Pith. sign in

REVIEW 3 major objections 5 minor 23 references

Design and Analysis of Power Consumption Models for Open-RAN Architectures

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Performing baseband processing at the data center instead of at the radio unit cuts per-user power by about 80 percent in O-RAN, according to a transaction-based model.

desk verdict Load-bearing typo in Eq. (1) flips the central result; the O-RAN modeling framework is useful but the quantitative claims need recomputation. read the letter →

arxiv 2505.24552 v1 pith:LEMPMO2J submitted 2025-05-30 cs.NI

Urooj Tariq , Rishu Raj , Dan Kilper This is my paper
classification cs.NI
keywords powerconsumptionopenRANfuturenetworksbasebandprocessingfanoutgreencommunicationtransaction-basedmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks where in a centralized open radio access network (O-RAN) baseband processing—turning digitized radio signals into user data—should be located to minimize power per user. The authors build a transaction-based model of the full path from radio unit to data center, separating processing power from transmission power, and compare four candidate locations for baseband processing: the radio unit, the distributed unit, the central unit, and the data center. Their central finding is that, with current commercial equipment and the fanout limits that restrict how many radio units can share a distributed unit, the data center is the most power-efficient location, cutting total power per user by roughly 80 percent compared with doing the same processing at the radio unit when there are 10 users per radio. This matters because O-RAN operators deciding where to centralize baseband currently lack an end-to-end power comparison that includes both compute and transport.

What carries the argument

The central object is a transaction-based power model that provisions equipment according to a target user traffic and radio coverage. Processing power $P_{pr}$ is a three-branch formula: before baseband processing it charges only node transport scaled by overprovisioning $\alpha_n$, overhead $\sigma_n$, and coverage $\rho_n$; at the baseband node it adds the server core term $M_c P_c / C_c$; after baseband it charges transport on the much smaller baseband traffic $C_u$. Transmission power $P_{tr}$ counts switch, WDM-link, and router hops with a factor $\varepsilon_l$ that equals the large eCPRI coverage traffic before baseband processing and the small user traffic after it. These two formulas, together with the equipment parameter tables, carry the entire comparison across the four baseband locations.

What would settle it

Re-run the model with the edge overprovisioning factor reduced from 5 to 1.3 (matching the data center) or with edge servers upgraded to 20 cores; if data-center processing no longer yields the lowest power per user, the central claim fails. Alternatively, measure commercial O-RAN deployments that process at the O-DU and at the data center under identical traffic and compare per-user power; the predicted ~80 percent advantage for the data center should be observable.

Watch

Extended reading notes

Core claim

The paper claims that for O-RAN deployments using the 7.2 functional split, eCPRI fronthaul, and 10 users per O-RU, total power per user is minimized when full baseband processing is done at the data center. Processing power at the data center is low because its servers are provisioned with 20 cores and 5 Gbps of baseband capacity per server, versus 4 cores and 1 Gbps at the edge, and because the data center carries lower overprovisioning and overhead factors (1.3 and 1.5) than the edge nodes (5 and 2). Although sending the large eCPRI signal to the data center increases transmission power, processing power dominates the total, so the net effect is an approximately 80 percent reduction in power per user compared with baseband processing at the O-RU. The paper also shows that raising O-DU and O-CU fanout lowers processing power at those nodes through better utilization, and that transmission power is always highest when processing is at the data center because the unmultiplexed fronthaul traffic travels farthest.

Load-bearing premise

The conclusion assumes that data-center servers are used much more efficiently than edge servers: they are provisioned with little spare capacity and overhead, and each server does five times the baseband processing, numbers taken from one vendor's equipment and never varied in the paper.

