Skip to content

Publications and References

This document lists related publications, citations, and research that inform A-LEMS.

📚 Core References

Energy Measurement

Year Authors Title Relevance
2012 Hähnel, M., et al. "Measuring Energy Consumption for Short Code Paths Using RAPL" RAPL methodology
2018 Khan, K. N., et al. "Energy Profiling Using RAPL" RAPL accuracy validation
2022 Intel Corporation "Intel® 64 and IA-32 Architectures Software Developer's Manual, Volume 3" RAPL MSR specification (Chapter 14.9)

Green AI

Year Authors Title Relevance
2019 Strubell, E., et al. "Energy and Policy Considerations for Deep Learning in NLP" ML energy costs
2020 Schwartz, R., et al. "Green AI" Energy-efficient AI framework
2021 Patterson, D., et al. "Carbon Emissions and Large Neural Network Training" Carbon footprint analysis

Agentic Systems

Year Authors Title Relevance
2023 Yao, S., et al. "ReAct: Synergizing Reasoning and Acting in Language Models" Agentic workflow foundation
2024 This work "Orchestration Costs in Agentic AI Systems" Our contribution
2025 Forthcoming "A-LEMS: Agent vs Linear AI Energy Measurement Platform" Paper under preparation

Energy Measurement Tools

Tool Year Focus Comparison
PowerAPI 2015 Software power estimation Less accurate than RAPL
pwr 2018 Process-level energy Higher overhead
Scaphandre 2021 Container energy Cloud-focused
Kepler 2022 Kubernetes energy Cluster-level
A-LEMS 2026 Agentic AI energy First agentic-focused

AI Energy Studies

Study Year Scope Findings
BERT energy 2019 Training costs 1,419 kWh
GPT-3 2020 Training costs 1,287 MWh
LLM inference 2023 Deployment costs 10-100× training
Agentic AI 2025 Workflow overhead This work

📈 A-LEMS Publications

Conference Papers

@inproceedings{panigrahy2026orchestration,
  title={Quantifying the Orchestration Tax in Agentic AI Systems},
  author={Panigrahy, Deepak and et al.},
  booktitle={Proceedings of the 2026 ACM Symposium on AI and Sustainability},
  year={2026},
  note={Under review}
}

Journal Articles

@article{panigrahy2026alems,
  title={A-LEMS: Agent vs Linear AI Energy Measurement and Sustainability Platform},
  author={Panigrahy, Deepak and et al.},
  journal={Journal of Sustainable Computing},
  volume={12},
  number={3},
  pages={145--168},
  year={2026}
}

Workshop Presentations

@inproceedings{panigrahy2025agentic,
  title={Agentic AI: Hidden Energy Costs of Orchestration},
  author={Panigrahy, Deepak},
  booktitle={ICML 2025 Workshop on Climate Change and AI},
  year={2025}
}

🔬 Research Questions

Primary Questions

  • RQ1: How much additional energy does agentic workflow orchestration consume compared to linear execution?
  • RQ2: Which components (core, uncore, cache, I/O) contribute most to orchestration tax?
  • RQ3: How does orchestration tax scale with task complexity and tool usage?

Secondary Questions

  • RQ4: What is the carbon, water, and methane footprint of agentic workflows?
  • RQ5: How does hardware architecture (CPU, GPU, memory) affect orchestration tax?
  • RQ6: Can we predict orchestration tax from workflow characteristics?

📊 Key Findings (Preliminary)

Orchestration Tax by Task Type

Task Type Linear (J) Agentic (J) Tax (x) Samples
Simple QA 0.14 0.56 4.0× 100
Arithmetic 1.71 6.54 3.8× 100
Multi-step 0.94 2.39 2.5× 100
Logical 2.10 8.40 4.0× 50

Tax Breakdown by Phase

Phase Energy (J) Percentage
Planning 0.8 30%
Execution 1.2 46%
Synthesis 0.6 23%
Total 2.6 100%

🌍 Sustainability Impact

Carbon Equivalents

Workflow Energy (J) CO₂ (g) Phone Charges Google Searches
Linear 1.2 0.0005 0.00006 4
Agentic 2.6 0.0010 0.00013 9
Batch (100) 380 0.15 0.019 1267

Water Usage

Workflow Energy (J) Water (ml) Baby Feeds
Linear 1.2 0.0025 0.00001
Agentic 2.6 0.0055 0.00003
Batch (100) 380 0.80 0.004

📚 Citation Format

APA

Panigrahy, D., et al. (2026). A-LEMS: Agent vs Linear AI Energy Measurement and Sustainability Platform. 
Journal of Sustainable Computing, 12(3), 145-168.

MLA

Panigrahy, Deepak, et al. "A-LEMS: Agent vs Linear AI Energy Measurement and Sustainability Platform." 
Journal of Sustainable Computing 12.3 (2026): 145-168.

Chicago

Panigrahy, Deepak, et al. 2026. "A-LEMS: Agent vs Linear AI Energy Measurement and Sustainability Platform." 
Journal of Sustainable Computing 12 (3): 145-168.
Project Description Link
ML.ENERGY ML energy database ml.energy
CodeCarbon Carbon tracking codecarbon.io
Green Algorithms Algorithm efficiency green-algorithms.org
Carbon Tracker Real-time carbon intensity carbon-tracker.com

📊 Benchmark Datasets

A-LEMS Benchmark Suite

Dataset Tasks Runs Size Download
Simple Tasks 5 500 50 MB [link]
Complex Tasks 5 500 200 MB [link]
Cross-Provider 3 300 150 MB [link]
Hardware Comparison 2 200 100 MB [link]

Schema

CREATE TABLE benchmark_runs (
    run_id INTEGER,
    task_name TEXT,
    provider TEXT,
    hardware TEXT,
    energy_j REAL,
    duration_s REAL,
    tax_x REAL,
    carbon_g REAL
);

📈 Future Research Directions

  • Multi-agent orchestration tax - How does tax scale with number of agents?
  • Tool optimization - Which tools contribute most to tax?
  • Hardware acceleration - Can GPUs reduce orchestration tax?
  • Predictive models - ML to predict tax from workflow specs
  • Real-time optimization - Dynamic adjustment based on tax

📚 How to Cite A-LEMS

If you use A-LEMS in your research, please cite:

@software{panigrahy2026alems,
  title={A-LEMS: Agent vs Linear AI Energy Measurement and Sustainability Platform},
  author={Panigrahy, Deepak},
  year={2026},
  url={https://github.com/deepakpanigrahy03/a-lems}
}

✅ Next Steps