Pith. sign in

REVIEW 4 major objections 4 minor 30 references

A Vertical Approach to Designing and Managing Sustainable Heterogeneous Edge Data Centers

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that vertical integration across the architecture, system, and runtime layers of an edge data center can minimize total carbon footprint while treating performance as a co-optimization objective.

desk verdict A clear vision paper that repackages the group's earlier single-layer results under a vertical-integration banner, but never tests the integration itself. read the letter →

arxiv 2506.01712 v1 pith:PORDUVDO submitted 2025-06-02 eess.SY cs.SY

Aikaterini Maria Panteleaki , Varatheepan Paramanayakam , Vasileios Pentsos , Andreas Karatzas , Spyros Tragoudas , Iraklis Anagnostopoulos This is my paper
classification eess.SYcs.SY
keywords edgedatacenterscarbon-awarecomputingembodiedcarbonoperationalverticalintegrationDNNacceleratordesignapproximateintensityscheduling
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 argues that the carbon footprint of edge data centers cannot be minimized by optimizing any single layer of the computing stack, because embodied emissions are fixed at hardware design time while operational emissions vary with energy use and grid carbon intensity. It proposes a vertical integration principle in which the architecture, system, and runtime layers exchange information, so hardware design choices, workload scheduling, and dynamic power and model-selection policies are decided together. The intended payoff is a design methodology that reduces both embodied and operational carbon while treating performance as an equal objective rather than a worst-case constraint. The reported evaluations are per-layer, and the paper's central claim is that these layers, when composed, yield combinations that layer-independent approaches would not achieve.

What carries the argument

The load-bearing mechanism is the vertical information flow across three optimization layers, anchored by two closed-form carbon models: embodied carbon as a function of die area and packaging, and operational carbon as $C_{\text{operational}} = CI \times E$, where $CI$ is grid carbon intensity and $E$ is energy. The Carbon Delay Product (CDP) is the compound objective that couples embodied footprint to latency at the architecture layer, while the transformer-based estimator and the carbon-intensity-aware power governor carry the coupling at the system and runtime layers. The approximate-multiplier library generated by gate-level pruning and bit-width reduction supplies the area-reduction knob that makes embodied-carbon savings compatible with maintained accuracy.

What would settle it

Run an end-to-end comparison on the same edge platform: one stack uses the paper's architecture, system, and runtime methods together with cross-layer information flow, and the other uses the same methods independently with no information shared between layers; if total embodied-plus-operational carbon over a 24-hour carbon-intensity cycle is statistically indistinguishable, the vertical-integration thesis is falsified.

Watch

Extended reading notes

Core claim

The central claim is that simultaneous, information-sharing optimization across the architecture, system, and runtime layers produces sustainable edge data center designs that no single-layer carbon-optimization method can reach. At the architecture layer, embodied carbon is modeled as a function of manufactured die area plus silicon wastage and packaging, and a genetic algorithm minimizes a Carbon Delay Product while optionally inserting approximate multipliers that trade small accuracy loss for area savings. At the system layer, a transformer-based estimator predicts throughput and power for fine-grained layer-level DNN mappings, and scheduling is throttled by 24-hour carbon-intensity forecasts. At the runtime layer, dynamic batching, concurrency, frequency scaling, and fallback to lighter or lower-precision models respond to real-time conditions. The paper reports per-layer gains: up to 65% embodied-carbon reduction from approximation, 30% operational-carbon reduction and 25% Carbon Delay Product improvement at the system level, and 52% operational-carbon reduction for LLM execution at the runtime level.

Load-bearing premise

The load-bearing premise is that the three layer frameworks, each validated separately, can be composed into a single system where cross-layer information flow yields carbon and performance gains beyond optimizing each layer on its own.

