Pith. sign in

REVIEW 2 major objections 4 minor 34 references

This paper claims that in O-RAN, end-to-end energy and latency can be jointly modeled and optimized as functions of where baseband processing and AI inference run, with edge placement minimizing latency and centralized placement minimizing

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 15:24 UTC pith:GVZ66OP6

load-bearing objection Solid modeling extension with a load-bearing simplification in the latency stack; worth a referee but the high-load results need rework. the 2 major comments →

arxiv 2608.02082 v1 pith:GVZ66OP6 submitted 2026-08-03 cs.NI

Energy-Latency Trade-offs in O-RAN with Distributed Baseband Processing and AI Inference

Urooj Tariq , Rishu Raj , Shashi Raj Pandey , Merim Dzaferagic , Petar Popovski , Dan Kilper This is my paper
classification cs.NI
keywords O-RANfunctional splitsenergy-latency trade-offbaseband processing placementAI inferenceM/M/1 queueingMILP optimizationfronthaul/midhaul/backhaul
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper develops a unified, throughput-based model of energy and delay for O-RAN deployments, decomposing both into baseband processing, transport, and AI-inference components across fronthaul, midhaul, and backhaul segments. It then formulates a mixed-integer program that selects the best functional configuration—where baseband processing and AI inference are placed—under an energy-per-bit budget, with server frequency and network load as adjustable operating points. Numerical results show the central trade-off: processing at the edge (O-RU) yields the lowest latency but the highest energy per bit, while centralized processing at the data center is most energy-efficient but accumulates transport delay, becoming impractical for latency-critical services under heavy load. The contribution is a systematic framework for choosing placement and operating point in AI-native O-RAN deployments under QoS and energy constraints.

Core claim

The central claim is that the end-to-end model (Eqs. 1–11) together with the optimization (18–23) jointly captures baseband, transport, and AI-inference energy and latency, and that placing baseband processing closer to the network edge significantly reduces latency but incurs higher energy consumption, whereas centralized deployments at the O-CU or data center achieve lower energy per bit at the cost of increased transport latency. The paper further establishes that network load and server frequency are the two operating variables shaping feasibility regions for latency budgets: raising frequency compensates for queuing-induced delay under moderate load, but centralized configurations remai

What carries the argument

The key objects are (i) the four functional configurations {G, F, M, B} that place baseband processing and co-located AI inference at the O-RU, O-DU, O-CU, or data center; (ii) closed-form per-bit energy terms for processing (Eqs. 1–3) and transport across backhaul, midhaul, and fronthaul (Eqs. 5–7); (iii) an M/M/1 queueing model for Ethernet switch delay (Eq. 8), combined with propagation, eCPRI encapsulation, baseband, and AI-inference latency terms; and (iv) a hierarchical binary tuple (χ, η, ω) mapped to a one-hot configuration-selection vector x, turning the placement choice into a small MILP that minimizes end-to-end latency subject to an energy budget. The MILP is solved exactly by en

Load-bearing premise

Every switch along the path is assumed to experience the same network load ρ, so queueing delay at each hop is computed with the same M/M/1 utilization; per-switch arrival rates are never derived from user traffic.

