Energy Consumption in Next Generation Radio Access Networks

1

Energy Consumption in Next Generation Radio Access Networks

arXiv:2605.11899v1 [cs.NI] 12 May 2026

Urooj Tariq, Rishu Raj, Merim Dzaferagic and Dan Kilper CONNECT Centre, Trinity College Dublin, Ireland. Email: {tariqu, rajr, merim.dzaferagic, dan.kilper}@tcd.ie

Abstract—The radio access network (RAN) accounts for the largest share of energy consumption in mobile networks, making it essential to understand how and where this energy is used, particularly as future networks move toward higher levels of densification. Open radio access networks (O-RAN) have emerged as a promising approach to support this evolution through open interfaces that enable a multivendor environment, support for hierarchical intelligent controls, and simplified, cost-effective radio units that facilitate large-scale deployments. This paper examines the energy consumption in next-generation RAN architectures through transaction-based energy models. The model captures both processing and transmission energy components and evaluates how energy use varies with the placement of baseband processing (BBP) across network nodes and with different levels of network densification. Results indicate that processing energy dominates total consumption and that the location of BBP strongly influences overall energy efficiency. These insights can inform the design of future RAN deployments that balance flexibility, cost, and sustainability. Index Terms—Energy Consumption, Radio Access Networks, O-RAN, Energy Modeling.

I. I NTRODUCTION OMMUNICATION networks have evolved from systems primarily designed for human communication to infrastructures increasingly dominated by connected devices. While billions of people continue to use the network for everyday services, an even larger number of devices are now connected, generating and exchanging data autonomously. This evolution has driven a transition towards a world in which the number of connected devices exceeds the human population, creating significant challenges in terms of scalability, latency, and energy efficiency. According to estimates, there are between 25 to 50 billion devices being used, which produce 79.4 zettabytes of data annually [1]. In 2023, 65.7% of the global population [2] was using the Internet for applications with diverse service requirements, including high data rates, low latency, and high mobility. The development of the fifth generation (5G) standard introduced a new paradigm that aimed to address these diverse performance requirements by defining three service categories: enhanced mobile broadband (eMBB), ultra-reliable low latency communication (URLLC), and massive machine type communication (mMTC). However, as deployment progressed, it became clear that the boundaries between these categories were not distinct, since emerging applications often require multiple capabilities simultaneously. Consequently, research for sixth generation (6G) networks is focusing on developing an AI-native architecture, where

C

artificial intelligence is embedded into the network design to enable real-time adaptation to dynamic vertical requirements and service conditions. This increased reliance on virtualization and software-defined control has also benefited the development of open networking standards such as the open radio access network (O-RAN) specification (www.o-ran.org). O-RAN defines open and interoperable interfaces that allow a multivendor environment and enable intelligent control and management applications, which in the future are expected to rely heavily on AI methods. Its architecture disaggregates the network into radio units, and distributed and centralized processing units, simplifying the deployment of cost-efficient radio units that are well-suited for dense network scenarios expected in future 6G systems. The exponential growth of Internet traffic has frequently drawn concern over the potential for a rapid expansion of equipment deployments and corresponding increases in energy use. Much attention was focused on this more than a decade ago, when it was first recognized that the energy efficiency improvements in network equipment were slowing and not keeping up with traffic growth [3]. In order for traffic to scale exponentially, the network equipment efficiency, measured in energy per bit (transported), must reduce exponentially. This scaling requirement was originally met by scaling the transmission speeds or capacity exponentially, with little to no increase in power consumption. This capacity-based scaling slowed as both wireless and then wireline systems approached the Shannon capacity limits. At the same time, electronic processors used in network equipment were facing their own scaling challenges due to the ’one volt wall’ and other changes that slowed the energy efficiency improvements associated with Moore’s Law [4]. These studies predicted significant increases in network equipment energy use over time. As a result, one would expect that with the speed and capacity increases of networking equipment of the past decade, since those studies, the energy use and thermal density of telecommunications equipment have shown a significant increase. In fact, the most recent generation of 800G optical transceivers is being offered with liquid cooling in order to address the high thermal densities [5]. The impact of these equipment efficiency trends on actual commercial network energy use is complicated due to the relative mix of legacy and new equipment deployments and efficiencies in business operations specific to each operator. Nevertheless, a number of studies have found evidence of industry-wide network sector increases in electricity use, even growing faster than computers and data

2

Fig. 1. Energy consumption per user in access switches over time with a technology improvement rate of 20%.