Editorial extensions

If this is right

  • If the vertical-integration thesis holds, embodied-carbon accounting can become a first-class hardware design objective, making chip area and approximation explicit levers alongside latency and energy.
  • Fine-grained layer-level scheduling with learned predictors could lower operational carbon without expensive online profiling, which would make the approach practical for resource-constrained edge servers.
  • Real-time carbon-intensity forecasts can be folded into runtime power management, with threshold-based updates, such as the paper's 10% change trigger, preventing oscillatory behavior.
  • The Carbon Delay Product objective gives a coherent way to compare carbon-performance trade-offs across heterogeneous edge platforms, potentially replacing constraint-based views of performance.
  • Composing the layers would allow an edge data center to adapt model quality, operating frequencies, and hardware mappings as grid carbon intensity shifts, reducing both embodied and operational emissions together.

Reading between the lines

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

  • A direct corollary of the paper's thesis that is not reported is an end-to-end experiment: build the full three-layer stack and compare its total embodied-plus-operational carbon against a baseline where each layer is optimized independently; the absence of such a comparison is the main open validation gap.
  • If the area-based embodied-carbon model is accurate, the same vertical logic could be extended to other physical design choices, such as 3D chiplet stacking, packaging technologies, and cooling, using carbon formulas analogous to the paper's TSV-aware extension.
  • The 10% carbon-intensity trigger suggests a general hysteresis rule for intermittent renewable integration, but the paper does not derive an optimal threshold; a sensitivity study varying that threshold would clarify the trade-off between responsiveness and stability.
  • The transformer-based estimator's classification formulation implies that the value of cross-layer information depends on how well predicted throughput and power distributions track real hardware behavior, so deployment drift could weaken the claimed synergy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a 'vertical integration' methodology for sustainable heterogeneous edge data centers, spanning architecture, system, and runtime layers. At the architecture layer, it advocates carbon-aware genetic-algorithm-based DNN accelerator design with approximate multipliers to reduce embodied carbon. At the system layer, it proposes a transformer-based multi-DNN scheduler that adapts power thresholds to forecasted carbon intensity to reduce operational carbon. At the runtime layer, it proposes dynamic scheduling for DNN inference and carbon-aware LLM function-calling controllers. The paper reports per-layer quantitative results (up to 65% embodied-carbon reduction, 30% and 52% operational-carbon reduction in separate experiments) and claims that information flow among the layers yields cross-layer decisions that layer-independent approaches cannot achieve. However, the evaluation in Section IV consists solely of isolated single-layer experiments drawn from prior work; no integrated experiment, combined carbon objective, or end-to-end simulation is presented.

Significance. If the vertical-integration claim were demonstrated, the paper would address a genuine gap in sustainable computing: most existing frameworks optimize a single layer and treat performance as a constraint rather than a co-optimization objective. The paper also usefully emphasizes direct embodied-carbon reduction rather than amortization, and it introduces a carbon-delay product as an optimization metric. These are valuable ideas. However, the manuscript as submitted provides no evidence for its central contribution. The quantitative claims are inherited from prior layer-specific papers, and the integrated cross-layer benefit is asserted without measurement, derivation, or sensitivity analysis. The paper therefore reads as a position or vision statement rather than a validated research contribution. No code, data, or machine-checked artifacts are provided.

