{"id":"255074d7-722a-41ea-874b-8eccf5532361","arxiv_id":"2505.22469","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-informed residual neural network trained on per-unit power labels cuts power estimation error by 74 to 85 percent over the unsupervised ABPI baseline on Jetson hardware and a simulated heterogeneous SoC.","lead":"The paper tests a temperature-based power estimator on an Nvidia Jetson chip, finds it inaccurate, then adds a small neural network to correct it, cutting power prediction error by roughly 74 to 85 percent. The result points to a practical way to get accurate per-component power readings without adding power sensors to future chips.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Supervised-vs-unsupervised confound: CPINN-ABPI uses p_true labels, so the 84.7%/73.9% gain over blind ABPI does not establish the physics-informed contribution.","rationale":"The paper has two separable contributions: the first empirical validation of ABPI on commercial hardware, and CPINN-ABPI as a physics-informed correction. The first is useful and plausible, but the headline accuracy numbers for the second are measured against a baseline that lacks the most important input the proposed model receives: per-unit ground-truth power labels. The reader's weakest assumption correctly identifies this supervised-versus-unsupervised confound. A supervised network could plausibly learn a direct mapping from temperature readings to power, especially on a board with clean unit-level power sensors, and the residual branch plus fine-tuned A' and B' gives the model enough capacity to act as that mapping. Therefore, the experiment as designed cannot attribute the reported improvement to the physics-informed mechanism. This is not an internal contradiction in the method, but a missing control condition in the evaluation. Since the paper does not prove that the physics branch is necessary, yet claims that the physics-informed design produces the gains, the conditional verdict is appropriate and no verdict change is needed. The identical ABPI CPU/GPU metrics in Table II are an independent red flag that should be resolved in revision, but the primary load-bearing concern remains the absence of a supervised baseline and physics-loss ablation.","tokens_in":9889,"tokens_out":4573,"duration_ms":56373,"concrete_test":"Re-run the Jetson AGX Xavier experiment with two additional baselines: (1) a plain feedforward network with the same residual-branch architecture trained on the same inputs [t_prev; t_curr] and optionally P_est, using only L_data, with no physics branch and with lambda_phys = lambda_guide = 0, under identical 10-fold splits and NSGA-II selection; (2) CPINN-ABPI with lambda_phys = lambda_guide = 0 to ablate the physics losses. If either baseline achieves CPU/GPU MAE and WMAPE statistically indistinguishable from the reported 0.57/0.98 W and about 12% (for example within 10% relative), then the claimed role of the physics-informed design is not supported and the comparison reduces to supervised MLP versus unsupervised ABPI. Also recompute the ABPI CPU and GPU rows in Table II from raw traces; identical MAE and MSE to two decimals would indicate a trace-processing or reporting error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a physics-informed residual PINN dramatically improves ABPI is not established by the reported experiments, because the comparison is confounded by supervision. CPINN-ABPI is trained with per-unit ground-truth labels: Eq. 3 defines L_data = (1/N) sum ||p_final - p_true||^2, Algorithm 1 lines 14-17 minimize this loss, and NSGA-II in Algorithm 2 selects architectures using MAE against p_true on held-out data. The ABPI baseline is blind and unsupervised, using only temperatures and total power. Thus the 84.7%/73.9% MAE reductions and the simulated 85-99% gains could be achieved by any supervised residual network that directly learns p_true from t_prev and t_curr; they do not by themselves validate the physics branch, the physics consistency loss (Eq. 4), or the guidance loss (Eq. 5). Moreover, because A' and B' are trainable and the residual network input includes P_est, the combined model can act as an arbitrary function of the temperature sequence, making the physics branch largely an initialization rather than a verified inductive bias. Table II also reports exactly identical ABPI MAE (3.74 W) and MSE (32.10 W^2) for CPU and GPU, indicating a possible data or reporting inconsistency that should be checked. A plain supervised MLP baseline and a lambda_phys = lambda_guide = 0 ablation are needed before the physics-specific contribution can be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CPINN-ABPI, a