{"id":"f6e453e0-9bfa-4113-af0e-f698e0422f26","arxiv_id":"2505.22149","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"On a real 5G testbed, offloading a CNN with early exits and splits to a MEC server reduces inference delay by up to 2.5x and energy by up to 2.6x versus local processing, but the analytical models are fitted without independent validation.","lead":"This paper tests what happens when a car sends parts of its road-sign recognition network to a nearby server over a real 5G link, instead of running the whole network on the car. It finds that splitting the network can cut time and energy, but the new mathematical models need more checking before they can be used for prediction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) charges Pcomm over the full ttotal, so local-only S=5 predicts nonzero radio energy despite the paper stating Ecomm=0; the energy model is internally inconsistent and unvalidated.","rationale":"The paper's primary empirical result, that offloading early-exit/split CNN inference over the real 5G MEC testbed reduces total delay and energy, is supported by direct repeated measurements and is plausible given the asymmetric compute power. However, the paper's third stated contribution is the analytical model, and Eq. (7) is internally inconsistent: it charges the 5G modem's power Pcomm during the entire processing time, including phases when no communication occurs. The text for Fig. 8 explicitly says Ecomm=0 for S=5, but the formula yields Pcomm * t_total > 0 for that case. This is not a matter of overfitting or generalization; it is a structural error in the model's definition. Since the conclusion and abstract advertise the models, the modeling contribution needs correction and re-validation before acceptance. The reader's weakest_assumption already identified the Pcomm issue, and the recommended CONDITIONAL verdict remains appropriate; I see no reason to move it to REJECT because the measurements themselves still support the feasibility claim, and no reason to ACCEPT because the modeling contribution is not sound as written. The proposed concrete test directly checks the inconsistency and validates a corrected model. Other potential concerns (e.g., parameter fitting to one testbed, NTP synchronization accuracy, the fairness of comparing a 7 GFLOPS Atom to a 52 TFLOPS GPU) are weaker: they affect generality but not the internal correctness of the reported experiments. Thus, the single most load-bearing concern is the Eq. (7) energy-model contradiction.","tokens_in":8993,"tokens_out":8283,"duration_ms":88353,"concrete_test":"Plug the S=5 row of Table II into Eq. (7): set t_comm=0, t_MEC=0, use t_total from Fig. 5 and Pcomm=0.79 W; compute the Pcomm * t_total term. If it is nonzero, Eq. (7) is inconsistent with the paper's Ecomm=0 statement. Then refit the energy model with the last term changed to Pcomm * (t_ul + t_dl + t_MEC) (or only t_ul + t_dl) and compare leave-one-out predictions against the measured E_total for all (E,S) in Figs. 7-8; the corrected model should reproduce the 2.6x ratio if the measurements are robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is in the energy model, Eq. (7): E_total = (t_comm + t_MEC) * P_idle + t_prep * P_prep + t_proc * P_comp + t_total * P_comm. The last term multiplies the communication power by the entire processing time, including local CPU computation, preprocessing, and idle phases. For the full-local baseline S=5, t_comm=0 and t_MEC=0, yet t_total>0, so the model returns a nonzero 'communication' energy. This directly contradicts Section V (Fig. 8 discussion), which states that for S=5 'communication energy E_comm is zero.' Because Eq. (7) is the basis for the claimed 2.6x energy reduction whenever the model is used or extended, the model cannot be trusted as a predictive tool without correcting this term. The direct power-meter/RAPL measurements in Figs. 7-8 are not invalidated by this, so the feasibility claim retains empirical support, but the analytical-model contribution, a stated major contribution, is not sound as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a real-world 5G MEC testbed in which a VGG-16 road-sign classifier with early exits and split points runs on an autonomous vehicle, with partial or full inference offloaded to an edge server via OpenAirInterface. The authors measure total processing delay, AV energy consumption, and classification accuracy across split/exit combinations, reporting reductions of up to 2.5 times in delay and up to 2.6 times in energy relative to full local processing, and then fit analytical models for delay and energy using parameters listed in Table II. The direct measurements form the main empirical contribution; the analytical models are intended to support future theoretical work on offloading.","tokens_in":9256,"tokens_out":5528,"duration_ms":62043,"significance":"The empirical core is a valuable and relatively rare real-hardware check of offloading claims for CNN inference with early exits and splits. Strengths include 250 repetitions per configuration, NTP-synchronized clocks, RAPL and USB power-meter measurements, consideration of both uplink and downlink communication, and the stated public release of code and data. The headline delay and energy reductions are measured directly and do not depend on the analytical models, so the feasibility claim