What would settle it

Measure end-to-end latency in a small O-RAN testbed or simulation with realistic per-switch utilization tapering (e.g., fronthaul at 0.9, midhaul at 0.6, backhaul at 0.3) and compare against the paper's single-ρ M/M/1 prediction: if the latency ordering among configurations F, M, and B at high load reverses, the global-load assumption is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If a deployment targets sub-millisecond latency, processing must remain at the O-RU or O-DU; configurations M and B cannot meet tight latency budgets even at the highest server frequency and zero load.
  • Under near-saturation load (ρ=0.99), centralized configurations degrade sharply—configuration B stays above roughly 7 ms even at the best frequency—so latency-critical services must stay at the edge or rely on frequency scaling where possible.
  • Raising the energy budget shifts the optimal configuration from centralized to distributed, creating an explicit energy-versus-latency trade-off curve that operators can use for placement decisions.
  • Server frequency scaling can compensate for moderate queuing delay; for example, configuration F maintains about 2 ms latency as load rises from 0.01 to 0.99 by increasing frequency from 1.2 to 1.6 GHz.
  • The optimization is cheap to solve (four enumerations with low-dimensional continuous subproblems), enabling re-optimization as load and compute conditions change.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the single-global-load assumption were relaxed to per-switch arrival rates derived from user traffic, the configuration ordering at high load could change: realistic load tapering toward the aggregation/core would reduce queueing delay on midhaul and backhaul switches, likely making centralized configurations more competitive than the paper's ρ=0.99 results suggest.
  • The framework treats AI inference as a fixed compute term co-located with baseband; decoupling inference placement (e.g., edge baseband with cloud inference) would introduce a new optimization degree of freedom and could shift the Pareto frontier for AI-native workloads.
  • Replacing the deterministic average-delay abstraction with quantile-based or closed-loop latency constraints—identified in the paper as complementary—could make the feasibility regions stricter for reliability-critical services and change which configurations are selected under low energy budgets.
  • The model's energy gap between edge and central processing depends on assumed processor efficiencies (6 W/core at O-RU/O-DU/O-CU vs. 5.5 W/core at DC); if edge hardware becomes more energy-proportional, the energy penalty of distributed processing would shrink, altering the optimal placement boundary.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper develops an end-to-end energy and latency model for O-RAN with four functional configurations (G, F, M, B) that place baseband processing and co-located AI inference at the O-RU, O-DU, O-CU, or DC. The energy part combines a previously published throughput-based model with a new AI-inference energy term; the latency part adds M/M/1 switch queueing, propagation, baseband processing, and AI inference components. The authors then formulate a mixed-integer program that minimizes end-to-end latency subject to an energy-per-bit budget over configuration choice and operating point (server frequency f and network load rho). Numerical results show that edge processing gives lower latency but higher energy, while centralized processing gives lower energy but higher latency, with the gap widening under high load. The central trade-off claim is plausible and consistent with prior work, but two modeling aspects—the shared global load rho in the queueing analysis and the monotone structure of the optimization—need to be addressed before the quantitative conclusions can be accepted.

Significance. If the model were fully validated, the paper would provide a useful system-level framework for O-RAN deployment studies, especially the explicit inclusion of AI inference energy and latency alongside conventional baseband and transport components. The decomposition into processing, transport, and inference terms is a reasonable contribution, and the paper correctly identifies that configuration choice couples compute and transport costs. The paper also gives a clear, low-dimensional enumeration procedure for the MIP, which is practical. However, the load-dependent latency results and the claimed joint optimization of operating point are not yet supported by the current formulation. The low-load trade-off (edge vs. centralized) is robust and useful; the high-load saturation behavior, which is emphasized in the conclusions, depends on an unverified uniformity assumption about switch loads. No code or data are provided, so numerical claims cannot be checked directly. Overall the framework has potential, but the central quantitative and optimization claims require revision.

major comments (2)
  1. [Section III-B and Eqs. (5)-(7)] The latency model applies one global load rho to every switch along the path: L_FH = tau_o d_f + H_f,s L_sw, L_MH = tau_o d_m + H_m,s L_sw, and L_BH = tau_o d_b + H_b,s L_sw + H_b,r L_ro, where L_sw uses the same M/M/1 delay L_q(rho) from Eq. (8) on every hop. Yet the energy model in Eqs. (5)-(7) uses segment-specific traffic scaling factors gamma_f, gamma_m, gamma_b that differ by configuration, implying that the carried traffic per segment is known to be different. No derivation connects per-switch arrival rates to these gamma factors or to user traffic. Consequently, the high-load ordering in Figs. 4-7—especially Conf. B reaching ~7 ms at rho=0.99—is largely a mechanical consequence of multiplying a chosen divergent queue delay by hop counts. With realistic load tapering toward aggregation/core, the ordering of configurations under saturation could change. Please derive per-switch uti
  2. [Section IV-C, Eqs. (18)-(23)] The optimization minimizes L(x,f,rho) over configuration x and operating point (f,rho). The objective is monotone in both continuous variables: L_bp in Eq. (9) decreases as f increases (roughly as 1/f^2), and L_sw in Eq. (8) increases with rho. The energy constraint (19) uses per-configuration constants E_i that do not depend on f or rho. Therefore, for any feasible configuration i, the optimal operating point is trivially f=f_max and rho=rho_min; there is no frequency-load trade-off to optimize. The claim that the method 'jointly selects' the operating point is thus not realized by the formulation. Either make the energy model depend on f and rho (e.g., frequency-dependent processor power, load-dependent network equipment energy), add a coupling term that makes the objective non-monotone, or treat rho as an exogenous scenario parameter and state that the only decision is configuration s
minor comments (4)
  1. [Section III-B2] Units error: the text says 'service rate mu, measured in packets per second', then gives 'mu=0.138 Mbps'. For a 10 Gbps line card with jumbo frames, the service rate should be approximately 138,888 packets/s (or the equation should be written in bits/s). Please correct the units and verify that any numerical latency results that use mu are dimensionally correct.
  2. [Eq. (9)] The typesetting of L_bp = N_rb f^2 sum zeta_j i_m^j is ambiguous. Please clarify whether the intended formula is L_bp = (N_rb / f^2) * sum_j zeta_j i_m^j, and state the units of N_rb, zeta_j, and f explicitly, along with the source of the coefficient values used in the numerical evaluation.
  3. [Section III-A and Eqs. (5)-(7)] The notation gamma_u in Eq. (1) and gamma_(.) in Eqs. (5)-(7) is overloaded. It would help readers to use distinct symbols (e.g., gamma_u for node scaling and g_f, g_m, g_b for segment scaling) and to state explicitly in one place how each gamma depends on the BBP location.
  4. [Section V-B2] Typographical errors: 'fro configuration G' and 'the latency behavior fro configuration G' should be 'for'. Please proofread the results section.

