Pith. sign in

REVIEW 4 major objections 4 minor 26 references

CARB: A Characterization-Guided Framework for CNN Inference Cost Prediction and Deployment Screening

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A single model predicts CNN energy, latency, and memory at R2 ≈ 0.99

desk verdict A credible R2~0.99 prediction study with a real, but addressable, validation gap in the screening workflow and baselines too weak to prove superiority. read the letter →

arxiv 2608.10506 v1 pith:FKCTE7I4 submitted 2026-08-11 cs.AR cs.LGcs.PF

Linh Nguyen , Zhixin Pan This is my paper · ORCID
classification cs.ARcs.LGcs.PF
keywords CNNinferencecostpredictionenergy-latencydivergenceGPUtelemetrycascadeensembledeploymentscreeningParetooptimizationpeakmemorycross-GPUtransferability
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 tries to establish that the three practical costs of running a CNN on a GPU—energy, latency, and peak memory—cannot be safely inferred from FLOPs or from each other, and that a data-driven ensemble can predict all three before deployment. From 13,419 configurations measured on two GPU platforms, it reports that energy and latency diverge by a factor of about 3 under heavy computational demand, and that energy and latency need platform-specific models while peak memory transfers across the tested GPUs. The proposed framework, CARB, cascades memory, energy, and latency predictions through a blended tree ensemble and corrects residuals in the low-batch regime, reaching $R^2 \approx 0.99$ for all three targets. If correct, this lets practitioners score thousands of design candidates in seconds and reserve hardware profiling for a Pareto-prioritized shortlist.

What carries the argument

The central object is CARB, a cascade-blended ensemble with three structurally diverse tree learners blended per target and chained so the memory prediction feeds the energy model, and the energy prediction feeds the latency model. Multiplicative interaction features (e.g., batch size times SM utilization) encode compounding hardware-load effects, and per-target residual correctors are trained only in the low-batch regime—where residual analysis shows structured error—with a 0.9 damping factor to avoid overcorrection. All targets are learned in log space. The cascade is the load-bearing mechanism: the paper reports that the upstream energy prediction ranks first among latency features, acting as a compressed summary of execution-time determinants.

What would settle it

Measure the same configurations on a third GPU (or on the same GPUs) with SM utilization sampled during evaluation-mode inference only, and retrain CARB; if the $R^2 \approx 0.99$ accuracy, the 3× energy-latency divergence, or the cross-GPU slope pattern (energy 1.61×, latency 2.09×, memory ≈1) does not reproduce, the characterization-guided design loses its empirical foundation.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that energy is not a shadow of latency or FLOPs: across batch-size tiers under high SM utilization, energy scales 35.4× while latency scales only 11.2×, so a latency-only proxy underestimates energy severalfold in deployment-relevant regimes. Cross-GPU analysis shows energy and latency follow non-unit transfer slopes (1.61× and 2.09×), whereas peak memory follows slope ≈1, meaning memory models transfer between the two tested platforms but energy and latency models must be per-GPU. CARB operationalizes this characterization: a cascade-blended ensemble predicts peak memory first, then energy using memory, then latency using both, achieving $R^2$ of 0.997, 0.993, and 0.992 on held-out test configurations. A two-stage screening workflow exploiting the near-monotonic ranking across GPUs reduces a 3,072-configuration space to seven Pareto-prioritized candidates, with 95.8% budget-classification accuracy and 2.1% false accepts at a 75 J threshold.

Load-bearing premise

The whole case rests on the assumption that SM utilization recorded during a training-mode forward-backward pass faithfully represents hardware load during evaluation-mode inference, because the interaction features and regime correctors are built on that proxy.

Editorial extensions

If this is right

  • A practitioner who profiles only latency will mis-budget energy by roughly 3× in high-load, large-batch regimes; energy must be predicted or measured directly.
  • Cross-GPU energy and latency estimates need a per-GPU model, while a single peak-memory model can be shared across the tested GPUs.
  • FLOPs alone cannot drive reliable cost screening: a FLOPs-only linear baseline achieves $R^2 < 0.38$ for all three targets.
  • Design spaces of thousands of CNN configurations can be reduced to a handful of Pareto candidates in under a second of scoring, before any hardware profiling.
  • Budget classification is reliable enough for pre-screening (95.8% accuracy, 2.1% false accepts), with borderline candidates flagged for measurement.