centers [6]. The global climate crises, likewise have significantly shifted our attention towards scrutinizing, discussing, and reducing the energy consumption and carbon footprint of the information and communication technology (ICT) sector, which was 1.27 Gt of carbon dioxide emissions (CO2e ), or 2.3 % of global emissions in 2020. These trends motivate a re-examination of the energy use that has occurred since those previous predictions to determine whether the predicted rates of increase have borne out. Ethernet switches are an important category of network equipment that depend strongly on semiconductor scaling and have not seen a significant shift in their architecture or use. Assuming semiconductor scaling is the dominant contributor to energy use, an exponential time-dependent model has been used for the prediction of energy consumption (or energy per bit) expressed as E(t) = E0 (1 − µ)t−t0 where µ is the annual rate of technological improvement, E0 is the energy consumption of the access switches at t0 . In previous studies, the technology improvement rate was expected to continue at 20% per annum [7]. We carried out an analysis of a family of Ethernet switches from a common vendor developed for access and aggregation networks. Using a single vendor and switch family ensures that design changes and vendor differences are not a major factor. In Fig. 1, we show the energy consumption of these access and aggregation switches based on public datasheets from the year 2008 to 2023. We also show a linear trend on the log-linear scale corresponding to µ = 0.2, with an R2 of 0.848 in the log-domain. This indicates good agreement with the observed data and corroborates previous predictions, while also supporting the use of µ = 0.2 as a representative value. Understanding the energy consumption of network equipment is complex and has been approached through several different methods. The most commonly used method is a socalled ‘top-down’ economic-based model that estimates the overall network power consumption from publicly reported network operator business operations figures or equipment sales figures [8]. This gives an accurate indication of energy use for the sector overall, but does not provide insights into how the network is configured or operated in order to arrive at the total energy consumption numbers, which might include

legacy equipment or redundant deployments. In order to better understand the evolution of network efficiency and energy use, bottom-up methods can be used, which model architectures based on available equipment or equipment models [9]. These methods can be used to study architectural trade-off spaces and a wide range of network optimization and techno-economic analyses. One such bottom-up approach is the transactionbased modeling technique, which models data traffic flow through a network based on mean quantities such as the mean number of hops and typical equipment configurations. Similar to lifecycle analysis methods, transaction-based network models follow the data path of a service and give an estimate of the accumulated energy impact. This type of modeling has been employed to evaluate the energy consumed by a particular service [10] and to observe the effect of variations in the traffic volume of the network [9]. Transaction-based models have also been used to calculate optical IP network power [7] and different access networks in wired and wireless systems [11]. Such methods have been used to study wireless access networks, which are estimated to be the most energy-consuming segment of the overall global network infrastructure. In order to continue to increase capacity, wireless networks must rely on using more bandwidth (spectrum) and spatial diversity by reducing the cell size or distance to the access point. Both approaches require more energy. The introduction of centralized radio access network (C-RAN) architectures is in part done to reduce the energy requirements of access point densification by removing the baseband processing to more efficient, shared processing locations in data centers. It also allows for greater use of virtualized network functions. Numerous studies of C-RAN architectures have considered the optimization of both wireless and compute resources[12]. With commercial equipment becoming available for O-RAN deployments, energy optimization of the wireless network through C-RAN and virtualization has become a key use case and stimulated renewed attention. It also motivates the study of such architectures using transaction-based methods and revisiting previous studies with updated equipment and new architectures. Recently [5], this approach was used to study the power consumption in O-RAN architectures for different network topologies and nodal fanouts. Using these new transaction-based models for 5G architectures, including the full front- and back-haul networks, it is possible to analyze their evolution and connection to previous architectures and past predictions of energy use. Here we perform such an analysis by first considering both the processing energy at different nodal units and the transmission energy associated with high-capacity evolved common public radio interface (eCPRI) data [13]. The eCPRI protocol used in 5G carries digitized radio signals along with control, management, and synchronization information. We update previous models [11] by using the most recent generation of network equipment, and then perform a comparative evaluation between conventional access network designs and the new ORAN-based architectures for different user access rates. From an access technology perspective, the energy benefits of the new centralized architectures are readily apparent, showing

3

Fig. 2. Overview of different RAN architectures.

efficiency similar to passive optical networks at access rates of 100 Mbps. Furthermore, we examine the effect of placing baseband processing at different centralization points in the RAN architecture and evaluate performance under varying levels of network densification for a fixed number of users, providing a framework for understanding the impact of densification and its trade-offs with respect to energy use. The results highlight how the placement of baseband processing and the degree of network densification influence overall energy efficiency. The results show that in these new small cell architectures, the processing energy, rather than the radio unit energy, dominates total consumption. The transactionbased models capture the corresponding architectural tradeoffs and show that centralized architectures exhibit the potential to achieve significant savings when resources are efficiently shared. II. RAN A RCHITECTURES A. Evolution of RANs Radio Access Networks have evolved significantly across successive generations of wireless communication systems to meet growing connectivity demands. Early implementations were characterised by closed, hardware-dependent architectures, while current designs emphasize openness, flexibility, and energy efficiency. Fig. 2 depicts the different types of RANs. In the first generation (1G) and second generation (2G) of communication networks, the radio equipment, including antennas and RF amplifiers were connected by copper cables to custom radio signal processing hardware that performed processing to baseband digital signals. These signals were then aggregated through a backhaul network to various service gateway systems in the core that admitted traffic onto the internet. This macro-cell base station configuration with colocated RF and radio processing equipment is often referred to as distributed RAN (D-RAN) today. This architecture was efficient for large macrocell deployments aggregating radio traffic over 10-20 km distances. The third generation (3G) exhibited improved energy efficiency due to the use