remains credible even where the modeling section is flawed. If the models are repaired and validated, the paper would offer a useful basis for future offloading designs; as written, the modeling contribution is not sound enough to be used as published.","major_comments":[{"comment":"The last term, t_total * Pcomm, charges communication power over the entire processing time, including local computation, preprocessing, and idle phases. For the full-local baseline S=5, Eq. (4) gives tcomm=0 and Eq. (5) gives tMEC=0, yet the model returns a nonzero communication energy equal to t_total * Pcomm. This directly contradicts the statement in Section V and Fig. 8 that communication energy is zero for S=5. Because Eq. (7) is the basis of the energy model, the model cannot be trusted until the phase durations are corrected, for example by multiplying Pcomm by tcomm rather than ttotal, or by explicitly modeling the modem power over the communication phases only.","section":"Section VI, Eq. (7)"},{"comment":"The MEC computing term in Eq. (5) is written as a sum over i=1..min(S,E) of C_i^D / C_MEC, which makes the local and MEC subtasks overlap for every split point. The surrounding text and Section VI, Eq. (9) define the MEC term as a sum over i=S+1..E. As printed, Eq. (5) does not represent the described split architecture and cannot be used as the delay model underlying the paper's analysis.","section":"Section II, Eq. (5)"},{"comment":"The delay and energy models are fitted on the same measurements used to report the headline results, with no held-out validation, no confidence intervals on the fitted parameters, and no residual analysis. Since deriving practical models is a stated major contribution, the paper should include a validation procedure, such as leave-one-configuration-out cross-validation or a train/test split, and report prediction errors. Without this, the models' predictive value beyond the specific testbed is unsupported.","section":"Section VI and Table II"}],"minor_comments":[{"comment":"The parameter labeled dAS appears where Eq. (8) uses dAV; please unify the notation.","section":"Table II"},{"comment":"Classification accuracy is reported without error bars or confidence intervals; since each configuration is repeated 250 times, reporting dispersion would strengthen the claim that accuracy is not impaired by split points and compression.","section":"Section V, Fig. 4"},{"comment":"The phrase 'up 2.5 times' should read 'up to 2.5 times'.","section":"Abstract and Conclusion"},{"comment":"The sentence beginning 'While the number of processed tasks is minimized, the total processing delay is not considered' is garbled and should be revised.","section":"Introduction"},{"comment":"Energy components would benefit from uncertainty bars, given the stated ±1.41% power-meter accuracy and the use of RAPL measurements.","section":"Section V, Fig. 8"},{"comment":"Reference [4] is cited for the AV architecture and MEC integration; please clarify the incremental contribution of this paper relative to that prior work.","section":"Section III"}],"recommendation":"major_revision","confidential_remarks":"The direct measurements appear credible and the paper has strong reproducibility assets. The main obstacle is the modeling section: Eq. (7) is internally inconsistent, Eq. (5) has an indexing error, and the fitted models lack validation. These are fixable, so I recommend major revision rather than rejection. Please also verify the novelty and overlap with the companion VTC2025 paper [4]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis is a genuinely useful empirical paper, and I'd send it to a serious referee. But the analytical model has an equation-level error that needs to be fixed before the model is used.\n\nWhat's new: the authors built a real 5G testbed (OpenAirInterface + USRP N310) with an actual vehicle-like device and a MEC server, implemented a VGG-style CNN with early exits and split points, and measured end-to-end delay, energy, and accuracy over the air. Prior work on early exits and splits for MEC is almost entirely simulation. They also explicitly capture downlink delay, which most offloading studies ignore, and show it's a substantial cost. The measurement methodology is solid: 250 repetitions per configuration, synchronized clocks, RAPL energy readings plus a separate power meter for the modem. Those measurements support the headline claims: full offloading cuts total delay by up to 2.5x and energy by up to 2.6x versus local processing, with accuracy roughly flat across split points for a given exit. Code and data are public on GitLab.\n\nThe soft spot is the modeling. Equation (7) multiplies Pcomm by ttotal rather than by the communication time tcomm. For the all-local case (S=5), tcomm=0 and tMEC=0, yet ttotal is positive, so the formula returns a nonzero \"communication\" energy. That directly contradicts the text and Figure 8, where Ecomm is zero for S=5. The stress-test note is correct; this isn't cosmetic. Any optimization or theoretical study that uses this model will inherit the error. Also, the models in Section VI are fitted to the same measurements with parameters in Table II, and there's no held-out validation or uncertainty on the fitted constants. As written, they're descriptive curves, not predictive tools. Minor points: \"dAS\" in Table II should be \"dAV\", the abstract's \"not impairing classification accuracy\" is true per-exit but could be misread, and the introduction has a garbled sentence about task minimization.