major comments (4)
  1. [Section III and Section IV] The paper never tests the central claim that vertical integration across architecture, system, and runtime layers enables cross-layer decisions that minimize total carbon footprint. Figure 2 reports three independent experiments: (a) architecture-level accelerator design, (b) system-level scheduling, and (c) runtime-level LLM control. Each experiment uses its own baseline and metrics; no experiment runs the three layers together, and no equation in Section III shows how the output of one layer enters the optimization of another. Consequently, the statement in Section III.C that runtime results 'validate our vertical integration approach' is not supported by the evidence presented.
  2. [Section III.A-C and Section IV] The paper defines separate embodied-carbon and operational-carbon models (C_embodied and C_operational = CI × E) but never introduces or evaluates a combined total-carbon objective. It reports embodied-carbon reductions up to 65% and operational-carbon reductions of 30% and 52% in separate experiments, but no single workload is measured under the integrated framework. Without a total-carbon figure or a Pareto analysis over the two carbon components, the claim of minimizing total carbon footprint remains unquantified.
  3. [Section III.B and III.C] The load-bearing assumption that the three previously developed layer frameworks can be composed synergistically is asserted, not demonstrated. The text says the system layer 'receives detailed hardware characteristics from the architecture layer' and 'forwards the optimized configurations to the runtime layer,' but no interface specification, closed-loop mechanism, or end-to-end simulation is provided. The runtime results cited in Section III.C come from a standalone LLM scheduling framework [26] and are not connected to the approximate accelerators [18] or the MapFormer scheduler [23]. Since the evidence for each layer comes from prior layer-specific papers, the integrated benefit is attributed to the coexistence of independent results rather than to a measured interaction.
  4. [Section III.B] The claimed operational-carbon savings depend on two ad-hoc threshold parameters: the 10% carbon-intensity-change trigger and the <2% accuracy-drop tolerance. No sensitivity analysis is reported, so the paper does not show how the 30% and 52% reductions vary as these thresholds change. This makes the robustness of the central result impossible to assess.
minor comments (4)
  1. [Figure 2] The three subfigures show only aggregate averages without error bars or confidence intervals, and the axis labels are difficult to read; please provide detailed plots with variance information.
  2. [Section III.A] The equation for C_embodied is typeset as 'X Cdie + Cpackaging'; the summation symbol is garbled, and the units of CFPA and CFPASi are not defined.
  3. [Section III.A] The paper does not report the actual measured inference accuracy of the approximated accelerators; it only states 'typically < 2%', which is too vague to support the claim of negligible accuracy loss.
  4. [Table I] Table I lists 'This paper' with full marks for all three optimization levels, but the comparison is based on the cited prior works rather than on a direct experimental comparison; a column noting the evaluation status would clarify that the vertical integration is not yet measured.

Circularity Check

2 steps flagged · score 7.0 of 10

The central claim of vertical integration is validated only by re-citing the authors' own single-layer results; no cross-layer experiment, joint objective, or derivation connects the layers, so the integrated prediction reduces by construction to the sum of its self-cited inputs.

  1. self citation load bearing [Section III.C (Runtime-Level Optimizations), final paragraph]
    "These results validate our vertical integration approach by demonstrating how runtime adjustments build upon architecture and system-level optimizations to yield globally optimal sustainable solutions that layer-independent approaches would not achieve."

    The 'results' in this sentence are the runtime experiments of the authors' prior standalone LLM function-calling framework [26]; those experiments contain no architecture-layer approximate accelerators, no system-layer MapFormer scheduler, and no cross-layer information exchange. The conclusion that runtime adjustments 'build upon' the other two layers and produce a global optimum is therefore not derived from any integrated quantity; it is the single-layer result restated as validation of the integration. The claim reduces, by construction, to the presence of the standalone runtime framework plus the assertion of synergy.

  2. self citation load bearing [Section IV, Evaluation, first paragraph]
    "We evaluate our vertical integration approach by assessing the frameworks of each layer. The results demonstrate the effectiveness of our carbon-aware methodologies across architecture, system and runtime levels for the key sustainability and performance metrics."

    The evaluation never runs the layers together. Each quantitative result cited in this section comes from the authors' prior single-layer works ([16], [18], [21] for architecture; [23], [24] for system; [25], [26] for runtime). The paper's new contribution is the vertical integration, but its 'evaluation' is exactly the sum of the component frameworks, which are the inputs of the proposed composition. The integrated prediction of cross-layer decisions is equivalent to re-citing the separate layer results, with no joint objective, interface implementation, or end-to-end experiment connecting them.