Fig. 3. Schematic of the access network architecture for various technological platforms.

of digital wideband coding, faster power control, and link adaptation to transmit only the power needed for a given channel condition. Like the previous generations, 3G also employed a D-RAN architecture where each cell site’s radio (NodeB) was co-located with its own baseband processing, so every site hosted a complete radio and baseband unit. This provided low latency between the radio and its processing unit, but relied on dedicated processing hardware placed in an increasing number of remote locations as the density of cell sites increased. The introduction of fourth generation (4G) long-term evolution (LTE) brought a fully packet-switched architecture with evolved NodeB (eNodeB) units delivering robust broadband connectivity. The energy efficiency improved over 3G because LTE removes the external radio network controller (RNC), along with lean reference signals instead of always-on pilots, and widespread multiple-input multipleoutput (MIMO)/beamforming to deliver more bits for the same power. In addition, discontinuous reception (DRX) and more efficient radio-frequency (RF)/baseband hardware reduced both user equipment (UE) and site power, especially under variable load. LTE benefited from a transition to remote radio units for which the radio equipment was connected to the baseband processing unit through an optical fiber connection (from the top of the cell tower to the base), saving significant RF transmission power. Despite LTE’s energy gains, architectural constraints at the cell site remain the bottleneck for scaling and cost due to D-RAN limitations, as each eNodeB operates independently and relies on proprietary systems. To overcome these issues, the centralised RAN (C-RAN) architecture was introduced, which decoupled radio units (RUs) from BBUs using the long reach of the fiber connection to distribute RUs at remote cell sites while pooling BBUs in a centralised location [14]. This arrangement is designed to allow multiple RUs to share processing resources, improving efficiency and reducing operational costs. Centralised BBUs also enable more flexible scaling, since additional RUs can be deployed without replicating processing units at each site. An optical fronthaul network, typically based on the CPRI proto-

4

col over fibre, connects the distributed RUs with the centralised BBU pool. Although C-RAN might improve scalability and reduce energy consumption, it introduces new challenges, including a single point of failure at the BBU pool and the need for high-capacity fronthaul connections, which can increase deployment costs. In addition, reliance on proprietary interfaces limits multi-vendor interoperability, maintaining the vendor lock-in. Virtualised RAN (vRAN) and Cloud-RAN ease vendor lock-in and let the same hardware be shared across RAN variants, so deployments can be more flexible, easier to scale, and quicker to adapt. Operationally, vRAN emerged in late 4G/LTE deployments and has become more prevalent with the fifth generation (5G), where wider bandwidths and large-scale MIMO make pooling and automation particularly valuable. Cloud-RAN is largely a 5G-era development, as operators relocate functions of distributed units (DUs) and centralised units (CUs) to edge data centers or centralized clouds to improve utilization and orchestration. Consistent with 5G targets, energy efficiency improves significantly over 4G, driven by lean carriers, deeper sleep states, and massiveMIMO/beamforming, though the exact gain depends on band, load, and feature usage. vRAN virtualizes DU/CU processing on COTS platforms to improve utilization and agility, but it does not, by itself, guarantee openness across vendors nor improved efficiency given that GPUs may not be more efficient than dedicated radio processing ASICs. Building on virtualization, the O-RAN Alliance was formed in 2018 to define a fully open and disaggregated framework for 5G RAN. Unlike standardised architectures focused solely on functionality, O-RAN emphasises openness and modularity by disaggregating hardware and software and enabling network functions to run on commercial off-the-shelf hardware. This reduces dependence on proprietary systems while enhancing flexibility. The O-RAN architecture is composed of open radio units (O-RUs), open distributed units (O-DUs), and open central units (O-CUs) [14], as depicted in Fig. 4. ORAN makes energy a controllable resource by disaggregating O-RU/O-DU/O-CU and exposing policy controls through near-/non-real time (RT) RAN intelligent controllers (RIC) for cell muting, carrier thinning, MIMO layer capping, and deep sleep orchestration. By allowing each element to be scaled and power-managed independently, O-RAN is meant to translate 5G’s theoretical energy efficiency gains into systemlevel savings, making O-RANs adaptable to varied deployment environments. The O-RAN architecture also incorporates the RAN intelligent controller (RIC), which is meant to support advanced controls, including machine learning algorithms to adjust the operation of RAN nodes, for example, based on traffic patterns [14]. The RIC comprises two subcomponents: the near-realtime (RT) RIC and the non-RT RIC, as illustrated in Fig. 2. To execute its near real-time control functions, the near-RT RIC uses modular applications known as xApps. These vendordeveloped software modules operate within the RIC platform and perform specific control tasks such as load balancing and handover optimisation. xApps acquire telemetry data from the network, apply policies and machine learning algorithms, and issue control commands back to the RAN nodes. The non-RT

