Skip to content
GreenKube

Energy & Carbon

Carbon is reported next to cost on every workload, recommendation and measured change. This page explains where the energy and CO₂e figures come from. They are model-based estimates, not power-meter readings; every row records the provenance of each input so you can tell which parts are estimated. The formulas are part of the versioned Impact Model; the reference documents are docs/power_estimation_methodology.md and docs/methodology/impact-model.md.

For each node and collection step, GreenKube estimates power from three components:

node_cpu_cores = max(root cgroup CPU of the node, Σ pod CPU)
node_utilization = min(node_cpu_cores / node_vcores, 1)
cpu_power_watts = min_watts + node_utilization × (max_watts − min_watts)
memory_power_watts = node_memory_gib × 0.392 W/GiB
gpu_power_watts = Σ GPUs: measured DCGM board power, else a TDP estimate

Known instance types use the Cloud Carbon Footprint per-vCPU coefficients of their provider multiplied by the instance’s vCPU count:

ProviderMin W/vCPUMax W/vCPU
AWS0.743.50
GCP0.714.26
Azure0.783.76
OVH0.803.52
Scaleway0.602.78

OVH and Scaleway have no CCF defaults: their values are the average of the CPU micro-architectures they document, following CCF’s method for unknown micro-architectures. Each row records where the node profile came from (power_source):

  1. instance_profile: the instance type is in the bundled profile data;
  2. node_capacity_derived: the node’s CPU capacity × its provider’s per-vCPU coefficients (the average of the listed providers, 0.73–3.56 W per vCPU, when the provider is unknown); flagged as estimated;
  3. default_profile: no instance type and no capacity known; the configured default instance (1 vCPU, 1–10 W) is used, flagged as estimated and counted in greenkube_unknown_power_profile_nodes.

CPU used outside pods (kubelet, container runtime, kernel) counts in dynamic power when the node’s root cgroup series exceeds the pods’ sum; that system share goes to the node’s idle row.

The Cloud Carbon Footprint memory coefficient, 0.392 W per GiB (impact.memoryWattsPerGiB), applied to the node’s memory capacity. When the capacity is unknown, only the reserved memory is counted.

  • With the NVIDIA DCGM exporter, GreenKube reads the measured board power (DCGM_FI_DEV_POWER_USAGE) and utilization of each GPU per step. A GPU attributed to a pod is charged to that pod; unattributed GPUs go to the node’s idle row. When DCGM reports utilization but no power, power is idle + util × (TDP − idle) from bundled board-power profiles (idle is 15% of TDP).
  • Without DCGM, GPUs declared in the node capacity (nvidia.com/gpu, amd.com/gpu, gpu.intel.com/*) are detected. Each GPU a pod requests is charged to it at the TDP formula with an assumed 50% utilization (config.power.gpuAssumedUtilizationPct); other GPUs count at the idle floor on the idle row. GPUs of an unknown model are flagged “GPU power not measured”.

Rows carry gpu_source (dcgm, gpu_tdp_estimate, unmeasured), and the Footprint nodes page badges nodes whose GPU power is estimated or not measured.

Nodes on the same physical host (same bootID, or kind:// and k3d:// provider IDs) are treated as one host: each of N nodes gets 1/N of the host’s idle power, memory power and embodied emissions, with a “Shared host” estimation reason. Opt out with config.power.sharedHostDetection: false.

Each workload reserves max(request, usage) of CPU and memory. The node’s power and emissions are split as follows:

QuantitySplit
Idle CPU power (min_watts)By CPU reservation share
Dynamic CPU powerBy CPU usage share
Memory powerBy memory reservation share
GPU powerTo the pod DCGM (or the GPU request) attributes it to
Embodied emissionsBy CPU reservation share
Operational CO₂eEnergy × PUE × grid intensity, same split as the energy

What no pod reserves (unreserved cores and memory, GPUs without a pod, system CPU) is written to an explicit __idle__ row per node and step, so pod rows plus the idle row add up to the node totals for energy, operational and embodied CO₂e. Idle power and embodied emissions follow reservations, like cost, so a request cut lowers a workload’s allocated CO₂e the same way it lowers its allocated cost.