full rationale

The component-level results are not themselves shown to be circular: each prior framework has its own baseline and benchmark, so the architecture, system, and runtime numbers can be taken at face value as single-layer evidence. The circularity is confined to the paper's central claim. Section III asserts that 'these three layers exchange information and work synergistically,' but the only support offered is Fig. 1 (a diagram) and the citation chain to the same group's prior papers [16,18,21,23-26]. Section III.C then states that standalone runtime results 'validate our vertical integration approach' and 'yield globally optimal sustainable solutions that layer-independent approaches would not achieve.' There is no equation or experiment in which outputs of one layer enter the optimization of another; there is no combined objective summing embodied and operational carbon for a single workload; and the evaluation section explicitly assesses 'the frameworks of each layer' separately. The claimed cross-layer advantage therefore reduces by construction to the existence of the three separate single-layer results, making the self-citation load-bearing for the paper's novelty claim. Score 7 reflects that the central claim, not the peripheral descriptions, is supported mainly by the authors' own prior results.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce new physical entities, but it depends on several domain assumptions about carbon modeling, forecast accuracy, and the composability of separately developed optimizations. The only hand-chosen thresholds appear in the system layer (10% CI trigger) and architecture layer (2% accuracy tolerance).

free parameters (2)
  • CI-change update threshold = 10%
    In Section III.B, updates to the maximum power threshold are triggered when CI changes exceed 10% of the estimated range. This threshold is chosen by hand, and no sensitivity analysis is provided.
  • Accuracy-drop tolerance = <2%
    In Section III.A, approximate multiplier configurations are selected subject to a typically less than 2% inference accuracy drop. The choice is a design constraint rather than a fitted value, but it affects the reported carbon savings.
assumptions (4)
  • domain assumption Embodied carbon of a die is proportional to chip area and can be modeled as C_die = CFPA*A_die + CFPASi*A_wasted
    Invoked in Section III.A as the basis for carbon-aware design space exploration; this model is taken from ACT [3] without verification in this paper.
  • domain assumption Carbon intensity forecasts with sufficient accuracy are available for the runtime scheduling decisions
    The runtime and system layers rely on 24-hour CI forecasts (Section III.B) and real-time CI adaptation; the accuracy and availability of these forecasts are assumed.
  • domain assumption Approximate multipliers with small accuracy degradation do not significantly affect DNN inference accuracy for the target models
    The architecture layer assumes error resilience of DNNs to substitute approximate multipliers, keeping accuracy loss below 2%, as stated in Section III.A.
  • ad hoc to paper The results from the individual layer frameworks can be combined additively or synergistically without conflicting constraints
    The paper's central claim of vertical integration assumes that architecture-level designs, system-level schedulers, and runtime controllers can be composed to yield combined benefits; no integrated experiment is presented to support this composition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Vertical Approach to Designing and Managing Sustainable Heterogeneous Edge Data Centers." pith.science (2026). https://pith.science/paper/PORDUVDO

@misc{pith2026250601712,
  author       = {Pith},
  title        = {Pith review of: A Vertical Approach to Designing and Managing Sustainable Heterogeneous Edge Data Centers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PORDUVDO}},
  note         = {Machine review of arXiv:2506.01712}
}
read the original abstract

The increasing demand for Artificial Intelligence (AI) computing poses significant environmental challenges, with both operational and embodied carbon emissions becoming major contributors. This paper presents a carbon-aware holistic methodology for designing and managing sustainable Edge Data Centers (EDCs), based on three design principles that challenge the state-of-the-art optimization paradigms. Our approach employs vertical integration across the architecture, system, and runtime layers, balances operational and embodied carbon emissions while considering EDC performance as a co-optimization objective, rather than a constraint. At the architecture level, we propose carbon-aware and approximate accelerator designs to reduce embodied carbon. At the system level, we enhance resource utilization and adapt to real-time carbon intensity variations to minimize operational emissions. Finally, at the runtime level, we develop dynamic scheduling frameworks that adjust execution, based on energy constraints and carbon intensity.

Figures

Figures reproduced from arXiv: 2506.01712 by the authors.