RIC employs equivalent applications known as rApps, which focus on long-term optimisation tasks such as training AI models and generating network policies[14]. This modularity allows operators to embed intelligence directly into the network with the potential to adapt performance in response to real-world conditions. These software control features require compute resources that need to be accounted for in terms of energy use. The actual control signaling and processing is relatively infrequent and low capacity, so in principle, the energy cost might be minimal. In practice, the provision of computer resources and open stacks can be substantial, and early operational implementations may be inefficient. Deployments are still limited in scale and have neither benefited from multiple rounds of evolution nor economies of scale. An accounting of their energy use at this stage is important in order to assess the current state and understand what further improvements are needed. Mobile networks using a C-RAN architecture can be organized into fronthaul, midhaul, or backhaul networks that extend across the edge, metro, regional, and longhaul hierarchical areas. Fronthaul connections are typically point-topoint links between the O-RUs and the nearest O-DU/O-CU aggregation and processing nodes. If the radio signals are processed down to baseband, then the connections become conventional backhaul links. The term midhaul applies to the transport of signals that have been partially processed and are sent on to another node for full processing, typically between an O-DU and O-CU. The combination of front/mid- and back-haul networks is often referred to as the xhaul network. Signals are increasingly multiplexed through dense wavelength-division multiplexing (DWDM) as they are aggregated deeper into the network. The DWDM links also interconnect core nodes through gateway routers, enabling routing functions across the entire network. To allow for the flexibility and virtualization of radio processing at different locations in the network, 3GPP introduced functional splits, which divide the radio processing into its various steps separated by standardized split points. Processing a signal to a given split point requires processing resources and associated power consumption, but reduces the transport capacity requirement to carry the digitized radio signal, which typically includes far more information than the actual baseband user data. When C-RAN architectures were first introduced, the tradeoff between inefficient processing closer to the edge versus higher capacity transport of digitalized signals was identified and well studied [12]. [12] 3GPP specifies eight split options for the distribution of the 5G gnodeB (gNb) across the network [15]. In the simplest case, the DU and CU, which comprise the gNb are located at the RU location. At the other extreme, using Split 8, the analog radio waveform is transported (typically using a digitized and packetized analog signal format such as eCPRI) from the RU to a remote gNb. Split 1 to Split 7 vary in the level of network function distribution between RU and DU/CU. In Split 6, the functional boundary is placed at the MAC/PHY interface, retaining the entire PHY in the RU and lowering fronthaul demand by transmitting MAC–PHY transport blocks instead of raw I/Q radio data. Split 7 supports important physical layer

5

Fig. 4. O-RAN architecture showing baseband processing (BBP) at O-CU.

centralization benefits, such as distributed MIMO processing, and it is further divided into 7.1, 7.2, and 7.3. In O-RAN, Split 7.2 is recommended between the RU and DU. Many combinations of processing splits can be used among the different x-haul network connections, allowing for flexible tradeoffs between processing, transport, and functionality that need to be captured in order to understand the consequences on network energy use. III. E NERGY U SE M ODELS A. Access Networks Various power consumption models exist in the literature for different network architectures as discussed in Section I. In order to explore the evolution of network energy use over the past decade, we start with the transaction-based models introduced in 2011[10], where the authors analyze both wired and wireless access network technologies. The energy consumption per user is given as [10],   2 PTU PRN PCPE EU = + + (1) RU NTU NRN 2 where RU is the user access rate, NTU and NRN are the number of users at the terminal unit and remote node, respectively, and PCPE , PRN and PTU are the power of the customer premises equipment (CPE), the remote node and the terminal unit in the central office, respectively. In the original work, this model was generalized to include passive optical network (PON), point-to-point (PtP) network, Fiber to the node (FTTN), and worldwide interoperability for microwave access (WiMAX). Fig. 3 shows the schematic illustration of these access technologies, dividing them into three sections: the terminal unit (located in the local exchange office), the remote node, the base station, and the CPE. Since the mid-1990s, PON technology has been used to provide fiber to the home, and today it is widely deployed for residential access. Every household in the PON access network