hybrid power estimation method for MPSoCs that combines the ABPI linear thermal state-space model with a residual neural network correction, trained with a multi-component loss (data fidelity, physics consistency, and physics guidance) and tuned by NSGA-II for the accuracy/compute tradeoff. The authors report the first empirical validation of ABPI on an NVIDIA Jetson AGX Xavier board, showing that ABPI has large errors (MAE 3.74 W, WMAPE 47–81%), and claim that CPINN-ABPI reduces MAE by 84.7% (CPU) and 73.9% (GPU), with WMAPE about 12%, and similar 85–99% MAE improvements on a simulated heterogeneous SoC, while keeping inference sub-millisecond.","tokens_in":10200,"tokens_out":3902,"duration_ms":45415,"significance":"If the central claim were fully established, the paper would make a useful empirical contribution: it would be the first public validation of ABPI on commercial hardware, and it would demonstrate that a supervised residual network trained with a physics-inspired loss can substantially improve blind power identification. The paper also contributes a new Jetson dataset and an NSGA-II-based architecture search for real-time power models. However, the experimental design as presented does not separate the effect of supervision from the effect of the physics-informed branch, and the physics constraint itself is not shown to be an independent prior because the thermal matrices are re-estimated during training. The reported gains are therefore not yet attributable to the proposed method's distinctive features, and a plain supervised baseline and a proper ablation are needed before the significance claimed in the title and abstract can be accepted.","major_comments":[{"comment":"The headline accuracy gains are confounded by supervision: CPINN-ABPI is trained with per-unit ground-truth labels (L_data in Eq. 3, Algorithm 1 lines 14–17), whereas the ABPI baseline is blind and unsupervised, using only temperature and total power. On this evidence alone, the 84.7% CPU and 73.9% GPU MAE reductions in Table II could be achieved by any supervised residual network and do not establish the contribution of the physics branch, Eq. 4, or Eq. 5. The paper should add a plain supervised baseline (e.g., an MLP with the same inputs and training protocol) and an ablation with λ_phys = λ_guide = 0.","section":"II-B, Algorithm 1, Table II"},{"comment":"The physics consistency loss is not an independent constraint: the matrices A' and B' in Eq. 4 are initialized from ABPI estimates on the same data and then further optimized by gradient descent (Algorithm 1 lines 21–22). Consequently L_phys can be minimized by changing A' and B' to conform to the network's power predictions, so it does not independently enforce thermodynamic consistency. To support the 'physics-informed' claim, the authors should either fix A and B to independently identified values, or demonstrate that the trained A', B' remain close to physically meaningful values and that the accuracy gain is not lost when L_phys is replaced by a purely data-driven regularization.","section":"II-A, Eq. 4, Algorithm 1"},{"comment":"The NSGA-II selection in Algorithm 2 evaluates fitness on D_test (line 9) and selects the Pareto-optimal architecture using that test set. Reporting the selected model's performance on the same D_test in Section IV therefore gives optimistically biased estimates. The 10-fold cross-validation applied after selection does not remove this selection bias because the architecture was chosen using the test data. The evaluation protocol should use a separate held-out test set that is never touched during NSGA-II, or use nested cross-validation.","section":"Algorithm 2, Section IV-A"},{"comment":"Table II reports identical ABPI MAE (3.74 W) and MSE (32.10 W^2) for both CPU and GPU, which is implausible for two units with different power profiles and workloads. This suggests a transcription or computation error in the baseline results; because the improvement percentages are computed relative to these ABPI numbers, the baseline must be verified and corrected.","section":"Table II, Section IV-B"}],"minor_comments":[{"comment":"The caption reads 'Experimental setup overflow' instead of 'overview'.","section":"Figure 1"},{"comment":"The heading 'Physics Guidance Loss::' contains a double colon; the extra colon should be removed.","section":"Eq. 5"},{"comment":"The notation for the physics branch is inconsistent: the text uses p_physics = Δt·(B^{-1})^T and t_prev^T A, while Algorithm 1 writes P_physics ← ΔT·(B'^{-1})^T and A'^T T_prev; please unify the transpose conventions and clarify the shapes of A, B, and the vectors.","section":"Section II-A"},{"comment":"The paper does not report the