Circularity Check

0 steps flagged

No significant circularity: the energy model is cited from prior work [5] with externally specified hardware parameters, and the latency/AI components come from independent literature; the central trade-off is not fitted or self-referential.

full rationale

I walked the paper's claimed derivation chain. The energy-per-bit formulas in Eqs. (1)-(7) are taken from the authors' own prior work [5], but the paper states this explicitly ('we start from the throughput-based O-RAN energy model in [5] and extend it') and the parameters (P_sc, C_sc, P_ls,u, C_ls,u, α, σ, φ, γ) are specified hardware/overprovisioning values, not quantities fitted to produce the paper's conclusions. The latency model is assembled from independent sources: M/M/1 queuing from [22], propagation/switching constants from [21], baseband-processing curve fit from [11], and AI inference latency decomposition from [27]. The optimization (18)-(23) enumerates four fixed configurations and compares stored per-configuration constants; no parameter is fitted to make the energy-latency trade-off appear. The only self-citation is [5], and although it is load-bearing for the energy component, it is not circular: it does not assume the target result, and its parameter values are externally specified. The reviewer concern that a single global ρ is applied to every switch (Eq. 8 feeding L_FH, L_MH, L_BH) is a modeling assumption that may affect the quantitative high-load ordering, but it is an assumption about traffic, not an equivalence between the model's inputs and outputs by construction; that belongs to correctness risk rather than circularity. No circular step meets the evidentiary bar of Eq. X = Eq. Y by definition or a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim rests on models inherited from prior work: the throughput energy equations (1),(4)-(7) come from the authors' own [5], the BBP latency polynomial (9) from curve fitting in [11], and the inference latency decomposition from [27]. No new physical entity is postulated. The main parameters are either set to literature values (lambda_in=1, link distances, switch latency constants) or adopted without tabulation (alpha,sigma,phi,gamma,zeta_j). The single-global-load assumption and the co-location of AI with BBP are the two most load-bearing domain assumptions.