\n\nNone of this undermines the direct measurements. The feasibility result is real. But the model contribution isn't sound as written.\n\nFor peer review: I'd accept with major revision—fix Eq. (7), validate the models, clean up the prose. The empirical core is worth publishing, and it gives the community someone else's real numbers to test their own offloading algorithms against.","headline":"A genuinely useful real-world 5G testbed study of early-exit/split CNN offloading, but the analytical energy model has an equation-level inconsistency that needs fixing.","tokens_in":9733,"tokens_out":2804,"would_cite":true,"duration_ms":27466,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Offloading early-exit CNNs to a real 5G MEC server cuts total processing delay by up to 2.5 times and energy by up to 2.6 times compared with full local processing on the vehicle, while classification accuracy for each exit stays nearly…","keywords":["multi-access edge computing","computation offloading","early exit","split computing","CNN inference","autonomous vehicles","real-world measurement","delay and energy modeling"],"falsifier":"Repeat the measurement protocol on a second testbed with a different vehicle processor or a different 5G bandwidth, and compare observed total delay and energy for every split and exit against predictions from Equations (6) to (10) with the Table II constants; systematic divergence beyond measurement error would show the model is testbed-specific rather than general.","tokens_in":8812,"feed_emoji":"📡","tokens_out":3555,"duration_ms":37766,"temperature":0.7,"pith_summary":"The paper tries to establish that computation offloading of convolutional neural networks with early exits and split points is not just a simulation idea but works on real hardware. Using a real autonomous vehicle, a software-defined 5G link, and a MEC server, the authors measure that offloading reduces total processing delay by up to 2.5 times and total energy consumption by up to 2.6 times compared with full local processing. They also measure that, for a given early exit, classification accuracy remains nearly constant across split points, meaning the split can be chosen for delay or energy without sacrificing accuracy. From these measurements they derive analytical models of total delay and energy that include fixed communication overheads, preprocessing, computing, and idle power. A sympathetic reader would care because this gives realistic, testbed-derived equations for future offloading research, which previously relied on optimistic simulations.","feed_headline":"Offloading early-exit CNNs cuts delay 2.5x on real 5G","feed_subtitle":"Measured on a real AV and MEC testbed, split-and-exit CNN inference also cuts energy 2.6x with accuracy intact.","key_machinery":"The central object is a VGG-style CNN organized into five blocks, with four early exits and six split points co-located between blocks, plus autoencoders at each split to compress the feature maps before transmission. The argument is carried by decomposing the total inference cost into per-phase delays: local computing with a fixed overhead per block, MEC computing with its own fixed overhead, preprocessing delay linear in compressed data volume, and uplink and downlink communication delays that include constant network overheads. The fitted parameters in Table II, measured on the real testbed, make the model concrete and testable against future deployments.","core_discovery":"The paper claims that a CNN with multiple early exits and split points can be flexibly offloaded from a resource-limited autonomous vehicle to a MEC server over a real 5G network, and that this reduces the total processing delay by up to 2.5 times and total energy consumption by up to 2.6 times compared to full local processing. The measurements further show that classification accuracy depends mainly on which exit is used, rising from about 32 percent at the first exit to 93 percent at the main exit, while being nearly flat across split points for a fixed exit. Based on these experiments, the authors propose refined equations for total delay and energy, adding constant per-block overheads, a preprocessing delay proportional to compressed data volume, and fitted power values for idle, preprocessing, computing, and communication phases.","pith_inferences":["Inference: The fitted constants, especially the power values and fixed delays, are likely tied to this specific CPU, network stack, and software version; transferring the model to another autonomous platform would require re-calibration or an online parameter-estimation step.","Inference: The energy model applies communication power to the entire total processing time, which assumes the modem stays active while the device waits for results; a pipelined implementation that overlaps local computing with communication could change that term and make offloading even more attractive.","Inference: The near-constant accuracy across splits suggests that later split points can be used freely, but the low accuracy of early exits means those exits are only useful on easy inputs unless confidence-based gating is added.","Inference: Combining the fitted delay and energy equations with real-time