number of samples in the training, validation, and test sets, the duration of each workload trace, or the train/test split ratio; these details are needed to interpret the reported 10-fold cross-validation and the final test errors.","section":"Section III"},{"comment":"The NSGA-II description gives the number of generations and population size, but not the crossover and mutation rates or the ranges of the searched hyperparameters, which limits reproducibility.","section":"Section IV-A"},{"comment":"Figure 8 shows a 2500-second test dataset, but the paper does not state whether this is a single continuous trace or a concatenation of the ten workloads in Table I, nor how many independent runs were averaged; please clarify.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is not yet supported because the comparison is confounded and the physics constraint is not shown to be an independent inductive bias. I would encourage the editor to request the additional baselines and ablations before further consideration. There is also a possible double use of the test set in the NSGA-II selection, which should be caught in revision. The identical ABPI numbers for CPU and GPU in Table II should be checked for a data error."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper gives us the first hardware validation of ABPI on a commercial MPSoC (NVIDIA Jetson AGX Xavier), and it shows ABPI's blind estimates are far off on real silicon. That is a genuine new data point. The second part, CPINN-ABPI, reports large accuracy gains, but the experiment design does not support the claim that the physics-informed branch is what makes the gains happen.\n\nWhat the paper does well: the Jetson validation is new and useful. The residual PINN correction is a sensible engineering idea, and the sub-millisecond inference target is appropriate for power management. The simulated heterogeneous SoC results, though less compelling, suggest the method generalizes. The writing is clear, and the method is described in enough detail to be reproduced (modulo missing code/data).\n\nThe soft spots are real, and they are load-bearing for the paper's central claim. CPINN-ABPI is trained with per-unit ground-truth power labels (Eq. 3, Algorithm 1), while ABPI is unsupervised and blind. The reported 84.7%/73.9% MAE reductions could plausibly be achieved by any supervised residual network that directly learns p_true from temperatures; the results do not validate the physics branch, the physics consistency loss (Eq. 4), or the guidance loss (Eq. 5). A plain supervised MLP baseline and a lambda_phys=lambda_guide=0 ablation are needed before we can attribute anything to the physics.\n\nSecond, the thermal matrices A and B are estimated from the same data by ABPI and then fine-tuned during training. So the physics consistency loss enforces consistency with a fitted model, not an independent physical law. That weakens the 'physics-informed' framing considerably. I'd call this a circularity problem, though not a fatal one if the paper were reframed as a hybrid data-driven estimator.\n\nThird, Table II lists identical ABPI MAE and MSE for CPU and GPU (3.74 W and 32.10 W^2). That looks like a typo or a reporting inconsistency and must be checked. Fourth, the latency measurement is on an NVIDIA A100 GPU, not on the Jetson, so the real-time claim for the target platform is not demonstrated. No code or data is released, which limits reproducibility.\n\nThese are serious gaps for attribution, but the empirical result itself—a supervised residual network can correct ABPI's large errors on real hardware—is plausible and useful. The paper is coherent on its own terms; it just lacks the right controls.\n\nMy recommendation: this deserves peer review, not desk rejection. A good referee should push for the supervised baseline, the ablation, and a fix of Table II. The first hardware validation of ABPI alone makes it worth engaging.","headline":"First hardware validation of ABPI is a real contribution, but the physics-informed advantage is unproven due to a supervised-vs-unsupervised confound.","tokens_in":10739,"tokens_out":2538,"would_cite":false,"duration_ms":27531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper presents the first on-chip validation of ABPI and proposes CPINN-ABPI, a physics-informed neural network that cuts power-estimation error by 84.7% while keeping sub-millisecond inference.","keywords":["power estimation","physics-informed neural networks","multiprocessor systems-on-chip","blind power identification","thermal modeling","NSGA-II","NVIDIA Jetson AGX Xavier","heterogeneous SoC"],"falsifier":"Hold out a set of unseen workloads on the same Jetson AGX Xavier, keep the