Reading between the lines

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

  • If the rank-preservation result (Spearman ρ = 0.95 between GPU rankings) holds beyond the two tested GPUs, CARB-style screening could order candidates for multi-device deployment from a single ranking model—an extension the paper does not test.
  • The telemetry-free mode's near-parity with full telemetry suggests static architectural features may sustain accurate cost prediction on unseen devices, but this is only shown within the same architecture family and two GPUs; extending to new block types or vendors is untested.
  • The low-batch residual asymmetry implies kernel-launch and driver overheads dominate small-batch energy; adding explicit launch-overhead or occupancy features could shrink the remaining error—a testable modification of the feature set.
  • The 3× energy-latency divergence under high SM utilization predicts that energy-aware NAS rankings will differ materially from latency-aware NAS in high-batch regimes; comparing the two Pareto fronts on a real deployment would test the practical consequence.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents a workload characterization of 13,419 ResNet-style CNN configurations on RTX 5090 and RTX 3080 GPUs under controlled clock settings, reporting that energy and latency diverge substantially under high computational demand (35.4× vs 11.2× as batch size grows) and that cross-GPU transferability is target-dependent (energy slope 1.61, latency slope 2.09, memory slope ≈0.92). It then proposes CARB, a cascade-blended ensemble of XGBoost, LightGBM, and ExtraTrees with regime-specific residual correctors, which jointly predicts peak memory, energy, and latency with held-out R² ≈ 0.99 across the three targets. Finally, it demonstrates a two-stage deployment screening workflow that reduces a 3,072-configuration grid to a seven-configuration Pareto shortlist for an RTX 3080 deployment budget, with three of three matched configurations receiving correct budget classifications on real hardware.

Significance. If the prediction and screening claims hold, CARB would be a practically useful tool for pre-deployment, energy-aware CNN selection, and the characterization results would strengthen the case against using FLOPs or latency as proxies for energy. The paper has several genuine strengths: a large and carefully collected dataset with clock locking, cooling intervals, and averaged measurements; a stratified split plus a leave-one-batch-tier-out evaluation (R² 0.956–0.991) showing generalization beyond interpolation; an ablation demonstrating that a telemetry-free mode is within 0.0012 R² of the full model; and an honest acknowledgment that the screening validation is re-run on measured feature vectors. However, the screening workflow is validated with a protocol mismatch—imputed telemetry in deployment versus measured telemetry in validation—and the hardware check covers only a narrow architectural corner. These gaps must be closed before the deployment-screening claims are fully credible.

