computational effort per transport block (TB) is dictated by the chosen MCS, channel state (instantaneous SNR $\gamma$), and decoder algorithmic complexity.

Letting $r(\gamma, \Delta\gamma)$ denote the rate adaptation function (modulation/coding as a function of $\gamma$ and SNR-margin), each code block of $D_k$ bits and $C_k$ code blocks is decoded with an average number of iterations $E[L_r(\gamma, \Delta\gamma)]$ over $S_\mathrm{RE}$ resource elements (REs), yielding the normalized per-TB complexity: $$\mathcal{C}(\gamma, \Delta\gamma) = \frac{D_k\,C_k\,E[L_r(\gamma, \Delta\gamma)]}{S_\mathrm{RE}}$$ (Rost et al., 2015). For $N_\mathrm{c}$ RAPs aggregated at the data center, the total instantaneous load is $\gamma$0. This sum must be served by a total virtual CPU (vCPU) budget $\gamma$1, where $\gamma$2 is the per-RAP provisioned complexity resource (Rost et al., 2015).

Key performance metrics introduced include:

• Computational outage probability $\gamma$3: Probability that total computational demand exceeds the provisioned budget before decoding deadlines, i.e., at least one TB cannot be processed in time.

$\gamma$4

• Outage complexity $\gamma$5: Minimum per-RAP complexity that achieves a target outage.
• Computational gain $\gamma$6: Ratio of total computational resource needed with versus without pooling.
• Computational diversity $\gamma$7: Sensitivity of outage to pooling size.
• Complexity-rate tradeoff: Bit-iteration overhead needed to increase throughput by 1 bps/Hz (Rost et al., 2015).

Pooling a moderate number of RAPs (e.g., 8–16) realizes 80% or more of the asymptotic complexity gain, typically reducing required CPU provisioning by 30–50% per cell and boosting per-vCPU throughput (Rost et al., 2015).

3. Functional Split Selection, Fronthaul Constraints, and Dynamic Adaptation

Selecting the optimal location of functional splits is a core challenge in C-RAN. Centralizing more PHY functionality yields higher pooling and coordination gains but imposes stringent fronthaul bandwidth and latency requirements. Conversely, distributing more PHY baseband processing at the edge (e.g., at the DU) relaxes fronthaul requirements but diminishes centralization gains (Pérez-Romero et al., 2024, Rihan et al., 2023).

Recent work defines the functional split placement as a constrained optimization problem: minimizing total energy (or compute cost) subject to fronthaul bandwidth and latency limits. The per-sector split vector $\gamma$8 is chosen to minimize $\gamma$9 under fronthaul capacity constraints (Pérez-Romero et al., 2024). Dynamic split adaptation, orchestrated via O-RAN’s Service Management & Orchestration (SMO), rApps, and NFV-MANO, can realize 4–25% energy savings versus static splits while maintaining feasibility under variable site load and transport constraints.

When fronthaul is a bottleneck, shifting low-PHY functions (e.g., FFT/IFFT, RE mapping) to edge devices drastically reduces transport requirements. For example, transitioning from fully centralized (CPRI-style Option 8, $r(\gamma, \Delta\gamma)$07–10 Gb/s) to intra-PHY split 7.3 can cut fronthaul demand by $r(\gamma, \Delta\gamma)$150$r(\gamma, \Delta\gamma)$2 for a 20 MHz cell, making high-density RAN deployment practical (Rodriguez et al., 2021, Simeone et al., 2015).

4. Advanced Coordination, Mobility, and Cooperative Processing

Centralized processing enables advanced inter-cell coordination strategies, notably Coordinated MultiPoint (CoMP) transmission, joint scheduling, and global interference management (Simeone et al., 2015). C-RAN systems can utilize global channel state information, as all baseband traffic from multiple RAPs is available to the central scheduler. This enables frame-by-frame centralized weighted sum-rate optimization under power, fronthaul, and QoS constraints (e.g., via weighted MMSE or group-sparse formulations to activate/deactivate RRHs) (Simeone et al., 2015).

Centralized control extends to SDN-based architectures where all RRC/RRM and mobility decisions (attach, handover) are executed centrally in the core, with RAPs acting as pure data-plane nodes. Such structures can reduce end-to-end signaling overhead and handover delay by 10–30%, improve attach times, and enable global, load-aware mobility management algorithms (Nayak et al., 2018).

Centralization also allows for user-centric clustering, in which dynamic, overlapping clusters of RRHs jointly serve each user, thereby mitigating the classical "cluster-edge" penalty. When acquiring global CSI is infeasible, architectures rely on partial instantaneous CSI within user-centric clusters and coarse large-scale CSI externally, using robust optimization techniques (successive convex approximation, chance/worst-case constraints) for beamforming under estimation error and fronthaul limits (Pan et al., 2017).