Editorial extensions

If this is right

  • O-RAN operators should place full baseband processing in data centers to minimize per-user energy, given current server efficiency ratios and fanout limits.
  • Because processing power dominates, lowering the number of server cores needed per O-RU connection (for example through more efficient baseband software) directly reduces total network power.
  • As fanout increases (more O-RUs per O-DU), fronthaul transmission capacity requirements grow quickly, eventually making transmission power the limiting factor again.
  • Performing baseband at intermediate O-DU or O-CU nodes keeps transmission energy low while still gaining from efficient processing at high utilization, a middle ground the model identifies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If edge servers ever match data-center server efficiency (same overprovisioning and overhead factors), the model's ranking could flip; the paper does not test this sensitivity, so the 80 percent figure is conditional on the assumed vendor configuration.
  • The same model structure can be extended to compare other functional splits or to add latency constraints, since the paper's 7.2-split assumption is a choice, not a necessity.
  • Energy-aware O-RAN controllers could treat baseband processing location as a live optimization variable: the model implies that moving processing toward the data center saves more energy than any radio-side sleep mode currently considered.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper develops transaction-based power consumption models for a centralized O-RAN architecture, splitting total power into processing and transmission components. It compares performing baseband processing (BBP) at the O-RU, O-DU, O-CU, and data center as a function of the number of O-RUs and of different nodal fanout configurations. The central quantitative conclusion is that, with 10 users per O-RU, performing BBP at the DC reduces total power per user by about 80% compared with BBP at the O-RU, because efficient shared data-center servers are said to dominate the additional transport cost of fronthaul eCPRI traffic.

Significance. The paper addresses a timely and practically important question: where to place O-RAN baseband processing to minimize end-to-end energy consumption. Its strengths are the end-to-end scope (covering the UE, fronthaul, midhaul, backhaul, and data center), the use of commercial equipment datasheets, and the explicit treatment of eCPRI versus baseband traffic in the transmission model. The fanout analysis in Section III.B is a useful way to connect centralization, sharing, and power consumption. However, the central result currently rests on a definitional inconsistency in the processing-power equation and on parameter values that are not subjected to sensitivity analysis; until these issues are fixed, the quantitative claims cannot be regarded as established. If corrected, the framework could be a useful reference for O-RAN energy studies.

major comments (3)
  1. [II.B.1, Eq. (1)] In the 'BBP on node' case of Eq. (1), the server power term is written M_c P_c / C_c, with P_c and C_c defined as the power and capacity of each server core. For a server with M_c cores, the correct total server power divided by total server capacity is (M_c P_c)/(M_c C_c) = P_c/C_c, so the term as written is a factor M_c too large. Using the paper's stated configurations (4 cores/1 Gbps at edge nodes; 20 cores/5 Gbps at the DC), a literal evaluation gives 96 W/Gbps for the edge server and 440 W/Gbps for the DC server, which would make DC processing roughly 4.6 times worse than O-RU processing rather than the most efficient option. Since Fig. 2(c) and the ~80% savings in the Conclusion show the opposite, the authors must have used total server power divided by total server capacity. The equation and the definitions of P_c and C_c must be made consistent (for example, define C_c as total server capacity, or replace M_c P_c/C_c by P_c/C_c), and all figures and numerical claims must be recomputed from the corrected equation.
  2. [III.A-III.B, Table I] The ranking of BBP locations is driven by the entries of Table I and by the server configurations taken from a single vendor source [20], yet the paper provides no sensitivity analysis over these values. In particular, the DC advantage in Fig. 2(c) relies on the favorable overprovisioning and overhead factors at the DC (α=1.3, σ=1.5) compared with those at edge nodes (α=5, σ=2), together with the 20-core/5-Gbps server. If the DC were overprovisioned to a degree closer to the edge nodes, or if edge servers were configured with more cores per gigabit, the DC-over-edge ordering could reverse. Because the central conclusion is a general statement that DC processing is most power-efficient, the authors should provide a one-way sensitivity analysis over the parameters in Table I (or at least report the cross-over values) and discuss the provenance of these numbers for O-RAN equipment, rather than relying on parameter values taken directly from prior work [9].
  3. [II.B.2, Eq. (2)] The transmission power model depends on the hop counts H_sh, H_lh, and H_rh, but the paper never specifies how these are obtained for the topologies used in Figs. 2-4 or for the five fanout cases in Table III. Without a definition or a table of hop counts as a function of N_RU, N_DU, N_CU, and N_DC, an independent reader cannot reproduce the transmission power numbers or the processing-versus-transmission tradeoff that underlies the conclusion. Please provide explicit expressions or a table for the hop counts in the evaluated configurations.
minor comments (5)
  1. [III.A] The per-user power in Fig. 2 decreases with N_RU because the O-CU and DC counts are implicitly held at one while N_RU grows; this assumption should be stated explicitly in the text.
  2. [IV] The '~80%' reduction is not tied to a specific configuration or figure; specify the exact case (for example, the largest N_RU in Fig. 2(c)) from which this number is obtained.
  3. [Acknowledgments] The word 'Reasearch' should be 'Research'.
  4. [Table III] The fanout case labels C-1 through C-5 are used in the text before the table is introduced; add a pointer such as '(Table III)' at first use.
  5. [II.B.1] Define C_n explicitly as the total eCPRI rate entering the node from all child O-RUs, not just as 'network eCPRI traffic,' to avoid ambiguity in Eq. (1).