major comments (4)
  1. [§7.1, §7.3, Table 9] The screening validation does not validate the screening protocol as run. Section 7.1 Step 1 fixes runtime telemetry features to training-set medians stratified by batch-size tier, so the 3,072 candidates are scored without GPU measurements. Section 7.3 then states that CARB is "re-run on actual feature vectors—not the synthetic architectural approximations used during screening" before comparing against the 75 J budget. The 3/3 correct classifications and 2.49 J mean absolute error therefore certify a model that receives real telemetry, not the deployed model that receives imputed medians. The Section 6.3 ablation compares retrained architecture-only and full-telemetry models and does not quantify how median imputation shifts predicted energies near the 75 J boundary or reorders the Pareto front. Please run the actual screening pipeline with imputed telemetry on the held-out RTX 3080 configurations and report budget-classification accuracy and the resulting shortlist, or provide an explicit sensitivity analysis of the 75 J decisions and Pareto ranking to the imputation.
  2. [§7.2, §7.3, Table 9] The hardware validation covers only one corner of the screened space. All three matched configurations in Table 9 are basic-block depth-14, batch size 1, FP32, and differ only in width multiplier. The candidate grid in Section 7.2 includes four block types, three precision formats, and four batch sizes, and the Pareto shortlist spans three unique architecture signatures with input-resolution and early-downsampling variants. The 3/3 success therefore does not support the claim that the seven-configuration shortlist is feasible, nor the broader statement that "screening decisions are reliable." Please validate at least one representative configuration from each architecture signature in the shortlist, or narrow the screening claim to the subspace actually validated.
  3. [§3.3, §4.3, Fig. 2] The SM utilization used to characterize energy–latency divergence is measured in a different execution mode from the targets. Section 3.3 records GPU utilization over a 2-second window during a training-mode forward-backward pass, while energy and latency are measured in evaluation mode. Figure 2 bins configurations by this training-mode SM utilization and uses the resulting high/low separation to support the 35.4× vs 11.2× divergence claim, and Section 5.1 builds interaction features (batch_x_sm, flops_x_sm) from the same quantity. If training-mode utilization does not represent inference hardware load, the characterization finding and the physical-grounding narrative of the cascade are unsupported. Please report eval-mode SM utilization for the same configurations and re-run the Figure 2 analysis; if the divergence changes materially, revise the characterization claims accordingly.
  4. [§4.5, Fig. 4] The cross-GPU transferability asymmetry is reported without uncertainty quantification. The energy fit (y = 1.61x + 52.8), latency fit (y = 2.09x − 2.7), and memory fit (y = 0.92x + 6.7) are presented as point estimates, with the claim that memory transfers "well" because its slope is ≈1. Since the design of Section 7 (shared memory model across GPUs, per-GPU models for energy and latency) rests on the distinction between slope > 1 and slope ≈ 1, please report confidence intervals for the slopes and intercepts, the number of configurations in each fit, and a rank-correlation or residual diagnostic to support the rank-preservation claim.
minor comments (4)
  1. [§6.1] The statement that feature importances "independently reproduce" the Section 4 findings is overstated, because the interaction features and cascade order in Section 5.1 were designed from those same findings; this is a confirmatory consistency check, not an independent replication. Please rephrase to reflect that.
  2. [§5.5, §6.5] The residual standard deviation analysis that identifies the low-batch regime as the shared axis of elevated error is described only verbally; please include the underlying analysis (e.g., residual SD by batch tier or a figure) that justifies restricting the corrector to batch size ≤ 8.
  3. [Table 2, §3.4] The notation M_param and M_activation is used without explicit definition; Section 3.4 defines activation memory but not the parameter-memory symbol. Please define both quantities in the text or in the table caption.
  4. [Figure 1] The secondary axis of Figure 1 is labeled "Latency / Memory," but the two quantities have different units (milliseconds versus megabytes); please use separate axes or explicit units to avoid conflating them.

Circularity Check

1 steps flagged · score 2.0 of 10

Central R2 claim is grounded in held-out hardware measurements and is not circular; the only self-referential burden is the interpretability claim that feature importances independently rediscover Section 4 findings, when the features and cascade were engineered from those same findings.

  1. self definitional [Section 6.1, final paragraph and Section 6.4]
    "These rankings independently reproduce the principal findings of Section 4 without any direct exposure to that analysis during training. ... The feature importance rankings provide the empirical validation of this design: pred_peak_memory_MB ranks third for energy prediction, and pred_energy_J ranks first for latency prediction."

    The claimed 'independent reproduction' is not independent because the model inputs and cascade structure were explicitly designed from Section 4. Section 5.1 constructs interaction features such as batch_x_sm and flops_x_sm specifically 'to capture compounding hardware-load effects', and Section 5.3 builds the cascade as 'The empirical dependency structure ... is captured via cascade prediction'. Observing that these engineered interaction features and cascade predictions appear high in feature importance is therefore a check that the model uses the features the authors put into it, not an independent rediscovery of the energy-latency divergence or memory transferability findings.

full rationale

CARB's headline predictive claims are not circular. The R2 values are evaluated on a stratified held-out test set (Section 5.2), and the leave-one-batch-tier-out experiment further isolates generalization. Energy, latency, and peak memory are measured on real hardware, and the cascade and residual correctors are trained on training-set targets and then evaluated on unseen measurements. The regression baselines (FLOPs-only and latency-as-energy proxy) are external comparisons. The screening validation, while it has an external-validity gap because Section 7.1 scores candidates with median-imputed telemetry and Section 7.3 re-runs CARB on actual feature vectors, is not a circularity: the measurements used for validation are independent of the fitted model. There are also no load-bearing self-citations; all references are to external prior work. The one self-referential element is the interpretability claim in Sections 6.1 and 6.4 that feature importances 'independently reproduce' Section 4 findings, when the interaction features and cascade order were engineered from those findings. This is a genuine but mild circularity in an auxiliary validation claim, not in the central prediction result, so the appropriate score is 2.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest entirely on the measured dataset and the fitted models. The free parameters are blend weights, the damping factor, hand-selected regime thresholds, and the cross-GPU transfer slopes. The axioms are domain assumptions about measurement fidelity, the representativeness of the search space, and the validity of the SM utilization proxy. No new physical entities are introduced.

