Energy Estimation Methodology

GreenKube estimates per-pod energy consumption using a methodology based on the Cloud Carbon Footprint (CCF) project, adapted for real-time Kubernetes workload monitoring.

Overview

The energy estimation pipeline converts CPU utilization metrics into energy consumption (Joules) using known power characteristics of cloud instance types.

Input Metrics

📈 CPU Usage cores

💾 Instance Type metadata

🖥️ Node TDP watts/vCPU

Power Model

P_idle = min_watts × vCPUs

P_max = max_watts × vCPUs

P_node = P_idle + (P_max − P_idle) × U

U = CPU utilization (0.0 – 1.0)

Energy

E_node = P_node × Δt

Joules = Watts × seconds

E_pod = ^cpu_pod⁄_cputotal × E_node

Proportional distribution

Carbon

CO₂e = E_kWh × I × PUE

E_kWh = Joules / 3,600,000

I = gCO₂e/kWh (Electricity Maps)

PUE = datacenter overhead

Power Model

Instance Power Profiles

Each cloud instance type has a power profile defining its energy characteristics:

Parameter	Description	Unit
min_watts	Power consumption at idle (per vCPU)	Watts
max_watts	Power consumption at 100% utilization (per vCPU)	Watts
vcpus	Number of virtual CPUs	count

GreenKube includes built-in profiles for common instance types across:

• AWS — m5, c5, r5, t3, etc.
• GCP — n1, n2, e2, c2, etc.
• Azure — Standard_D, Standard_E, Standard_F, etc.
• OVH — Based on averaged micro-architecture data
• Scaleway — Based on averaged micro-architecture data

Power Calculation Formula

For each node, the power consumption is calculated as:

P_node = P_idle + (P_max - P_idle) × U

Where:
  P_idle = min_watts × vCPUs        (idle power)
  P_max  = max_watts × vCPUs        (max power)
  U      = CPU utilization ratio     (0.0 to 1.0)

This is a linear interpolation between idle and max power, which is a widely-used approximation validated by SPECpower benchmarks.

Energy Calculation

Energy is the integral of power over time:

E_node = P_node × duration_seconds    (Joules)

Per-Pod Distribution

Node-level energy is distributed to pods proportionally by CPU usage:

E_pod = (cpu_pod / cpu_total_node) × E_node

Where:

• cpu_pod — CPU usage of the specific pod
• cpu_total_node — Total CPU usage across all pods on the node

Carbon Emission Calculation

Energy is converted to CO₂ equivalent emissions:

CO₂e (grams) = E (kWh) × I (gCO₂e/kWh) × PUE

Where:
  E (kWh) = E (Joules) / 3,600,000
  I        = Grid carbon intensity (from Electricity Maps or default)
  PUE      = Power Usage Effectiveness (datacenter overhead)

Grid Carbon Intensity

The carbon intensity of electricity varies by:

• Geographic location (carbon zone)
• Time of day (renewable energy availability)
• Season (heating/cooling demand)

GreenKube uses the Electricity Maps API for real-time intensity data. When the API is unavailable, GreenKube follows a two-step fallback:

1. Per-zone default — A built-in table of average grid intensities per zone (e.g., France ≈ 85 gCO₂e/kWh, Germany ≈ 385 gCO₂e/kWh)
2. Global default — If no per-zone default exists, the configurable global fallback is used (default: 500 gCO₂e/kWh)

PUE (Power Usage Effectiveness)

PUE accounts for datacenter overhead (cooling, lighting, networking):

PUE = Total facility power / IT equipment power

Provider	PUE	Source
AWS	1.15	AWS Sustainability
GCP	1.09	Google Data Centers
Azure	1.18	Microsoft Sustainability Report 2025
OVH	1.26	OVHcloud Sustainability
Scaleway	1.37	Scaleway Environmental Leadership

OVH and Scaleway Methodology

Since the Cloud Carbon Footprint project doesn’t provide power constants for OVH and Scaleway, GreenKube derives provider-level estimates using the CCF methodology for “unknown micro-architectures”:

“When we don’t know the underlying processor micro-architecture, we use the average of all micro-architectures used by that cloud provider.”

OVHcloud

Identified architectures: Intel Haswell, Broadwell, Skylake, Cascade Lake; AMD EPYC Milan, Genoa

Micro-architecture	Min Watts	Max Watts
Haswell	1.90	5.99
Broadwell	0.71	3.54
Skylake	0.64	4.05
Cascade Lake	0.64	3.80
AMD EPYC Milan	0.45	1.87
AMD EPYC Genoa*	0.45	1.87

Average: Min Watts = 0.80, Max Watts = 3.52

Scaleway

Identified architectures: AMD EPYC Naples, Rome, Milan; Intel Skylake, Cascade Lake

Micro-architecture	Min Watts	Max Watts
AMD EPYC 1st Gen	0.82	2.55
AMD EPYC 2nd Gen	0.47	1.64
AMD EPYC 3rd Gen	0.45	1.87
Intel Skylake	0.64	4.05
Intel Cascade Lake	0.64	3.80

Average: Min Watts = 0.60, Max Watts = 2.78

Note

Genoa values estimated using Milan as a proxy due to lack of specific SPECpower data.

Fallback Behavior

When GreenKube cannot determine the exact instance type, it uses a configurable default profile:

# Default values in Helm
config:
  defaults:
    pue: 1.3
    instance:
      vcores: 1
      minWatts: 1.0
      maxWatts: 10.0

When default values are used, the metric is flagged as is_estimated: true with the reason recorded in estimation_reasons.

Embodied Emissions

GreenKube also supports embodied emissions — the carbon footprint of manufacturing the hardware — via the Boavizta API:

Embodied CO₂e per pod = (Server GWP_manufacture / Lifespan_hours) × Duration × (pod_share / node_total)

This data is cached in the EmbodiedRepository to minimize API calls.

Estimation Transparency

Every metric in GreenKube includes transparency flags:

Field	Description
is_estimated	true if any default/fallback was used
estimation_reasons	List of reasons why estimation was needed

Example reasons:

• "Unknown instance type, using default profile"
• "Carbon intensity API unavailable, using per-zone default"
• "Carbon intensity API unavailable, using global default 500 gCO₂e/kWh"
• "OpenCost data unavailable, cost set to 0"

Data Sources

Source	Data	Frequency
Prometheus	CPU, memory, network, disk metrics	Real-time (5m default)
Kubernetes API	Node metadata, pod specs	Real-time
Electricity Maps	Grid carbon intensity	Cached (hourly default)
Boavizta API	Hardware embodied emissions	Cached (long-term)
OpenCost	Cost allocation	Real-time
CCF Coefficients	Instance power profiles	Built-in (static)

References

• Cloud Carbon Footprint Methodology
• SPECpower Benchmarks
• Electricity Maps API
• Boavizta API