Figure 1
Figure 1. Overview of the vertical optimizations along with the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of Architecture, System, and Runtime-Level [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [26]

    CarbonCall: Sustainability-Aware Function Calling for Large Language Models on Edge Devices

    V . Paramanayakam et al. , “Carboncall: Sustainability-aware function calling for large language models on edge devices,” arXiv preprint arXiv:2504.20348, 2025

    work page Pith review arXiv 2025

  2. [18]

    Late breaking results: Leveraging approximate computing for carbon-aware dnn accelerators,

    A. M. Panteleaki et al. , “Late breaking results: Leveraging approximate computing for carbon-aware dnn accelerators,” in 2025 Design, Automa- tion & Test in Europe Conference (DATE)

    work page 2025

  3. [23]

    Mapformer: Attention-based multi- dnn manager for throughout & power co-optimization on embedded devices,

    A. Karatzas and I. Anagnostopoulos, “Mapformer: Attention-based multi- dnn manager for throughout & power co-optimization on embedded devices,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design , 2024

    work page 2024

  4. [1]

    Sustainable ai: Environmental implications, challenges and opportunities,

    C.-J. Wu et al. , “Sustainable ai: Environmental implications, challenges and opportunities,” Proceedings of Machine Learning and Systems , 2022

    work page 2022

  5. [2]

    Chasing carbon: The elusive environmental foot- print of computing,

    U. Gupta et al. , “Chasing carbon: The elusive environmental foot- print of computing,” in 2021 IEEE International Symposium on High- Performance Computer Architecture (HPCA) , 2021

    work page 2021

  6. [3]

    Act: Designing sustainable computer systems with an architectural carbon modeling tool,

    ——, “Act: Designing sustainable computer systems with an architectural carbon modeling tool,” in Proceedings of the 49th Annual International Symposium on Computer Architecture , 2022

    work page 2022

  7. [4]

    Offline energy-optimal llm serving: Workload-based energy models for llm inference on heterogeneous systems,

    G. Wilkins et al. , “Offline energy-optimal llm serving: Workload-based energy models for llm inference on heterogeneous systems,” ACM SIGENERGY Energy Informatics Review , 2024

    work page 2024

  8. [5]

    Clover: Toward sustainable ai with carbon-aware machine learning inference service,

    B. Li et al. , “Clover: Toward sustainable ai with carbon-aware machine learning inference service,” in International Conference for High Perfor- mance Computing, Networking, Storage and Analysis , 2023

    work page 2023

Show all 30 references
  1. [6]

    Greenllm: Disaggregating large language model serving on heterogeneous gpus for lower carbon emissions,

    T. Shi et al. , “Greenllm: Disaggregating large language model serving on heterogeneous gpus for lower carbon emissions,” arXiv preprint arXiv:2412.20322, 2024

    2024 arXiv

  2. [7]

    Greenscale: Carbon-aware systems for edge comput- ing,

    Y . G. Kim et al. , “Greenscale: Carbon-aware systems for edge comput- ing,” arXiv preprint arXiv:2304.00404 , 2023

    2023 arXiv

  3. [8]

    Ecoserve: Designing carbon-aware ai inference systems,

    Y . Li et al. , “Ecoserve: Designing carbon-aware ai inference systems,” arXiv preprint arXiv:2502.05043 , 2025

    2025 arXiv

  4. [9]

    Carbonedge: Leveraging mesoscale spatial carbon- intensity variations for low carbon edge computing,

    L. Wu et al. , “Carbonedge: Leveraging mesoscale spatial carbon- intensity variations for low carbon edge computing,” arXiv preprint arXiv:2502.14076, 2025

    2025

  5. [10]

    Carboncp: Carbon-aware dnn partitioning with con- formal prediction for sustainable edge intelligence,

    H. Ke et al. , “Carboncp: Carbon-aware dnn partitioning with con- formal prediction for sustainable edge intelligence,” arXiv preprint arXiv:2404.16970, 2024

    2024 arXiv

  6. [11]

    Exploring the efficiency of renewable energy-based modular data centers at scale,

    J. Sun et al., “Exploring the efficiency of renewable energy-based modular data centers at scale,” in Proceedings of the 2024 ACM Symposium on Cloud Computing , 2024

    2024

  7. [12]

    Carbon explorer: A holistic framework for designing carbon aware datacenters,

    B. Acun et al. , “Carbon explorer: A holistic framework for designing carbon aware datacenters,” in Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, V olume 2, 2023

    2023

  8. [13]

    Cordoba: Carbon-efficient optimization framework for computing systems,

    M. Elgamal et al. , “Cordoba: Carbon-efficient optimization framework for computing systems,” in 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) , 2025