free parameters (6)
  • Validation-calibrated specialist blend weights = memory (0.3, 0.4, 0.3); energy (0.3, 0.5, 0.2); latency (0.2, 0.3, 0.5)
    Grid-searched on the validation set (Table 5). The final predictions and reported R2 depend on these weights.
  • Residual corrector damping factor lambda = 0.9
    Chosen by hand in Equation 4 to prevent overcorrection. It directly scales the corrector contribution to the final predictions.
  • Low-batch regime cutoff = batch size <= 8
    Selected from residual standard deviation analysis (Sections 5.4 and 5.5). Residual correctors are trained only on this regime, which is a central design choice.
  • High-stress SM utilization threshold = avg_SM_util > 65%
    Used in the correctability analysis and in the stratified split definition. The threshold is hand-chosen and affects which residuals are considered learnable.
  • Cross-GPU linear transfer slopes = energy 1.61, latency 2.09, memory 0.92
    Fit to per-configuration scatter plots in Section 4.5. These slopes are the empirical basis for the claim that energy and latency need per-GPU models while memory transfers well, which directly shapes the framework design.
  • Stratification tier cutoffs = batch size tiers <=4, <=16, <=128; SM/mem >65%; depth tiers
    Hand-chosen cutoffs in the stratified 70/15/15 split (Section 5.2). They determine which configurations appear in train, validation, and test partitions, and hence the reported R2 values.
assumptions (5)
  • domain assumption NVML energy counter and PyTorch memory statistics accurately measure GPU inference energy and peak memory.
    All characterization and prediction targets are these measured quantities (Section 3.3). If the counter or memory stats are biased, the entire dataset and every downstream claim are affected.
  • domain assumption SM utilization measured during a training forward-backward pass is representative of inference hardware load.
    Invoked in Section 4.3 to define high computational demand and in Section 5.1 to create interaction features. The energy-latency divergence and the regime corrector design rely on this proxy.
  • domain assumption The ResNet-style search space is representative of practical CNN deployment candidates.
    Section 3.2 defines the space with two block types, depth, width, precision, batch size, input size, and early downsampling. The conclusions are framed generally despite this limited family.
  • domain assumption Stratified sampling with hand-chosen tier cutoffs produces unbiased train and test partitions.
    Section 5.2 relies on stratification keys to ensure edge cases are represented. The reported R2 values assume these partitions do not leak information across the boundary.
  • domain assumption Locked clocks, cooling intervals, and averaging over 100 runs remove measurement variance from the dataset.
    Section 3.1 describes these controls. The dataset is treated as clean ground truth for both characterization and training, so any residual thermal or DVFS effects are assumed negligible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CARB: A Characterization-Guided Framework for CNN Inference Cost Prediction and Deployment Screening." pith.science (2026). https://pith.science/paper/FKCTE7I4

@misc{pith2026260810506,
  author       = {Pith},
  title        = {Pith review of: CARB: A Characterization-Guided Framework for CNN Inference Cost Prediction and Deployment Screening},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKCTE7I4}},
  note         = {Machine review of arXiv:2608.10506}
}
read the original abstract

Accurate pre-deployment estimation of CNN inference cost--energy, latency, and peak memory--is increasingly critical as models are deployed on resource-constrained GPU platforms. Existing approaches rely on FLOPs, latency measurements, or single-device profiling as energy proxies, overlooking the non-linear interactions between architectural design and hardware load. We present a workload characterization study of 13 419 CNN configurations on two GPU platforms (RTX 5090 and RTX 3080) under GPU telemetry, revealing that energy, latency, and memory exhibit fundamentally distinct scaling behaviors: energy and latency diverge by 3x under high computational demand, and cross-GPU transferability differs by target--energy and latency require platform-specific models while memory transfers well across the two tested platforms. Building on these characterization findings, we develop CARB, a cascade-blended ensemble that jointly predicts all three targets with R2 ~0.99, and a two-stage deployment screening workflow that eliminates over 90% of candidates in seconds, reducing large design spaces to a Pareto-prioritized shortlist validated against real hardware.