free parameters (5)
  • Overprovisioning, overhead, coverage, and traffic scaling factors alpha, sigma, phi, gamma = from [5], not tabulated
    Used in Eqs. (1),(4)-(7); values are in the authors' prior paper [5], not in this manuscript.
  • Inference request rate lambda_in = 1 request/s
    Set in Section III-A1 from edge-AI studies [20]; directly scales AI energy per bit (Eq. 2).
  • BBP latency polynomial coefficients zeta_j = from [11], not listed
    Eq. (9) is a curve-fitted polynomial; coefficients are not reproduced, so L_bp(f) cannot be recomputed from this paper alone.
  • Transport link distances d_f, d_m, d_b = 20 km, 40 km, 100 km
    Chosen in Section V; they dominate the configuration latency ordering in Section V-B.
  • Switch service rate mu = example given as 0.138 Mbps (units inconsistent)
    Eq. (8) needs mu in packets/s; the text's example unit is wrong, so the queueing delay magnitude is not reliably specified.
axioms (6)
  • domain assumption Each egress switch port is an M/M/1 queue with Poisson arrivals and exponential service.
    Section III-B2 uses Eq. (8) from [22]; no justification is given for Poisson/exponential traffic in O-RAN fronthaul, midhaul, or backhaul.
  • domain assumption The same global network load rho applies to every switch in every traversed segment.
    The segment latency equations multiply hop counts by a single L_sw(rho); per-switch arrival rates are never derived from user traffic.
  • domain assumption AI inference latency is dominated by compute time, L_in approximately L_ic = N_op/T_ex.
    Section III-B5 states this tractability assumption; memory-bound LLM stages are explicitly not modeled.
  • ad hoc to paper BBP and AI inference are co-located at the selected configuration node.
    Section IV states BBP and AI inference are co-located; decoupled placement is deferred to future work.
  • ad hoc to paper One inference request per second per user (lambda_in=1).
    Section III-A1 adopts this from edge-AI literature [20].
  • domain assumption The curve-fitted BBP latency polynomial (9) from [11] holds for the modeled O-RAN servers.
    Adopted from [11],[26] without verification on the O-RU/O-DU/O-CU/DC platforms used here.

pith-pipeline@v1.3.0-daily-deepseek · 18572 in / 17808 out tokens · 115164 ms · 2026-08-04T15:24:12.034597+00:00 · methodology

0 comments
read the original abstract

The Open Radio Access Network (O-RAN) architecture introduces flexible functional splits and open interfaces that enable distributed and centralized deployment of baseband processing. While this flexibility offers opportunities for improved resource utilization, it also introduces fundamental trade-offs between energy efficiency and latency. In this paper, we develop a throughput-based end-to-end energy consumption model for O-RAN and extend it by incorporating detailed latency modeling and application-specific Artificial Intelligence/Machine Learning inference costs. The proposed end-to-end modeling framework provides a general representation of processing, transport, and inference-related energy and delay across the access, metro, and long-haul network segments. Building on this general model, we formulate an optimization problem that selects the placement of baseband processing and AI inference tasks across candidate O-RAN configurations to analyze energy-latency tradeoffs under network load, server frequency, and energy-budget constraints. Using representative hardware platforms and realistic traffic assumptions, we evaluate multiple baseband processing placements corresponding to different O-RAN functional configurations. Our results reveal how user quality of service requirements and network load conditions jointly determine the optimal placement of baseband processing and AI inference tasks, highlighting the inherent trade-off between energy efficiency and latency. The analysis provides practical insights for latency-aware and energy-efficient O-RAN deployments supporting emerging AI-driven services.

Figures

Figures reproduced from arXiv: 2608.02082 by Dan Kilper, Merim Dzaferagic, Petar Popovski, Rishu Raj, Shashi Raj Pandey, Urooj Tariq.

