ML-EcoLyzer: Machine Learning Environmental Impact Analysis Framework

ML-EcoLyzer is a reproducible benchmarking and analysis framework for quantifying the environmental cost of machine learning inference. It supports modern transformers, vision models, and classical ML pipelines, adaptable to both edge and datacenter-scale deployments.

---

Environmental profiling across tasks, models, and hardware tiers.

---

🌍 Key Features

• Inference-Centric Analysis: Quantifies CO₂ emissions, energy use, and water impact from real-time inference
• Cross-Hardware Profiling: Supports A100, T4, RTX, GTX, and CPU-only setups
• Model-Agnostic Framework: Runs LLMs, ViTs, audio models, and traditional ML
• ESS Metric: Introduces the Environmental Sustainability Score for normalized emissions comparison
• Quantization Insights: Analyzes FP16 and INT8 savings for sustainable deployment
• Frequency-Aware Monitoring: Adjusts sampling dynamically for short and long-running workloads
• Lightweight and Extensible: Runs on mobile, edge, and low-resource devices

---

📊 What It Measures

✅ CO₂ Emissions

• Based on PUE, regional carbon intensity, and power consumption
• Adaptive to cloud, desktop, or edge scenarios

✅ Energy Usage

• Instantaneous power profiling via NVIDIA-SMI, psutil, or RAPL
• Sample-level granularity for each inference configuration

✅ Water Footprint

• Derived from power-to-water coefficients by tier (e.g., datacenter vs. mobile)

✅ Environmental Sustainability Score (ESS) Metric

$$\text{ESS} = \frac{\text{Effective Parameters (M)}}{\text{CO₂ (g)}}$$

• A normalized environmental efficiency metric for sustainable ML comparisons

---

🔧 Installation

pip install ml-ecolyzer

---

🚀 Quick Example

from mlecolyzer import EcoLyzer

config = {
    "project": "sustainability_demo",
    "models": [{"name": "gpt2", "task": "text"}],
    "datasets": [{"name": "wikitext", "task": "text"}]
}

eco = EcoLyzer(config)
results = eco.run()

print(f"CO₂: {results['final_report']['analysis_summary']['total_co2_emissions_kg']:.6f} kg")

---

📚 Scientific Foundation

Built on rigorously defined environmental assessment literature and standards:

• IEEE 754 (numeric precision)
• ASHRAE TC 9.9 (thermal/infra cooling)
• JEDEC JESD51 (thermal/power envelopes)
• Strubell et al. (2019), Patterson et al. (2021), Henderson et al. (2020), Lacoste et al. (2019)

---

🔬 Benchmark Coverage

• 1,500+ inference runs
• 4 hardware tiers: GTX 1650, RTX 4090, Tesla T4, A100
• Tasks: text, audio, vision, classification, regression
• Model families: GPT-2, OPT, Qwen, LLaMA, Phi, Whisper, ViT, etc.
• Precisions: FP32, FP16, INT8

---

🛠️ Configuration Template

project: "ml_sustainability_benchmark"

models:
  - name: "facebook/opt-2.7b"
    task: "text"
    quantization:
      enabled: true
      method: "dynamic"
      target_dtype: "int8"

datasets:
  - name: "wikitext"
    task: "text"
    limit: 1000

monitoring:
  frequency_hz: 5
  enable_quantization_analysis: true

hardware:
  device_profile: "auto"

output:
  export_formats: ["json", "csv"]
  output_dir: "./results"

---

🧪 Research Insights

Quantization Efficiency

INT8 models show up to 74% higher ESS than FP32 equivalents.

Hardware Utilization

A100 performs worst in ESS when underutilized; RTX/T4 yield better emissions-per-parameter for single-batch workloads.

Task-Wise Trends

Traditional models like SVC or Logistic Regression incur high ECEP due to small parameter count, despite low energy.

---

📜 Citation

@inproceedings{mlecolyzer2025,
  title={ML-EcoLyzer: Comprehensive Environmental Impact Analysis for Machine Learning Systems},
  author={Minoza, Jose Marie Antonio and Laylo, Rex Gregor and Villarin, Christian and Ibanez, Sebastian},
  booktitle={Proceedings of the Asian Conference on Machine Learning (ACML)},
  year={2025}
}

---

ML-EcoLyzer — Advancing sustainable inference in resource-constrained and production-scale deployments. 🌱