Figures

Figures reproduced from arXiv: 2608.10506 by the authors.

Figure 1
Figure 1. Energy characterization across design knobs. Each panel shows the basic family (solid red line) and the bottleneck [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Mean energy and latency across batch-size tiers and GPU SM utilization levels (measured during training). The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Energy per GFLOP (J/GFLOP) across three dimensions. Left: by batch-size tier (log scale). Right two panels: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Per-configuration scatter: RTX 5090 (x-axis) vs. RTX 3080 (y-axis) for energy, latency, and memory. Points on the diagonal indicate identical cost. Linear fits characterize the cross-GPU scaling relationship for each target. therefore incurs a large per-GFLOP energy pe…
Figure 5
Figure 5. Figure 5: CARB framework overview. Left: both architectural and optional runtime features feed into all three cascade models (memory → energy → latency). Right: CARB predictions drive two-stage screening—Stage 1 (RTX 5090 ranking) shortlists candidates; Stage 2 (RTX 3080 budget …
Figure 6
Figure 6. Figure 6: Top-10 LightGBM feature importances (split count) per prediction target. Interaction features (f [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 16 canonical work pages

  1. [1]

    Maple: Microprocessor a priori for latency estimation

    Saad Abbasi, Alexander Wong, and Mohammad Javad Shafiee. Maple: Microprocessor a priori for latency estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2747–2756, 2022

    work page 2022

  2. [2]

    Neuralpower: Predict and deploy energy-efficient convolu- tional neural networks

    Ermao Cai, Da-Cheng Juan, Dimitrios Stamoulis, and Diana Mar- culescu. Neuralpower: Predict and deploy energy-efficient convolu- tional neural networks. InAsian Conference on Machine Learning, pages 622–637. PMLR, 2017

    work page 2017

  3. [3]

    Proxylessnas: Direct neural architecture search on target task and hardware.arXiv preprint arXiv:1812.00332, 2018

    Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware.arXiv preprint arXiv:1812.00332, 2018

    arXiv 2018

  4. [4]

    An analysis of deep neural network models for practical applications.arXiv preprint arXiv:1605.07678, 2016

    Alfredo Canziani, Adam Paszke, and Eugenio Culurciello. An analysis of deep neural network models for practical applications.arXiv preprint arXiv:1605.07678, 2016

    arXiv 2016

  5. [5]

    Energy-based tuning of convolu- tional neural networks on multi-gpus.Concurrency and Computation: Practice and Experience, 31(21):e4786, 2019

    Francisco M Castro, Nicolás Guil, Manuel J Marín-Jiménez, Jesús Pérez-Serrano, and Manuel Ujaldón. Energy-based tuning of convolu- tional neural networks on multi-gpus.Concurrency and Computation: Practice and Experience, 31(21):e4786, 2019

    work page 2019

  6. [6]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 785–794, New York, NY, USA, 2016. Association for Computing Machinery

    2016

  7. [7]

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE Transactions on Evolutionary Computation, 6(2):182–197, 2002

    2002

  8. [8]

    Extremely randomized trees.Machine Learning, 63(1):3–42, Mar 2006

    Pierre Geurts, Damien Ernst, and Louis Wehenkel. Extremely randomized trees.Machine Learning, 63(1):3–42, Mar 2006

    work page 2006

Show all 26 references
  1. [9]

    Dvfs- aware dnn inference on gpus: Latency modeling and performance analysis

    Yunchu Han, Zhaojun Nan, Sheng Zhou, and Zhisheng Niu. Dvfs- aware dnn inference on gpus: Latency modeling and performance analysis. InICC 2025-IEEE International Conference on Communica- tions, pages 1274–1279. IEEE, 2025

    2025

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  3. [11]

    Lightgbm: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in N...