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional, fitted-input, or imported-uniqueness circularity: the DC-more-efficient ranking is an emergent output of the stated parameter set; only a minor self-citation ([9], Table I) supplies load-bearing numerical factors, while the flagged Eq. (1) mismatch is a reproducibility defect, not circularity.

full rationale

The paper's central claim — that performing BBP at the DC is the most power-efficient location, cutting per-user power by roughly 80% versus an O-RU — is not equivalent to its model inputs by construction. It is an emergent evaluation of Eqs. (1)-(2) at the Table I/II parameter set: with the physically intended reading of the server term (total server power / total capacity, about 24 W/Gbps at the edge versus about 22 W/Gbps at the DC), the DC advantage is driven by the lower overprovisioning/overhead factors (alpha_n*sigma_n = 1.95 for DC versus 5 at the O-RU), by coverage-based sharing (rho_n = N_DC/N_u), and by the computed dominance of processing over transmission; the transmission component separately favors the edge, so the ranking is a nontrivial output, and changing alpha_n/sigma_n or the DC server configuration can reverse it. No step of the form 'Eq. X = Eq. Y by construction,' no fitted-parameter-renamed-as-prediction, and no imported uniqueness theorem occurs. The one self-citation with numerical weight is [9] (Kilper et al., including the present corresponding author), which supplies the Table I overprovisioning/overhead/coverage factors; these factors are transparently tabulated and materially affect the magnitude of the DC advantage, but they are inputs from a prior peer-reviewed study, not the target conclusion, and the result also depends on external datasheets ([16]-[23]) and on the fanout structure. This is a mild self-citation/validation concern, not circularity, so the score is 2 rather than 0. Separately, the reviewer-flagged defect in Eq. (1) — the printed term M_c*P_c/C_c yields 96 W/Gbps (edge) and 440 W/Gbps (DC), contradicting Fig. 2 — means the published equations as written cannot reproduce the published figures; that is a reproducibility/correctness defect in the derivation chain, not a circular-input issue, and it is excluded from this score per the rubric, though it should be corrected before the quantitative claims are relied upon.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The model rests on a substantial set of parameter choices from prior literature, vendor datasheets, and one O-RAN software vendor, none of which are validated against measured O-RAN power. The central ranking of BBP locations is an outcome of these choices rather than an independent empirical finding.

free parameters (7)
  • Overprovisioning factor for edge nodes (alpha_n) = 5
    From prior work [9]; applied to O-RU, O-DU, O-CU and fronthaul/midhaul. Not validated for O-RAN.
  • Overprovisioning factor for data center (alpha_n) = 1.3
    From prior work [9]; a key reason DC appears efficient.
  • Overhead factor for O-DU/O-CU (sigma_n) = 2
    From [9]; assumed for O-RAN nodes.
  • Server cores and capacity at edge = 4 cores, 1 Gbps
    From Accelleran dRAX [20].
  • Server cores and capacity at DC = 20 cores, 5 Gbps
    From Accelleran dRAX [20].
  • Fronthaul eCPRI rate per O-RU = 11 Gbps
    From Small Cell Forum [23] for 7.2 split.
  • O-DU fanout limit = 4 O-RUs per O-DU
    Current implementation constraint; used in provisioning rule.
assumptions (4)
  • domain assumption Network equipment operates at full rated power independent of traffic load
    Stated in Section II.B, citing [13]; this makes power per user decrease with more users and favors shared DC servers.
  • domain assumption eCPRI fronthaul traffic under 7.2 split is transported without multiplexing and provisioned for peak load
    Stated in Section II.B; leads to transmission power scaling with peak eCPRI rate.
  • domain assumption The node modeling parameters in Table I, originally from [9], apply to O-RAN equipment
    Used in Eq. (1) and (2); no O-RAN-specific validation.
  • ad hoc to paper The number of O-CUs and DCs is fixed (at 1 each) as the number of O-RUs and O-DUs grows
    In the O-RU scaling analysis, only O-DUs are added every 4 O-RUs; O-CU and DC counts are not scaled, so their coverage factor denominators grow with users.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Design and Analysis of Power Consumption Models for Open-RAN Architectures." pith.science (2026). https://pith.science/paper/LEMPMO2J