hardware's power sensors as ground truth, and compare CPINN-ABPI, ABPI, and a plain supervised network with the same labels and no physics branch. If CPINN-ABPI's WMAPE rises well above 12% or a plain network matches its MAE, the claim that the physics-informed correction is what delivers the gain is not supported.","tokens_in":9694,"feed_emoji":"⚡","tokens_out":7454,"duration_ms":68869,"temperature":0.7,"pith_summary":"This paper is trying to establish two things: that Alternative Blind Power Identification (ABPI), a leading method for estimating per-unit power from temperature alone, does not hold up on real hardware; and that a hybrid physics-informed neural network (CPINN-ABPI) can correct ABPI's errors well enough for practical use. The authors run the first empirical ABPI validation on a commercial MPSoC and report CPU mean absolute error falling from 3.74 W to 0.57 W and GPU from 3.74 W to 0.98 W, with weighted mean absolute percentage error dropping from 47-81% to about 12%. The result matters because unit-level power estimates are what enable predictive thermal management, dynamic voltage/frequency scaling, and thermal-attack detection on multicore chips.","feed_headline":"Physics-informed neural net cuts chip power-error by 84.7%","feed_subtitle":"On a Jetson AGX Xavier, the hybrid cuts CPU error 84.7% and GPU error 73.9% while running in 195 microseconds.","key_machinery":"The engine is a two-branch estimator: a physics branch implementing ABPI's thermal update $\\hat{t}=t_{\\mathrm{prev}}^T A$, $\\Delta t=t_{\\mathrm{current}}-\\hat{t}$, and $p_{\\mathrm{physics}}=\\Delta t\\cdot(B^{-1})^T$, with the thermal matrices $A,B$ initialized from ABPI and fine-tuned during training; and a residual network branch that maps $[t_{\\mathrm{prev}},t_{\\mathrm{current}},P_{\\mathrm{estimated}}]$ to a correction $\\Delta p$. The loss $L=L_{\\mathrm{data}}+\\lambda_{\\mathrm{phys}}L_{\\mathrm{phys}}+\\lambda_{\\mathrm{guide}}L_{\\mathrm{guide}}$ couples the branches, and NSGA-II optimizes layer count, width, activation, and the $\\lambda$ weights against the two objectives of low MAE and low MAC count.","core_discovery":"The central claim, stated on the paper's own terms, is that ABPI's linearized thermal model, $T_r(k)=AT_r(k-1)+BP(k)$, is the accuracy bottleneck on real MPSoCs: on a Jetson AGX Xavier it yields MAE of 3.74 W for both CPU and GPU and WMAPE of 47-81%, because it collapses the heat-diffusion equation into a simplified state-space update. The paper's remedy, CPINN-ABPI, keeps ABPI's physics branch but adds a parallel residual network that learns a correction $\\Delta p$, combined as $p_{\\mathrm{final}}=p_{\\mathrm{physics}}+\\Delta p$, with a three-term loss that fits ground-truth labels while enforcing thermal consistency and staying close to the physics estimate; NSGA-II selects the network size and loss weights. On the Xavier, this brings CPU MAE to 0.57 W (84.7% reduction) and GPU MAE to 0.98 W (73.9% reduction), keeps WMAPE near 12%, and runs in 195.3 $\\mu$s per inference; on a simulated six-component heterogeneous SoC the MAE improvements range from 85% to 99%.","pith_inferences":["The reported gains are purchased with per-unit ground-truth power labels for training; if such labels do not exist on a target platform, the method's advantage is not established, so deployment hinges on label availability.","The paper compares against unsupervised ABPI; a plain supervised neural network trained on the same labels might capture much of the same accuracy, leaving open how much the physics branch specifically contributes.","The residual-correction recipe is general: any differentiable physics-based power/thermal estimator could be wrapped with the same three-term loss and NSGA-II tuning, so the approach may extend beyond ABPI.","Cross-platform transfer is tested only as initialization of $A$ and $B$ from ABPI on the same platform; whether the learned correction transfers across workloads or chip generations is not addressed."],"forward_implications":["Unit-level power estimates on Jetson-class SoCs become accurate enough for real-time DVFS, thermal management, and thermal-Trojan detection, not just coarse server-level accounting.","ABPI's linear state-space thermal model is identified as the decisive source of error; correcting it with a learned residual is sufficient to cut MAE by 73.9-99% depending on the unit.","The approach keeps ABPI's key property: it still requires only consecutive temperature measurements and total power, not steady-state temperature.","The NSGA-II Pareto selection makes the accuracy-latency tradeoff explicit; the chosen Jetson model costs only 176 MACs per inference, which is why sub-millisecond operation