    2017

  4. [12]

    Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25, 2012

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems, 25, 2012

    2012

  5. [13]

    Forecasting gpu performance for deep learning training and inference

    Seonho Lee, Amar Phanishayee, and Divya Mahajan. Forecasting gpu performance for deep learning training and inference. InProceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1, pages 493–508, 2025

    2025

  6. [14]

    fvcore: Facebook’s core library for computer vision research, 2020

    Meta AI Research. fvcore: Facebook’s core library for computer vision research, 2020

    2020

  7. [15]

    Pytorch: An imperative style, high- performance deep learning library.Advances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high- performance deep learning library.Advances in neural information processing systems, 32, 2019

    2019

  8. [16]

    Latenrgy: Model agnostic latency and en- ergy consumption prediction for binary classifiers.arXiv preprint arXiv:2412.19241, 2024

    Jason M Pittman. Latenrgy: Model agnostic latency and en- ergy consumption prediction for binary classifiers.arXiv preprint arXiv:2412.19241, 2024

    2024 arXiv

  9. [17]

    Fine- grained energy profiling for deep convolutional neural networks on the jetson tx1

    Crefeda Faviola Rodrigues, Graham Riley, and Mikel Luján. Fine- grained energy profiling for deep convolutional neural networks on the jetson tx1. In2017 IEEE International Symposium on Workload Characterization (IISWC), pages 114–115. IEEE, 2017

    2017

  10. [18]

    Memory requirements for convolutional neural network hardware accelerators

    Kevin Siu, Dylan Malone Stuart, Mostafa Mahmoud, and Andreas Moshovos. Memory requirements for convolutional neural network hardware accelerators. In2018 IEEE International Symposium on Workload Characterization (IISWC), pages 111–121. IEEE, 2018

    2018

  11. [19]

    Multi-target regression via input space expansion: treating targets as inputs.Machine Learning, 104(1):55–98, Feb 2016

    Eleftherios Spyromitros-Xioufis, Grigorios Tsoumakas, William Groves, and Ioannis Vlahavas. Multi-target regression via input space expansion: treating targets as inputs.Machine Learning, 104(1):55–98, Feb 2016

    2016

  12. [20]

    Hyperpower: Power-and memory-constrained hyper- parameter optimization for neural networks

    Dimitrios Stamoulis, Ermao Cai, Da-Cheng Juan, and Diana Mar- culescu. Hyperpower: Power-and memory-constrained hyper- parameter optimization for neural networks. In2018 Design, Automa- tion & Test in Europe Conference & Exhibition (DATE), pages 19–24. IEEE, 2018

    2018

  13. [21]

    Energy and policy considerations for deep learning in nlp

    Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp. InProceedings of the 57th annual meeting of the association for computational linguistics, pages 3645–3650, 2019

    2019

  14. [22]

    Characterizing deep learning training workloads on alibaba-pai

    Mengdi Wang, Chen Meng, Guoping Long, Chuan Wu, Jun Yang, Wei Lin, and Yangqing Jia. Characterizing deep learning training workloads on alibaba-pai. In2019 IEEE international symposium on workload characterization (IISWC), pages 189–202. IEEE, 2019

    2019

  15. [23]

    Roofline: an insightful visual performance model for multicore architectures

    Samuel Williams, Andrew Waterman, and David Patterson. Roofline: an insightful visual performance model for multicore architectures. Communications of the ACM, 52(4):65–76, 2009

    2009

  16. [24]

    Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search

    Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. InProceedings of the IEEE/CVF conference on comp...

    2019

  17. [25]

    Designing energy- efficient convolutional neural networks using energy-aware pruning

    Tien-Ju Yang, Yu-Hsin Chen, and Vivienne Sze. Designing energy- efficient convolutional neural networks using energy-aware pruning. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5687–5695, 2017

    2017

  18. [26]

    Nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices

    Li Lyna Zhang, Shihao Han, Jianyu Wei, Ningxin Zheng, Ting Cao, Yuqing Yang, and Yunxin Liu. Nn-meter: Towards accurate latency prediction of deep-learning model inference on diverse edge devices. InProceedings of the 19th Annual International Conference on Mobile Systems, App...

    2021

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.