@misc{pith2026250524552,
  author       = {Pith},
  title        = {Pith review of: Design and Analysis of Power Consumption Models for Open-RAN Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LEMPMO2J}},
  note         = {Machine review of arXiv:2505.24552}
}
read the original abstract

The open radio access network (O-RAN) Alliance developed an architecture and specifications for open and disaggregated cellular networks including many elements that are being widely adopted and implemented in both commercial and research networks. In this paper, we develop transaction-based power consumption models of a centralized O-RAN architecture based on commercial hardware and considering the full end-to-end data path from the radio unit to the data center. We focus on recent fanout limitations and early baseband processing requirements related to current implementations of O-RAN and assess the power consumption impact when baseband processing is employed at different centralization points in the network. Additionally, we explore how greater fanout and sharing deeper into the network impact the balance of processing and transmission. Low processing fanout restrictions motivate greater centralization of the processing. At the same time, allowing for more open radio units per open distributed unit will quickly increase the transmission capacity requirements and related energy use.

Figures

Figures reproduced from arXiv: 2505.24552 by the authors.

Figure 1
Figure 1. Centralized RAN architecture used in O-RAN showing baseband [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Variation in (a) processing, (b) transmission, and (c) total power consumption per user when the number of O-RUs is increased with baseband processing [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Variation in power consumption per user for five different cases of nodal fanout (refer Table III) with baseband processing (BBP) at (a) O-DU, (b) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Variation in total power consumption per user with baseband processing (BBP) at different nodes when the nodal fanout case (a) C-1, (b) C-2, and (c) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [20]

    Accelleran’s 5G dRAX™ cloud-native Open RAN software

    “Accelleran’s 5G dRAX™ cloud-native Open RAN software.” Ac- cessed: Aug. 04, 2024. [Online]. Available: https://accelleran.com/ accellerans-5g-drax-cloud-native-open-ran-software-now-available/

    work page 2024

  2. [9]

    Power trends in communication networks,

    D. Kilper et al. , “Power trends in communication networks,” IEEE J. Sel. Topics Quantum Electron. , vol. 17, no. 2, pp. 275–284, 2010

    work page 2010

  3. [1]

    ORAN Alliance, 2023

    “ORAN Alliance, 2023.” Accessed: Oct. 16, 2024. [Online]. Available: https://www.o-ran.org/

    work page 2023

  4. [2]

    O-RAN Vertical Industries White Paper December 2023

    “O-RAN Vertical Industries White Paper December 2023.” https://mediastorage.o-ran.org/white-papers/O-RAN.WG1. Vertical-Industry-White-Paper-2023-12.pdf. (Accessed on 11/01/2024)

    work page 2023

  5. [3]

    Energy consumption in wired and wireless access networks,

    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

    work page 2011

  6. [4]

    Energy-efficient baseband unit placement in a fixed/mobile converged wdm aggregation network,

    N. Carapellese et al. , “Energy-efficient baseband unit placement in a fixed/mobile converged wdm aggregation network,” IEEE J. Sel. Areas Commun., vol. 32, no. 8, pp. 1542–1551, 2014

    work page 2014

  7. [5]

    Energy-efficient DU-CU deployment and lightpath provisioning for service-oriented 5G metro access/aggregation networks,

    Y . Xiao, J. Zhang, and Y . Ji, “Energy-efficient DU-CU deployment and lightpath provisioning for service-oriented 5G metro access/aggregation networks,” J. Lightw. Technol., vol. 39, no. 17, pp. 5347–5361, 2021

    work page 2021

  8. [6]

    Determining the right hardware for open RAN deployment,

    Intel Network Builders, “Determining the right hardware for open RAN deployment,” tech. rep., Intel Corporation, 2022. [Online]. Available: https://networkbuilders.intel.com/docs

    work page 2022

Show all 23 references
  1. [7]

    Cloud fog architectures in 6G networks,

    B. A. Yosuf et al. , “Cloud fog architectures in 6G networks,” in 6G Mobile Wireless Networks , pp. 285–326, Springer, 2021

    2021

  2. [8]