is preserved.","The same recipe transfers to a simulated big.LITTLE heterogeneous SoC, where WMAPE becomes stable at low double-digit to tens of percent across all six components."],"supporting_citations":[{"why":"Defines the blind power identification thermal model whose linearized state-space form is the baseline that CPINN-ABPI corrects.","marker":"[5]"},{"why":"Introduces ABPI, the method this paper validates on hardware and then extends with the neural residual branch.","marker":"[17]"},{"why":"Documents the Jetson AGX Xavier platform used for the hardware validation.","marker":"[19]"},{"why":"Supplies the physics-informed neural network training paradigm that motivates the custom loss.","marker":"[20]"},{"why":"Provides the NSGA-II algorithm used to optimize accuracy versus computational cost.","marker":"[21]"},{"why":"Connects NSGA-II to PINN training, supporting the multiobjective tuning of the model.","marker":"[22]"},{"why":"Provides the CPU stress workloads used on the Jetson board.","marker":"[24]"},{"why":"Provides the system monitoring tool used to validate workload behavior during trace collection.","marker":"[25]"},{"why":"Supplies the thermal traces for the simulated heterogeneous SoC evaluation.","marker":"[27]"},{"why":"Supplies the ground-truth power values for the simulated heterogeneous SoC.","marker":"[28]"}],"fun_headline_variants":["Real-hardware test exposes ABPI flaws; PINN fix cuts errors by 85%","First on-chip ABPI test fails; new PINN method cuts error 84.7%","Physics-informed NN overtakes ABPI on Jetson: ~85% less error","ABPI's real-world accuracy flaw fixed by CPINN: 84.7% error drop","New hybrid neural-physics model slashes MPSoC power error to 12%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"CPINN-ABPI is trained with per-unit ground-truth power labels, while the ABPI baseline is unsupervised; the comparison counts as a fair demonstration only if such labels are available in the settings where unit-level power estimation is needed.","fun_headline_variants_meta":{"raw":{"variants":["Real-hardware test exposes ABPI flaws; PINN fix cuts errors by 85%","First on-chip ABPI test fails; new PINN method cuts error 84.7%","Physics-informed NN overtakes ABPI on Jetson: ~85% less error","ABPI's real-world accuracy flaw fixed by CPINN: 84.7% error drop","New hybrid neural-physics model slashes MPSoC power error to 12%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000896,"raw_usage":{"total_tokens":3931,"prompt_tokens":1086,"completion_tokens":2845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":702,"completion_tokens_details":{"reasoning_tokens":2729}},"tokens_in":702,"tokens_out":2845,"duration_ms":21081,"temperature":1.0,"reasoning_tokens":2729,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:06:37.536223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a set of unseen workloads on the same Jetson AGX Xavier, keep the hardware's power sensors as ground truth, and compare CPINN-ABPI, ABPI, and a plain supervised network with the same labels and no physics branch. If CPINN-ABPI's WMAPE rises well above 12% or a plain network matches its MAE, the claim that the physics-informed correction is what delivers the gain is not supported.","supporting_citations":[{"cited_title":"Blind identification of thermal models and power sources from thermal measurements,","cited_arxiv_id":null,"evidence_quote":"Defines the blind power identification thermal model whose linearized state-space form is the baseline that CPINN-ABPI corrects."},{"cited_title":"Alternating blind identification of power sources for mobile socs,","cited_arxiv_id":null,"evidence_quote":"Introduces ABPI, the method this paper validates on hardware and then extends with the neural residual branch."},{"cited_title":"Nvidia jetson agx xavier: Embedded system for autonomous machines,","cited_arxiv_id":null,"evidence_quote":"Documents the Jetson AGX Xavier platform used for the hardware validation."},{"cited_title":"Nsga-pinn: A multi-objective optimization method for physics-informed neural network training,","cited_arxiv_id":null,"evidence_quote":"Connects NSGA-II to PINN training, supporting the multiobjective tuning of the model."},{"cited_title":"stress-ng,","cited_arxiv_id":null,"evidence_quote":"Provides the CPU stress workloads used on the Jetson board."},{"cited_title":"Jetson stats tool (jtop),","cited_arxiv_id":null,"evidence_quote":"Provides the system monitoring tool used to validate workload behavior during trace collection."},{"cited_title":"Hotspot: A temperature modeling tool for ics,","cited_arxiv_id":null,"evidence_quote":"Supplies the thermal traces for the simulated heterogeneous SoC evaluation."}],"review_version":1}