has an optical network unit (ONU) as the customer premises equipment (CPE). ONUs use fiber links to connect to the optical line terminal (OLT) in the central office through power splitters that enable a shared access medium. In contrast, PtP networks provide a faster access rate via dedicated fiber links to the end user or access point. Radio and enterprise networks use PtP before multiplexing into the DWDM networks. For the conversion of an optical signal into an electrical signal, an optical media converter (OMC) is used as the CPE, and an Ethernet switch is used at the network edge node. FTTN is a hybrid network that utilizes copper wires already installed by the network operator. In FTTN, fiber runs from the central office’s Ethernet switch to the mini-exchange near the customer’s home, serving as a remote node, which is then connected to the customer’s premises using copper wire. An ADSL modem is used as the CPE, and a multi-service access node (MSAN) is used as a hub serving 32 customers. WiMAX was used in previous studies, but today it is largely considered a legacy technology. In Fig. 5, we reproduce the energy consumption analysis of PON, FTTN, PtP, and WiMAX based on the network equipment specifications previously reported [10]. We also obtain the energy consumption of these access network technologies using the latest network equipment, including the Cisco Catalyst 1300 [5] as the Ethernet switch. Additionally, we use this model to obtain the energy consumption in ORAN (Split 8) and LTE networks, which were not included previously [10]. Note that this analysis only considers the access portion of the network [10] and does not incorporate the energy consumed in the transmission of data over the larger network connecting to data centers. Furthermore, the additional energy used for thermal management, protection resources, or mobility coverage is not included here [10]. Thus, this analysis provides a basic, coarse comparison of the energy use of different access technologies without accounting for the complexities of their deployment requirements and operation.

6

Fig. 5. Energy consumption of various access network technologies. The curves for PON, FTTN, PtP, and WiMAX are generated using the equipment in [10] (dotted lines) and also the latest network equipment (bold lines). LTE and O-RAN are not studied in [10].

B. Energy Consumption Model Given that recent radio architectures rely on network scale efficiency enhancements, the coarse model shown in Fig. 5 needs to be enhanced to include the overall network between the access point and the data center. Network-scale transactionbased models were developed previously. Here, we describe their application to recent reports of 5G C-RAN uplink scenarios, including use in O-RAN. These models consider the mean energy use of deployed equipment over a sufficiently long period of time that a stable average can be obtained (on the order of weeks). The mean energy per bit of each equipment type used to process a given internet transaction (e.g., mobile application execution) throughout the network is modified by deployment parameters. Here, an ‘over-provisioning’ factor is used to account for provisioning above that needed to handle the mean traffic, for example, to be able to accommodate diurnal and other variable traffic patterns or to support survivability functions [9]. An ‘overhead’ modifier is used to account for cooling, distribution losses, and other network equipment inefficiencies [10]. A coverage factor is used to account for radio network deployments needed to cover a wide area with a desired density with respect to the average number of users. We look at the end-to-end architecture and identify the different components that contribute to the overall energy consumption in the network. Energy is expended during wireless transmission of data from the user equipment to the RUs. The radio equipment located at the RUs also requires energy for operation. Additionally, data traffic consumes energy when it travels from one nodal unit to another. This energy consumption depends on the amount of data traffic, the type of operations performed on the data, and the network equipment it traverses. Today’s RAN extends from the RU to the DU and CU located in nearby data centers or processing sites, and application data might continue to a regional data center where the user application processing takes place. For a centralized RAN, this network includes fronthaul, midhaul, and backhaul transmission links. Moreover, at each nodal unit in the network, energy is consumed if processing operations

are performed on the user traffic. Therefore, the total energy consumption includes the radio energy consumed for wireless transmission and in the RU equipment, the processing energy at the network nodes where BBP is performed, and the transport energy consumed when data traffic travels across the network. 1) Radio Energy: Before entering the RAN, the wireless transfer of data from the user device to the radio unit consumes some energy Ew , which is the energy expended by the user equipment in transferring one bit of data over the wireless channel to the radio unit. In addition, the equipment (power amplifiers, antennas etc.) in RUs also consume energy expressed as Ee = Nr Pr /CU , where Nr is the number of RUs, Pr is the power consumption of the radio equipment at each RU and CU is the user traffic. So, the radio energy consumption is obtained as Era = Ew + Ee . 2) Processing Energy: This is the energy consumption of the processing server and is present only if BBP is being performed at the node. Let U denote the set of nodal units in a RAN, i.e., RU, DU, CU and DC. The energy consumed per bit for BBP at unit u ∈ U is expressed as EB,u = αu σu ρu γu NC