5. Resource Statistical Multiplexing, Latency, and Energy Considerations

Elastic sharing of computational resources across multiple RAPs captures statistical multiplexing gain—peaks in demand for one cell are absorbed by valleys in others, reducing overprovisioning. Queuing models show that long-term computational multiplexing (pooled server clusters with slow adaptation) can achieve nearly the same resource savings as idealized per-frame allocation, with much lower implementation complexity and negligible additional delay (for realistic frame durations and latency budgets) (Kalør et al., 2017).

The upper bound on energy/resource savings is $r(\gamma, \Delta\gamma)$3, where $r(\gamma, \Delta\gamma)$4 is the normalized load. For latency-constrained services (e.g., URLLC), pooling remains viable provided per-user queuing delays are dimensioned to meet required percentiles using the analytical models described above.

However, energy savings by centralizing BB baseband processing (BBP) must be weighed against increased fronthaul transmission power, especially for deep splits that push all 10–11 Gb/s eCPRI into the core and across multiple network hops (Tariq et al., 30 May 2025). Minimizing total power often favors deeper centralization (at the DC or O-CU), but only when high fanout (i.e., sufficient pooling across multiple RUs) offsets transport energy growth (Tariq et al., 30 May 2025).

6. System-Level Performance, Case Studies, and Design Guidelines

Empirical studies and field trials confirm significant benefits:

• Centralized processing of 8–16 cells can cut baseband CAPEX/OPEX by 20–40%, with edge throughput improvements of 20–40% under CoMP (Simeone et al., 2015).
• Adaptive split management (dynamic sector-by-sector split at the O-RAN SMO/non-RT RIC) reduces energy cost by up to 25% compared to static splits (Pérez-Romero et al., 2024).
• Soft Actor-Critic DRL policies for centralized resource management yield higher throughput-overhead-complexity (TOC) compared to static centralized/distributed policies, especially under variable network load (Nouruzi et al., 2022).
• Dynamic, parallelized channel decoding at the CO can aggregate tens of gNodeBs with strict real-time latency between radio edge and centralized processing (Rodriguez et al., 2019).

Design best practices include:

• Pool $r(\gamma, \Delta\gamma)$5–$r(\gamma, \Delta\gamma)$6 RAPs per data center cluster to capture pooling gains while keeping fronthaul latency low (Rost et al., 2015).
• Provision fronthaul headroom above the raw transport rates to avoid bottlenecking decoding deadlines (Rost et al., 2015, Rodriguez et al., 2021).
• Centralize BBP as deeply as fanout and transport allow, but move low-PHY edgewards when fronthaul costs/latency dominate (Tariq et al., 30 May 2025, Rodriguez et al., 2021).
• Operate at moderate SNR margins (e.g., $r(\gamma, \Delta\gamma)$7–$r(\gamma, \Delta\gamma)$8 dB), since tighter margins increase the risk of complexity outage, but overly loose margins waste spectral efficiency (Rost et al., 2015).

7. Open Challenges, Standardization, and Future Directions

Future research challenges focus on:

• Flexible, dynamic functional splits and orchestration: O-RAN-based split-selection agents, leveraging rApps and NFV-MANO, for real-time adaptation under traffic, fronthaul, and compute constraints (Pérez-Romero et al., 2024).
• AI-driven fronthaul adaptation and function placement, integrating cross-layer metrics (latency, energy, traffic forecasts) (Wang et al., 5 Nov 2025, Nouruzi et al., 2022).
• Hardware heterogeneity support (CPU, GPU, NPU allocation in the DC/edge) and ultra-tight synchronization for joint transmission and multi-point access (Wang et al., 5 Nov 2025).
• Full integration with non-terrestrial platforms (NTN), where high- and low-layer split selection must consider ultra-long RTTs and power-constrained payloads; dynamic orchestration is critical (Rihan et al., 2023).
• Standardization around open interfaces (eCPRI, O-RAN 7.2x), scalable packet-based fronthaul networks, and robust, vendor-neutral disaggregation (Hasabelnaby et al., 2024).

Centralized RAN architectures anchor the evolution of cellular networks toward intelligent, scalable, and energy-efficient infrastructures, enabling advanced cooperative transmission and flexible function placement. Ongoing advances in split optimization, statistical resource pooling, and AI-driven orchestration are central to meeting 5G-Advanced and 6G requirements for ultra-dense, heterogeneous, and globally-connected radio access (Rost et al., 2015, Pérez-Romero et al., 2024, Simeone et al., 2015, Hasabelnaby et al., 2024, Tariq et al., 30 May 2025, Wang et al., 5 Nov 2025).