    Green cloud computing: Balancing energy in processing, storage, and transport,

    J. Baliga, R. W. Ayre, K. Hinton, and R. S. Tucker, “Green cloud computing: Balancing energy in processing, storage, and transport,” Proceedings of the IEEE , vol. 99, no. 1, pp. 149–167, 2010

    2010

  3. [10]

    Energy consumption in optical IP networks,

    J. Baliga et al. , “Energy consumption in optical IP networks,” Journal of Lightwave Technology , vol. 27, no. 13, pp. 2391–2403, 2009

    2009

  4. [11]

    Energy consumption in access networks,

    J. Baliga et al. , “Energy consumption in access networks,” in Proc. Conference on Optical Fiber Communication , pp. 1–3, 2008

    2008

  5. [12]

    Next generation mobile fronthaul and midhaul architectures,

    T. Pfeiffer, “Next generation mobile fronthaul and midhaul architectures,” J. Opt. Commun. Netw. , vol. 7, no. 11, pp. B38–B45, 2015

    2015

  6. [13]

    Routers vs switches, how much more power do they really consume? A datasheet analysis,

    H. Mellah and B. Sanso, “Routers vs switches, how much more power do they really consume? A datasheet analysis,” in Proc. IEEE International Symposium WoWMoM, pp. 1–6, 2011

    2011

  7. [14]

    HSADR: A new highly secure aggregation and dropout- resilient federated learning scheme for radio access networks with edge computing systems,

    F. Wu et al. , “HSADR: A new highly secure aggregation and dropout- resilient federated learning scheme for radio access networks with edge computing systems,” IEEE Transactions on Green Communications and Networking, vol. 8, no. 3, pp. 1141–1155, 2024

    2024

  8. [15]

    Analyse or transmit: Utilising correlation at the edge with deep reinforcement learning,

    J. Hribar et al. , “Analyse or transmit: Utilising correlation at the edge with deep reinforcement learning,” in Proc. IEEE Global Communica- tions Conference, pp. 1–6, 2021

    2021

  9. [16]

    Cisco 8000 Series Routers - Cisco

    “Cisco 8000 Series Routers - Cisco.” Accessed: Aug. 16, 2024. [Online]. Available: https://www.cisco.com/site/us/en/products/networking/routers/ 8000-series/index.html

    2024

  10. [17]

    Cisco Catalyst Series Switches - Cisco

    “Cisco Catalyst Series Switches - Cisco.” Accessed: Aug. 14, 2024. [Online]. Available: https://www.cisco.com/c/en/us/support/switches/ category.html

    2024

  11. [18]

    1FINITY™ T600 - Fujitsu Network Communications : Fujitsu United States

    “1FINITY™ T600 - Fujitsu Network Communications : Fujitsu United States.” Accessed: Aug. 16, 2024. [Online]. Available: https://www. fujitsu.com/us/products/network/products/1finity-t600/

    2024

  12. [19]

    Benetel-RAN650-Product-Specification-Sheet-v.1.1.pdf

    “Benetel-RAN650-Product-Specification-Sheet-v.1.1.pdf.” Accessed: Aug. 03, 2024. [Online]. Available: https://benetel.com/wp-content/ uploads/2021/11/Benetel-RAN650-Product-Specification-Sheet-v.1-1. pdf

    2024

  13. [21]

    Intel Xeon Silver 4216 Processor 22M Cache 2.10 GHz Product Specifications

    “Intel Xeon Silver 4216 Processor 22M Cache 2.10 GHz Product Specifications.” Accessed: Aug. 01, 2024. [Online]. Available: https://ark.intel.com/content/www/us/en/ark/products/193394/ intel-xeon-silver-4216-processor-22m-cache-2-10-ghz.html

    2024

  14. [22]

    Intel Xeon Gold 6262V Processor 33M Cache 1.90 GHz Product Specifications

    “Intel Xeon Gold 6262V Processor 33M Cache 1.90 GHz Product Specifications.” Accessed: Aug. 01, 2024. [Online]. Available: https://ark.intel.com/content/www/us/en/ark/products/193972/ intel-xeon-gold-6262v-processor-33m-cache-1-90-ghz.html

    2024

  15. [23]

    Small Cell Virtualization: Functional splits and use cases,

    “Small Cell Virtualization: Functional splits and use cases,” in Small Cell F orum Release, vol. 6, p. 55, 2016

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.