Figure 1
Figure 1. Figure 1: End-to-end O-RAN architecture comprising O-RU, O-DU, O-CU, and DC nodes connected through fronthaul (FH), midhaul (MH), and backhaul (BH) segments. The total energy and latency are decomposed into processing components, namely baseband processing and AI inference, and transport components. In the representative configuration shown, both baseband processing and AI inference are executed at the O-DU. Across … view at source ↗
Figure 2
Figure 2. Figure 2: Representation of the candidate functional configurations G, F, M and B. Each configuration places the BBP and AI inference module at a different O-RAN node, from the O-RU to the DC, resulting in progressively deeper centralization and additional transport segments. computational demand of the model) and Tex is the execution throughput of the device measured in floating-point operations per second [27]. Re… view at source ↗
Figure 3
Figure 3. Figure 3: Energy and latency for different functional configurations based on the location of the processing node under an unloaded network condition (ρ = 0) and fixed server frequency (f = 2 GHz). contrast, processing at the O-DU, O-CU, or DC benefits from more centralized computational resources and more efficient processors, leading to lower energy consumption. However, the latency results exhibit an opposite tre… view at source ↗
Figure 4
Figure 4. Figure 4: Variation in end-to-end latency with increase in network load ρ at different server frequencies f and functional configurations over (a) the full load range ρ ∈ [0.01, 0.99], and (b) zoomed-in view of the high-load regime ρ ∈ [0.80, 0.99]. 0 2 4 6 8 10 0 0.2 0.4 0.6 0.8 1 (a) Latency (ms) CDF Conf. G Conf. F Conf. M Conf. B 0 2 4 6 8 10 0 0.2 0.4 0.6 0.8 1 (b) Latency (ms) CDF Conf. G Conf. F Conf. M Conf.… view at source ↗
Figure 5
Figure 5. Figure 5: Empirical CDF of end-to-end latency under varying network load ρ for each functional configuration at three different server frequencies: (a) f = 0.8 GHz, (b) f = 2 GHz, and (c)f = 3.6 GHz. 1 1.5 2 2.5 3 3.5 0 2 4 6 8 10 12 14 Server frequency, f (GHz) Latency (ms) Conf. G, ρ = 0.01 Conf. F, ρ = 0.01 Conf. G, ρ = 0.97 Conf. F, ρ = 0.97 Conf. G, ρ = 0.99 Conf. F, ρ = 0.99 Conf. M, ρ = 0.01 Conf. B, ρ = 0.01… view at source ↗
Figure 6
Figure 6. Figure 6: Variation in end-to-end latency with change in server frequencies f at different loading conditions ρ and functional configurations. and near-saturation conditions because processing at the O￾RU is largely independent of traffic load. However, there is a significant increase in the latencies for configuration M and configuration B which involve longer transport paths that become more pronounced under near-… view at source ↗
Figure 7
Figure 7. Figure 7: Empirical CDF of one-way end-to-end latency across each functional configuration under different server frequencies for three different network load levels: (a) ρ = 0.01, (b) ρ = 0.97, and (c) ρ = 0.99. delay on top of the fronthaul. In configuration B, the backhaul traverses a larger number of intermediate devices and a more complex transport path, making this option the most sensitive to load and congest… view at source ↗
Figure 8
Figure 8. Figure 8: Latency heatmaps for different operating conditions of network load ρ and server frequency f for (a) configuration G, (b) configuration F, (c) configuration M, and (d) configuration B. The heatmaps are overlaid with latency budget contours indicating boundaries of feasibility regions for representative latency requirements. 8 10 12 14 16 18 20 22 7 8 9 10 11 12 13 Energy budget, Emax (mJ/bit) Energy consum… view at source ↗
Figure 9
Figure 9. Figure 9: Energy-aware latency optimization, illustrating energy consumption in the optimal configuration. and the corresponding latency for different values of energy budget Emax. deployments. Building on these insights, several directions remain open for further investigation. Future work can further enhance the proposed framework by incorporating resource￾sharing factors at different O-RAN nodes, including the O-… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 2 linked inside Pith

  1. [1]

    O-RAN specifications,

    O-RAN Alliance, “O-RAN specifications,” 2025. Accessed: Jul. 21,

    2025

  2. [2]

    NN-Meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices,

    L. L. Zhanget al., “NN-Meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices,” inProc. Int. Conf. Mobile Syst., Appl., Services, WI, USA, pp. 81–93, 2021

    2021

  3. [3]

    O-RAN: Disrupting the virtu- alized RAN ecosystem,

    A. Garcia-Saavedra and X. Costa-Perez, “O-RAN: Disrupting the virtu- alized RAN ecosystem,”IEEE Commun. Standards Mag., vol. 5, no. 4, pp. 96–103, 2021

    2021

  4. [4]

    Energy consumption of machine learning enhanced open RAN: A comprehensive review,

    X. Lianget al., “Energy consumption of machine learning enhanced open RAN: A comprehensive review,”IEEE Access, vol. 12, pp. 81889– 81910, 2024

    2024

  5. [5]

    Design and analysis of power consumption models for Open-RAN architectures,

    U. Tariq, R. Raj, and D. Kilper, “Design and analysis of power consumption models for Open-RAN architectures,” inProc. IEEE Int. Conf. Commun. Workshops, Montreal, QC, Canada, pp. 1037–42, 2025

    2025

  6. [6]

    O-RAN: Analysis of latency-critical interfaces and overview of time sensitive networking solutions,

    E. Municio, G. Garcia-Aviles, A. Garcia-Saavedra, and X. Costa-P ´erez, “O-RAN: Analysis of latency-critical interfaces and overview of time sensitive networking solutions,”IEEE Commun. Standards Mag., vol. 7, no. 3, pp. 82–89, 2023