PC CC

(2)

where αu , σu and ρu are the overprovisioning factor, overhead factor and coverage factor, respectively, of the unit u. Here, γu = CN /CU , such that CN is the network eCPRI traffic passing through this unit. Moreover, NC is the number of server cores and PC and CC are the power consumption and processing capacity, respectively, of each server core. Note that this processing capacity is specific to the processing task, which might be the full BBP or else processing to a given split. The combination PC and CC can be obtained by measuring the processor power used to handle a given capacity of radio data. The total processing energy per bit across all nodal units P can be obtained as Epr = u∈U EB,u . 3) Transport Energy: The transport energy consumed in a RAN architecture includes two components: the equipment energy at the non-processing nodes and the transmission energy across the DWDM links. The first component models the energy consumption of the node transport equipment and is governed by the location of the node with respect to the BBP node location. The equipment implemented at unit u consumes energy per bit given as Eeq,u = αu σu γu

PI CI

(3)

where PI and CI are the power consumption and capacity, respectively, of the network interface card (NIC) or line system chassis present in the equipment. Here, if BBP has not been performed at a unit, then γu = ρu CN /CU , else for all other cases, γu = 1. This accounts for the fact that the fronthaul network is a dedicated network that operates at full capacity, independent of the user traffic, and is deployed based on coverage requirements. After baseband processing, the user data is carried in a shared packet network, and therefore, the power is directly proportional to the energy required to handle the corresponding user data. Note that, for the non-processing

7

nodes, EB,u = 0 as the signal passes transparently through these nodes without processing. The second component models the energy consumed when data traffic travels along the fronthaul, midhaul, and backhaul transmission links. The energy consumed per bit during transmission in the network backhaul is given as EBH = αB σB γB (ξB,s + ξB,l + ξB,r ) such that ξB,s = (HB,s + 1)PS /CS , ξB,l = HB,l PL /CL and ξB,r = (HB,r + 1)PR /CR , where αB and σB are the overprovisioning factor and overhead factor, respectively, of the backhaul network. Here, HB,s , HB,l and HB,r are the number of hops across switches, WDM links, and routers, respectively, in the backhaul. Moreover, PS , PL and PR are the power consumption in the switches, links, and routers, respectively, and CS , CL and CR are the corresponding capacities. Furthermore, γB = ρB CN /CU if BBP is being performed at DC, and γB = 1 otherwise, where ρB is the coverage factor for backhaul. Similarly, energy consumption during midhaul transmission is EMH = αM σM γM (ξM,s + ξM,l ), where αM and σM are the overprovisioning factor and overhead factor, respectively, of the midhaul network, and γM = ρM CN /CU if BBP is being performed at CU or at DC, and γM = 1 otherwise, such that ρM is the coverage factor for midhaul. Here, ξM,s = (HM,s + 1)PS /CS and ξM,l = HM,l PL /CL , where HM,s and HM,l are the number of hops across switches and WDM links, respectively, in the midhaul. Finally, transmission in the fronthaul consumes energy expressed as EFH = αF σF γF (ξF,s + ξF,l ), where αF and σF are the overprovisioning factor and overhead factor, respectively, of the fronthaul network, and γF = 1 if BBP is being performed at RU, and γF = ρF CN /CU otherwise, such that ρF is the coverage factor for fronthaul. Here, ξF,s = (HF,s + 1)PS /CS and ξF,l = HF,l PL /CL , where HF,s and HF,l are the number of hops across switches and WDM links, respectively, in the fronthaul. Therefore, the total transmissionPenergy consumed per bit is Etr = EFH +EMH +EBH + u∈U Eeq,u . Based on the above model, the total energy consumption per user is then calculated as ET = Era + Epr + Etr . The different components of energy consumption are also indicated in Fig. 4.

TABLE I N ETWORK E QUIPMENT [5] Device Router Core Switch Access Switch Fiber Link Radio

Model Cisco 8000 Cisco 9600 Cisco Catalyst 1300 1FINITY T600 Benetel 650

Rated Power (W) 172 3000 86.7 4265 110

Capacity (Gbps) 3200 25600 480 9600 11

Scenario 1: In this scenario, BBP is performed at the RU under a D-RAN option, where the full protocol stack is executed locally at the radio site. This represents a largely macrocell configuration with the gNB at the cell site. • Scenario 2: This is a typical O-RAN implementation where the BBP is performed at the O-DU. In this configuration, the O-RU and O-DU are co-located or within a short PtP connection of each other. Since baseband functions (RLC/MAC/PHY) reside in the O-DU, the functional separation with respect to the CU follows the standard Split 7.2 option (at RU-DU). • Scenario 3: Here, BBP is centralized at the CU hosted in an edge/regional cloud, with the RU acting as a lightweight RF front end and the remaining protocol stack executed at the CU. This Split 8-like edge deployment represents an emergent RAN architecture that pools processing closer to the RAN than a distant data center, retaining centralization gains while reducing transport distance and helping satisfy latency constraints. • Scenario 4: In this Split 8 configuration, BBP is fully centralized in the regional data center. The RU remains a lightweight RF front end, while the PHY and all higher protocol layers are executed centrally. This centralization simplifies the radio site and concentrates computing, scheduling, and optimization functions in the large-scale data center. This scenario is a C-RAN architecture, with the DU and CU (gNb) located in the data center and all other network nodes forwarding the radio signals through the network. •

