{
  "name": "Research Computing Energy and Carbon Telemetry Frameworks",
  "description": "Comparative technical matrix of open-source research computing energy telemetry and carbon estimation frameworks, analyzing CPU, GPU, and RAM measurement mechanisms, default telemetry sampling rates, and geographic grid intensity support.",
  "page": "https://casrai.org/guides/codecarbon",
  "publisher": "CASRAI",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "dateModified": "2026-11-16T22:04:26Z",
  "temporalCoverage": "2021/2026",
  "spatialCoverage": "Global",
  "columns": [
    {
      "name": "Framework",
      "unit": "",
      "description": "Software library or tool name"
    },
    {
      "name": "Primary Runtime",
      "unit": "",
      "description": "Primary programming language or execution environment"
    },
    {
      "name": "CPU Telemetry Backend",
      "unit": "",
      "description": "Interface used for CPU energy measurement or modeling"
    },
    {
      "name": "GPU Telemetry Backend",
      "unit": "",
      "description": "Interface used for GPU device telemetry"
    },
    {
      "name": "RAM Accounting Method",
      "unit": "",
      "description": "Approach used for system memory energy accounting"
    },
    {
      "name": "Default Polling Interval",
      "unit": "seconds",
      "description": "Default background telemetry sampling frequency"
    },
    {
      "name": "Cloud Region Detection",
      "unit": "",
      "description": "Native discovery of AWS, GCP, and Azure cloud instance metadata"
    },
    {
      "name": "Offline Mode Support",
      "unit": "",
      "description": "Capability to calculate emissions without internet access via static tables"
    }
  ],
  "rows": [
    {
      "Framework": "CodeCarbon",
      "Primary Runtime": "Python",
      "CPU Telemetry Backend": "Intel RAPL / Windows EMI / powermetrics / TDP fallback",
      "GPU Telemetry Backend": "NVIDIA NVML / AMD SMI",
      "RAM Accounting Method": "Modelled 5W per DIMM (x86) / PMIC rails (RPi5)",
      "Default Polling Interval": 15,
      "Cloud Region Detection": "Supported (AWS, GCP, Azure)",
      "Offline Mode Support": "Supported (ISO country codes)"
    },
    {
      "Framework": "Eco2AI",
      "Primary Runtime": "Python",
      "CPU Telemetry Backend": "Intel RAPL / AMD Zen RAPL sysfs",
      "GPU Telemetry Backend": "NVIDIA NVML (pynvml)",
      "RAM Accounting Method": "Static memory model based on capacity",
      "Default Polling Interval": 10,
      "Cloud Region Detection": "Manual cloud region assignment",
      "Offline Mode Support": "Supported (Static regional tables)"
    },
    {
      "Framework": "CarbonTracker",
      "Primary Runtime": "Python",
      "CPU Telemetry Backend": "Intel RAPL sysfs",
      "GPU Telemetry Backend": "NVIDIA NVML (pynvml)",
      "RAM Accounting Method": "Excluded from default calculation",
      "Default Polling Interval": 10,
      "Cloud Region Detection": "Supported (Cloud provider mappings)",
      "Offline Mode Support": "Supported (Pre-fetched intensity files)"
    },
    {
      "Framework": "Green Algorithms",
      "Primary Runtime": "Python / R / Web",
      "CPU Telemetry Backend": "TDP and core-hour mathematical modeling",
      "GPU Telemetry Backend": "TDP and accelerator runtime model",
      "RAM Accounting Method": "Modelled at 0.375 W per GB RAM",
      "Default Polling Interval": 0,
      "Cloud Region Detection": "Supported (Major provider regions)",
      "Offline Mode Support": "Supported (Local script calculations)"
    },
    {
      "Framework": "Scaphandre",
      "Primary Runtime": "Rust",
      "CPU Telemetry Backend": "Linux RAPL powercap sysfs (/sys/class/powercap)",
      "GPU Telemetry Backend": "Excluded (Host CPU and memory only)",
      "RAM Accounting Method": "Linux RAPL DRAM domain counters",
      "Default Polling Interval": 1,
      "Cloud Region Detection": "Infrastructure agent level",
      "Offline Mode Support": "Supported (Local metrics daemon)"
    },
    {
      "Framework": "Experiment Impact Tracker",
      "Primary Runtime": "Python",
      "CPU Telemetry Backend": "Intel RAPL sysfs",
      "GPU Telemetry Backend": "NVIDIA NVML (pynvml)",
      "RAM Accounting Method": "Modelled memory energy",
      "Default Polling Interval": 15,
      "Cloud Region Detection": "Supported (Cloud region lookup)",
      "Offline Mode Support": "Supported (Static fallback tables)"
    }
  ],
  "sources": [
    {
      "title": "CodeCarbon Documentation: Methodology & Output Reference",
      "url": "https://docs.codecarbon.io",
      "accessed": "2026-11-16"
    },
    {
      "title": "Schmidt et al., CodeCarbon: Estimate and Track Carbon Emissions from Machine Learning Computing",
      "url": "https://doi.org/10.5281/zenodo.4658425",
      "accessed": "2026-11-16"
    },
    {
      "title": "Budenny et al., Eco2AI: Carbon Emission Tracking of Machine Learning Models",
      "url": "https://arxiv.org/abs/2208.00406",
      "accessed": "2026-11-16"
    },
    {
      "title": "Anthony et al., Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models",
      "url": "https://arxiv.org/abs/2007.03051",
      "accessed": "2026-11-16"
    },
    {
      "title": "Lannelongue et al., Green Algorithms: Quantifying the Carbon Footprint of Computation",
      "url": "https://doi.org/10.1002/advs.202100707",
      "accessed": "2026-11-16"
    },
    {
      "title": "Hubblo: Scaphandre Energy Consumption Metrology Agent",
      "url": "https://github.com/hubblo-org/scaphandre",
      "accessed": "2026-11-16"
    }
  ]
}