    2023

  7. [7]

    On the road to 6G: Visions, requirements, key technologies, and testbeds,

    C.-X. Wanget al., “On the road to 6G: Visions, requirements, key technologies, and testbeds,”IEEE Commun. Surveys Tuts., vol. 25, no. 2, pp. 905–974, 2023

    2023

  8. [8]

    A survey of the functional splits proposed for 5G mobile crosshaul networks,

    L. M. P. Larsen, A. Checko, and H. L. Christiansen, “A survey of the functional splits proposed for 5G mobile crosshaul networks,”IEEE Commun. Surveys Tuts., vol. 21, no. 1, pp. 146–172, 2019

    2019

  9. [9]

    Toward sustainable O-RAN deployment: An in-depth analysis of power consumption,

    G. Baldiniet al., “Toward sustainable O-RAN deployment: An in-depth analysis of power consumption,”IEEE Trans. Green Commun. Netw., vol. 9, no. 2, pp. 429–444, 2025

    2025

  10. [10]

    Power-efficient baseband-function placement in latency-constrained 5G metro access,

    L. M. M. Zorello, M. Sodano, S. Troia, and G. Maier, “Power-efficient baseband-function placement in latency-constrained 5G metro access,” IEEE Trans. Green Commun. Netw., vol. 6, no. 3, pp. 1683–96, 2022

    2022

  11. [11]

    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

    2021

  12. [12]

    DRL-based energy-efficient baseband function deployments for service-oriented open RAN,

    H. Liet al., “DRL-based energy-efficient baseband function deployments for service-oriented open RAN,”IEEE Trans. Green Commun. Netw., vol. 8, no. 1, pp. 224–237, 2024

    2024

  13. [13]

    Energy-aware CPU orchestration in O-RAN: A dApp- driven lightweight approach,

    F. Crespoet al., “Energy-aware CPU orchestration in O-RAN: A dApp- driven lightweight approach,” 2025. arXiv:2508.00629. Accessed: Jul. 21, 2026. [Online]. Available: https://arxiv.org/abs/2508.00629

    Pith/arXiv arXiv 2025

  14. [14]

    Statistical characterization of closed-loop latency at the mobile edge,

    S. Sumanet al., “Statistical characterization of closed-loop latency at the mobile edge,”IEEE Trans. Commun., vol. 71, no. 7, pp. 4391–4405, 2023

    2023

  15. [15]

    Unified timing analysis for closed-loop goal-oriented wireless communication,

    L. Li, A. E. Kalør, P. Popovski, and W. Chen, “Unified timing analysis for closed-loop goal-oriented wireless communication,”IEEE Trans. Wireless Commun., vol. 24, no. 4, pp. 2814–2831, 2025

    2025

  16. [16]

    Statistical analysis for energy-efficient satellite edge computing with latency guarantees,

    N. D. Lyholmet al., “Statistical analysis for energy-efficient satellite edge computing with latency guarantees,” 2026. arXiv:2605.10215. Accessed: Jul. 21, 2026. [Online]. Available: https://arxiv.org/abs/2605. 10215

    Pith/arXiv arXiv 2026

  17. [17]

    Roofline: An insightful visual performance model for multicore architectures,

    S. Williams, A. Waterman, and D. Patterson, “Roofline: An insightful visual performance model for multicore architectures,”Commun. ACM, vol. 52, no. 4, pp. 65–76, 2009

    2009

  18. [18]

    MobileNetV2: Inverted residuals and linear bottle- necks,

    M. Sandleret al., “MobileNetV2: Inverted residuals and linear bottle- necks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Salt Lake City, UT, USA, pp. 4510–4520, 2018

    2018

  19. [19]

    How to evaluate deep neural network processors: TOPS/W (alone) considered harmful,

    V . Sze, Y .-H. Chen, T.-J. Yang, and J. S. Emer, “How to evaluate deep neural network processors: TOPS/W (alone) considered harmful,”IEEE Solid-State Circuits Mag., vol. 12, no. 3, pp. 28–41, 2020

    2020

  20. [20]

    Edge AI: On-demand acceler- ating deep neural network inference via edge computing,