B. Network Densification IV. RAN S CENARIOS Details of the network equipment used in this analysis are given in Table I. We assume that the average monthly data consumption of a user is 10 Gigabytes. Typical of RAN deployments, the servers at distributed radio locations in the network have 4 cores with a total BBP capacity of 1 Gbps each, whereas the servers in a DC have 20 cores, which gives a cumulative BBP capacity of 5 Gbps of fronthaul data per server [5]. The thermal design power per core for the servers at RU, DU and CU is 6 W, and for those at the DC, it is 5.5 W [5]. The energy expended in the UE is Ew = 25 nJ/bit [5]. These design parameters are fixed unless mentioned otherwise. A. Deployment Scenarios We analyse four different RAN configuration scenarios based on the location of the BBP in the RAN architecture.

To study network densification, we consider an average population of 100 users or CPEs supported by a RAN and increase the number of RUs serving these users from 1 to 100. As the number of RUs increases, we provision an additional DU for every fourth RU. This fan-out is commonly supported in open deployments. Larger fan-outs might be supported, but would correspond to a greater oversubscription of processing. With 4 RUs per DU, the DU is oversubscribed by 4 times. Consequently, the number of DUs increases from 1 to 25. Moreover, we assume there are two CUs in the network connected to a single DC in the network core. The variation in energy consumption (per user) with an increase in network densification for the four different deployment scenarios is depicted in Fig. 6. As the number of RUs is increased, the consumption of processing energy scales monotonously for all deployment scenarios except Scenario-2, as shown in Fig. 6(a). In the

8

Fig. 6. Variation in (a) processing, (b) radio & transport, and (c) total energy consumption with increase in network densification for different deployment scenarios. In (b), the transport energy for Scenario-1 (BBP at RU) is not plotted as it is fairly constant at a low value (∼ 40 nJ/bit).

case of Scenario-2, when there is a single RU in the network, the average energy consumption per bit is high, but it reduces as more RUs are added to reach the maximum sharing capability of the DU (which is 4 RUs per DU in our analysis). After every four RUs, when a new DU is added into the network, we observe a jump in the energy consumption. This periodic addition of DUs results in a step-like shape of the plot for Scenario-2. The frequency of the steps depends on the resource-sharing capability of the DU. Scenario-1 and Scenario-4 have the highest and lowest consumption of processing energy, respectively, because Scenario-4 benefits from multiplexing gain due to the sharing of resources at the core. However, there is a cross-over between the processing energy consumption of Scenario-2 and Scenario-3 at NRU ≈ 30. This is because the number of DUs keeps increasing with densification, whereas the number of CUs is fixed at 2. In Fig. 6(b), we depict the transmission energy consumption for the different deployment scenarios. In Scenario-1, the eCPRI data is processed at the RUs, so the transmission energy consumed in this case is very low and does not vary with densification. In all other scenarios, the transmission energy consumption is higher because the eCPRI data travels further into the network (across the fronthaul in Scenario-2, across the fronthaul and midhaul in Scenario-3, and up till the network core in Scenario-4). Moreover, in these three cases, the transmission energy consumption scales with densification. The eCPRI data hops through DUs in Scenario-3 and 4, which leads to a step-like shape of the curves for these two cases, due to the same reason as explained before for Fig. 6(a). Finally, the total energy consumption per user is shown in Fig. 6(c), which is essentially the sum of the energy consumption depicted in Figs. 6(a) and (b). We deduce that the D-RAN deployment in Scenario-1 is the least energy efficient. The other three deployment scenarios benefit from multiplexing gains and sharing benefits, leading to lower energy consumption. The energy consumed in these three cases is almost the same at high network densification, which is ∼75% less than the energy consumed in Scenario-1 when one RU is provisioned for each user. Across all scenarios, the processing energy dominates be-

cause when BBP is centralized (e.g., at the CU or a regional DC), the PHY/MAC operations get concentrated onto fewer, shared compute platforms, which raises the per-node processing load even as the fronthaul traffic per RU drops. That is, centralization shifts the energy burden upstream: the DU/CU infrastructure needs to handle aggregated waveform reconstruction, scheduling, and control-plane work for many RUs, so the compute contribution grows faster than the incremental savings in transmission energy. When BBP is distributed closer to the RU (e.g., Split 6 or 7.x), the processing load is spread among many lower-power nodes, which reduces each node’s compute energy but increases the number of nodes that must be powered. These counterbalancing trends explain why the total energy curve is most sensitive to the BBP location, and they reinforce the conclusion that intelligent placement, not merely densification, is what determines whether centralization actually reduces the network’s energy footprint. V. C ONCLUSION In this work, we explore the energy consumption of emergent RAN architectures. In particular, we update past results using today’s equipment and compare access technology energy efficiency. We also look at the historical efficiency of processor-dominant network equipment (edge routers) and find an efficiency improvement rate that is in line with predictions (and nevertheless slower than traffic growth rates). To consider today’s open and virtualized RANs, we need to include the full network energy from the access point to the data center. By including this end-to-end aspect, we show that the processing energy is taking up a greater portion of the overall energy in RANs, as the energy in the radio units is greatly reduced. Depending on the deployment strategy, greater sharing of the processing load is shown to yield significant energy efficiency improvements. These findings highlight that the placement of baseband processing within the network plays a crucial role in determining overall energy efficiency. While centralized architectures like C-RAN consume higher transmission energy due to longer data transport, their superior computational efficiency can offset this cost. Ultimately, designing future communication networks will require a careful balance be-