    2025

  9. [14]

    Adaptive capacity provisioning for carbon-aware data centers: a digital twin-based approach,

    Z. Cao et al. , “Adaptive capacity provisioning for carbon-aware data centers: a digital twin-based approach,” IEEE Transactions on Sustainable Computing, 2025

    2025

  10. [15]

    Carbon and reliability-aware computing for heteroge- neous data centers,

    Y . Zhang et al. , “Carbon and reliability-aware computing for heteroge- neous data centers,” arXiv preprint arXiv:2504.00518 , 2025

    2025 arXiv

  11. [16]

    Carbon-aware design of dnn accelerators: Bridging performance and sustainability,

    A. M. Panteleaki and I. Anagnostopoulos, “Carbon-aware design of dnn accelerators: Bridging performance and sustainability,” in 2024 IEEE Computer Society Annual Symposium on VLSI (ISVLSI) , 2024

    2024

  12. [17]

    Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks,

    Y .-H. Chen et al., “Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks,” IEEE journal of solid-state circuits, 2016

    2016

  13. [19]

    Approxtrain: Fast simulation of approximate multipliers for dnn training and inference,

    J. Gong et al. , “Approxtrain: Fast simulation of approximate multipliers for dnn training and inference,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems , 2023

    2023

  14. [20]

    Three-dimensional stacked neural network accelerator architectures for ar/vr applications,

    L. Yang et al. , “Three-dimensional stacked neural network accelerator architectures for ar/vr applications,” IEEE Micro, 2022

    2022

  15. [21]

    Carbon-efficient 3d dnn acceleration: Optimizing performance and sustainability,

    A. M. Panteleaki et al., “Carbon-efficient 3d dnn acceleration: Optimizing performance and sustainability,” arXiv preprint arXiv:2504.09851 , 2025

    2025 arXiv

  16. [22]

    T. F. Wu et al. , “11.2 a 3d integrated prototype system-on-chip for augmented reality applications using face-to-face wafer bonded 7nm logic at < 2µm pitch with up to 40% energy reduction at iso-area footprint,” in 2024 IEEE International Solid-State Circuits Conference (ISSC...

    2024

  17. [24]

    Ecomap: Sustainability-driven optimiza- tion of multi-tenant dnn execution on edge servers,

    V . Paramanayakam et al. , “Ecomap: Sustainability-driven optimiza- tion of multi-tenant dnn execution on edge servers,” arXiv preprint arXiv:2503.04148, 2025

    2025 arXiv

  18. [25]

    Energy-aware dnn task scheduling with dynamic batch- ing and frequency adjustment,

    V . Pentsos et al., “Energy-aware dnn task scheduling with dynamic batch- ing and frequency adjustment,” in 2025 IEEE International Symposium on Quality Electronic Design (ISQED) , 2025

    2025

  19. [27]

    Digamma: Domain-aware genetic algorithm for hw- mapping co-optimization for dnn accelerators,

    S.-C. Kao et al. , “Digamma: Domain-aware genetic algorithm for hw- mapping co-optimization for dnn accelerators,” in 2022 Design, Automa- tion & Test in Europe Conference & Exhibition (DATE) , 2022

    2022

  20. [28]

    Omniboost: Boosting throughput of heterogeneous embedded devices under multi-dnn workload,

    A. Karatzas and I. Anagnostopoulos, “Omniboost: Boosting throughput of heterogeneous embedded devices under multi-dnn workload,” in 2023 60th ACM/IEEE Design Automation Conference (DAC) , 2023

    2023

  21. [29]

    Gorilla: Large language model connected with massive apis,

    S. G. Patil et al., “Gorilla: Large language model connected with massive apis,” Advances in Neural Information Processing Systems , 2024

    2024

  22. [30]

    Less is more: Optimizing function calling for llm execution on edge devices,

    V . Paramanayakam et al., “Less is more: Optimizing function calling for llm execution on edge devices,” in 2025 Design, Automation & Test in Europe Conference (DATE) . IEEE, 2025

    2025

Pith tools

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