    E. Li, L. Zeng, Z. Zhou, and X. Chen, “Edge AI: On-demand acceler- ating deep neural network inference via edge computing,”IEEE Trans. Wireless Commun., vol. 19, no. 1, pp. 447–457, 2020

    2020

  21. [21]

    Edge-enhanced graph neural network for DU-CU place- ment and lightpath provision in X-Haul networks,

    R. Wanget al., “Edge-enhanced graph neural network for DU-CU place- ment and lightpath provision in X-Haul networks,”J. Opt. Commun. Netw., vol. 14, no. 10, pp. 828–839, 2022

    2022

  22. [22]

    C. G. Cassandras and S. Lafortune,Introduction to Discrete Event Systems. Springer, 2nd ed., 2008

    2008

  23. [23]

    Scal- ability and performance evaluation of edge cloud systems for latency constrained applications,

    S. Maheshwari, D. Raychaudhuri, I. Seskar, and F. Bronzino, “Scal- ability and performance evaluation of edge cloud systems for latency constrained applications,” inProc. IEEE/ACM Symp. Edge Comput., Bellevue, W A, USA, pp. 286–299, 2018

    2018

  24. [24]

    Evaluation of process bus data latency with PRP/HSR enabled ethernet switches,

    M. Han, P. A. Crossley, and S. Salyani, “Evaluation of process bus data latency with PRP/HSR enabled ethernet switches,” inProc. Int. Conf. Develop. Power Syst. Protection, Liverpool, U.K., pp. 1–6, 2020

    2020

  25. [25]

    A machine learning approach to estimating queuing delay on a router over a single-hop path,

    T. Rickeret al., “A machine learning approach to estimating queuing delay on a router over a single-hop path,” inProc. IEEE Int. Conf. Commun., Seoul, South Korea, pp. 2720–2725, 2022

    2022

  26. [26]

    Understanding the computa- tional requirements of virtualized baseband units using a programmable cloud radio access network testbed,

    T. X. Tran, A. Younis, and D. Pompili, “Understanding the computa- tional requirements of virtualized baseband units using a programmable cloud radio access network testbed,” inProc. IEEE Int. Conf. Auton. Comput., Columbus, OH, USA, pp. 221–226, 2017

    2017

  27. [27]

    Paleo: A performance model for deep neural networks,

    H. Qi, E. R. Sparks, and A. Talwalkar, “Paleo: A performance model for deep neural networks,” inProc. Int. Conf. Learn. Representations, Toulon, France, pp. 1–10, 2017

    2017

  28. [28]

    Efficient LLMs training and inference: An introduction,

    R. Liet al., “Efficient LLMs training and inference: An introduction,” IEEE Access, vol. 13, pp. 32944–32970, 2025

    2025

  29. [29]

    Mobile data traffic forecast—ericsson mobility report,

    Ericsson, “Mobile data traffic forecast—ericsson mobility report,” 2025. Accessed: Jul. 21, 2026. [Online]. Available: https://www.ericsson.com/en/reports-and-papers/mobility-report/ dataforecasts/mobile-traffic-forecast

    2025

  30. [30]

    Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices,

    Y .-H. Chen, T.-J. Yang, J. Emer, and V . Sze, “Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices,”IEEE J. Emerg. Sel. Topics Circuits Syst., vol. 9, no. 2, pp. 292–308, 2019

    2019

  31. [31]

    NVIDIA A100 tensor core GPU,

    NVIDIA, “NVIDIA A100 tensor core GPU,” Dec. 25, 2025. Ac- cessed: Jul. 21, 2026. [Online]. Available: https://www.nvidia.com/ en-us/data-center/a100/

    2025

  32. [32]

    L4 tensor core GPU for AI and graphics,

    NVIDIA, “L4 tensor core GPU for AI and graphics,” 2025. Ac- cessed: Jul. 21, 2026. [Online]. Available: https://www.nvidia.com/ en-us/data-center/l4/

    2025

  33. [33]

    Optical front/mid-haul with open access- edge server deployment framework for sliced O-RAN,

    S. Mondal and M. Ruffini, “Optical front/mid-haul with open access- edge server deployment framework for sliced O-RAN,”IEEE Trans. Netw. Service Manag., vol. 19, no. 3, pp. 3202–3219, 2022

    2022

  34. [2026]

    Available: https://www.o-ran.org/specifications

    [Online]. Available: https://www.o-ran.org/specifications