9

tween processing distribution and transmission distance to achieve sustainable energy performance. ACKNOWLEDGMENTS This work was supported in part by Research Ireland grants 18/CRT/6222 and 13/RC/2077/P2; and by the EU projects 6G-XCEL, ECO-eNET and ENGCoN. The projects 6G-XCEL and ECO-eNET have received funding from the SNS JU under grant nos. 101139194 and 101139133, respectively, and ENGCoN has received funding from the MSCA-PF scheme under grant no. 101155602.

R EFERENCES [1] Q. Qi, X. Chen, C. Zhong, and Z. Zhang, “Integration of energy, computation and communication in 6G cellular Internet of things,” IEEE Communications Letters, vol. 24, no. 6, pp. 1333–1337, 2020. [2] A. Badshah, A. Daud, H. U. Khan, O. Alghushairy, and A. Bukhari, “Optimizing the over and underutilization of network resources during peak and off-peak hours,” IEEE Access, vol. 12, pp. 82549–82559, 2024. [3] F. Sun, P. Wang, J. Zhao, N. Xu, J. Zeng, J. Tao, K. Song, C. Deng, J. C. Lui, and X. Guan, “Mobile data traffic prediction by exploiting time-evolving user mobility patterns,” IEEE Transactions on Mobile Computing, vol. 21, no. 12, pp. 4456–4470, 2022. [4] D. C. Kilper and H. Rastegarfar, “Energy challenges in optical access and aggregation networks,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 374, no. 2062, p. 20140435, 2016. [5] U. Tariq, R. Raj, and D. Kilper, “Design and analysis of power consumption models for Open-RAN architectures,” in 2025 IEEE International Conference on Communications Workshops (ICC Workshops), pp. 1037– 1042, 2025. [6] W. Vereecken et al., “The environmental footprint of data centers: The influence of server renewal rates on the overall footprint,” in Proceedings of the International Conference on Green Communications and Networks (GCN 2011), pp. 823–831, Springer, 2012. [7] J. Baliga, R. Ayre, K. Hinton, W. V. Sorin, and R. S. Tucker, “Energy consumption in optical IP networks,” Journal of Lightwave Technology, vol. 27, no. 13, pp. 2391–2403, 2009. [8] G. Raffin and D. Trystram, “Dissecting the software-based measurement of CPU energy consumption: A comparative analysis,” IEEE Transactions on Parallel and Distributed Systems, vol. 36, no. 1, pp. 96–107, 2025. [9] D. Kilper et al., “Power trends in communication networks,” IEEE Journal of selected topics in quantum electronics, vol. 17, no. 2, pp. 275– 284, 2010. [10] J. Baliga, R. Ayre, K. Hinton, and R. S. Tucker, “Energy consumption in wired and wireless access networks,” IEEE Communications Magazine, vol. 49, no. 6, pp. 70–77, 2011. [11] J. Baliga et al., “Energy consumption in access networks,” in Optical Fiber Communication Conference, p. OThT6, Optica Publishing Group, 2008. [12] M. Fiorani, S. Tombaz, J. Mårtensson, B. Skubic, L. Wosinska, and P. Monti, “Energy performance of c-ran with 5g-nx radio networks and optical transport,” in 2016 IEEE International Conference on Communications (ICC), pp. 1–6, IEEE, 2016. [13] T. Pfeiffer, “Next generation mobile fronthaul and midhaul architectures,” Journal of Optical Communications and Networking, vol. 7, no. 11, pp. B38–B45, 2015. [14] B. Agarwal, R. Irmer, D. Lister, and G.-M. Muntean, “Open RAN for 6G networks: Architecture, use cases and open issues,” IEEE Communications Surveys & Tutorials, 2025. [15] L. M. P. Larsen, A. Checko, and H. L. Christiansen, “A survey of the functional splits proposed for 5G mobile crosshaul networks,” IEEE Communications Surveys & Tutorials, vol. 21, no. 1, pp. 146–172, 2019.