estimates of uplink and downlink bitrate would enable the online split-and-exit selection that the paper lists as future work."],"forward_implications":["Offloading decisions for CNN-based tasks can be made at split-point granularity using measured constants rather than idealized simulation parameters.","Because accuracy stays nearly constant across split points for a given exit, the split can be chosen mainly to minimize delay and energy without degrading classification.","Early exits provide the larger delay and energy reduction, up to 4.2 and 4.4 times respectively, while offloading adds an additional factor of about 2.5 and 2.6 times versus full local processing.","Downlink delay and fixed network overheads are significant and should be included in any realistic offloading model, contrary to many prior studies that neglect them.","The open-source implementation and data allow other researchers to reproduce the measurements and re-fit the model for their own hardware."],"supporting_citations":[{"why":"Supplies the real autonomous-vehicle hardware, MEC integration, and measurement setup that the experiments are built on.","marker":"[4]"},{"why":"Introduces early-exit inference for deep neural networks, the mechanism whose delay and energy benefits are measured here.","marker":"[7]"},{"why":"Provides split computing as a way to partition CNN inference between device and edge, the core offloading mechanism evaluated.","marker":"[8]"},{"why":"Explores the combination of early exits and split computing, which this paper implements and measures on real hardware.","marker":"[12]"},{"why":"Supplies the supervised compression approach for split computing that the autoencoders at each split point rely on.","marker":"[14]"},{"why":"Defines the 5G modulation, coding, and resource-block parameters used to compute the uplink and downlink bitrates.","marker":"[16]"},{"why":"Provides the German Traffic Sign Recognition Benchmark dataset used for training and accuracy evaluation.","marker":"[20]"}],"fun_headline_variants":["Real 5G AV test: early-exit CNN offloading cuts delay 2.5x, energy 2.6x","Split-and-exit CNN on real AV: 2.5x faster, 2.6x less energy measured","Measured on 5G: offloaded early-exit CNN reduces delay 2.5x and energy 2.6x","AV offloads early-exit CNNs: real 5G cuts latency 2.5x, energy 2.6x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analytical models assume that a single set of constants fitted to one testbed, along with the assumption that communication power applies over the entire processing time, captures delay and energy across all split and exit combinations and across other deployments.","fun_headline_variants_meta":{"raw":{"variants":["Real 5G AV test: early-exit CNN offloading cuts delay 2.5x, energy 2.6x","Split-and-exit CNN on real AV: 2.5x faster, 2.6x less energy measured","Measured on 5G: offloaded early-exit CNN reduces delay 2.5x and energy 2.6x","AV offloads early-exit CNNs: real 5G cuts latency 2.5x, energy 2.6x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000379,"raw_usage":{"total_tokens":1977,"prompt_tokens":873,"completion_tokens":1104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":975}},"tokens_in":489,"tokens_out":1104,"duration_ms":10378,"temperature":1.0,"reasoning_tokens":975,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:13:42.586702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the measurement protocol on a second testbed with a different vehicle processor or a different 5G bandwidth, and compare observed total delay and energy for every split and exit against predictions from Equations (6) to (10) with the Table II constants; systematic divergence beyond measurement error would show the model is testbed-specific rather than general.","supporting_citations":[{"cited_title":"Danek, Z","cited_arxiv_id":null,"evidence_quote":"Supplies the real autonomous-vehicle hardware, MEC integration, and measurement setup that the experiments are built on."},{"cited_title":"Teerapittayanon, et al., ”Branchynet: Fast inference via early exiting from deep neural networks,” IEEE ICPR, 2016","cited_arxiv_id":null,"evidence_quote":"Introduces early-exit inference for deep neural networks, the mechanism whose delay and energy benefits are measured here."},{"cited_title":"Bakhtiarnia, et al., ”Dynamic Split Computing for Efficient Deep EDGE Intelligence,” IEEE ICASSP, pp","cited_arxiv_id":null,"evidence_quote":"Provides split computing as a way to partition CNN inference between device and edge, the core offloading mechanism evaluated."},{"cited_title":"Rauch, Z","cited_arxiv_id":null,"evidence_quote":"Explores the combination of early exits and split computing, which this paper implements and measures on real hardware."},{"cited_title":"Matsubar, et al., ”SC2 benchmark: Supervised compression for split computing,” Transactions on machine learning research, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised compression approach for split computing that the autoencoders at each split point rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 5G modulation, coding, and resource-block parameters used to compute the uplink and downlink bitrates."},{"cited_title":"Stallkamp, M","cited_arxiv_id":null,"evidence_quote":"Provides the German Traffic Sign Recognition Benchmark dataset used for training and accuracy evaluation."}],"review_version":1}