The Footprint page and GET /api/v1/metrics/summary, /metrics/timeseries and /report/summary show the allocated totals (workloads), the node idle share and the cluster total (allocated + idle) side by side, never mixed. See Footprint.

Manufacturing emissions are amortized per node-hour over a 4-year hardware lifespan: M_h = GWP × 1000 / (lifespan_years × 8760) gCO₂e per node-hour. The node’s GWP is resolved in this order and recorded as embodied_source:

  1. configured: the node annotation greenkube.io/embodied-kg (whole-node kg CO₂e, for on-premises hardware), then config.boavizta.embodiedKgByInstanceType;
  2. boavizta: the lifetime embedded GWP of the cloud instance type from the Boavizta API;
  3. fallback_per_vcpu: 12 kg CO₂e per vCPU (the Boavizta m5 family ratio, config.boavizta.embodiedKgPerVcpu) × cores, plus 0 kg per GiB by default (embodiedKgPerGib);
  4. fallback_default: 100 kg when the node’s capacity is unknown.

Rows of the two fallbacks are flagged as estimated (“Embodied emissions estimated”). A failed Boavizta lookup is retried at most every 24 hours.

Operational CO₂e is energy_kWh × PUE × grid intensity of the node’s electricity zone and hour. The provider is config.electricityProvider:

  • Electricity Maps (default), with a token (secrets.electricityMapsToken) for live values;
  • Wattnet, free and EU-funded, for 52 European zones at 15-minute granularity.

Each row records where its intensity came from (intensity_source):

ValueMeaning
consolidatedFinal provider value no older than INTENSITY_MAX_AGE (6 hours by default)
realtimeProvisional provider value; the hourly recompute replaces it once the provider consolidates it
zone_defaultNo fresh value: the bundled annual mean of the zone
country_defaultThe zone (a sub-zone such as US-CAL-CISO) has no bundled value: its country’s annual mean
staleOnly an older value exists and neither the zone nor its country has a default
global_defaultNothing else: config.defaultIntensity (500 gCO₂e/kWh by default)

The bundled defaults are the annual mean intensity of each country for 2024, with life-cycle emission factors, from Ember’s yearly electricity data (CC BY 4.0). An annual mean ignores hour-to-hour variation: set an Electricity Maps or Wattnet token for hourly values. Defaults are never written into the intensity history, so a missing token or a provider outage cannot pass for measured data.

On-premises, label nodes with their zone (topology.kubernetes.io/zone=FR) or set config.defaultZone, so they get their country’s value rather than the global default.

Power Usage Effectiveness multiplies IT energy to account for the data center’s cooling and power distribution. Each node’s PUE comes from its provider’s profile; providers without a specific value, and on-premises nodes (config.cloudProvider: on-prem), use config.defaults.pue (1.3 by default). On-premises, set it to the measured PUE of your site.

ScopeFieldContent
Scope 2 (location-based estimate)co2e_gramsGrid intensity × energy × PUE
Scope 3, category 1embodied_co2e_gramsHardware manufacturing emissions, amortized and allocated by CPU reservation

Scope 1 does not apply to Kubernetes workloads and is not tracked. Exports keep co2e_grams and embodied_co2e_grams in separate columns.

InputMeasuredEstimated
CPU usage, memory usage, requestsFrom Prometheus (cAdvisor, kube-state-metrics)—
CPU power—CCF power profiles × utilization (power_source)
Memory power—0.392 W per GiB
GPU powerDCGM board power (gpu_source: dcgm)TDP profile (gpu_tdp_estimate), or not counted (unmeasured)
Grid intensityProvider value (consolidated, realtime)Annual means and the global default
Embodied emissions—Operator value, Boavizta, or a per-vCPU fallback (embodied_source)
PUE—Provider profile or configured default

Rows built with any estimated input carry is_estimated and the reasons.

  • Modelled, not metered. Energy comes from power profiles, a constant memory power per GiB and, when available, DCGM GPU power. Storage and network energy are not modelled; orphaned PersistentVolume and LoadBalancer deletions project no CO₂e. Kepler-measured node power is on the roadmap.
  • Annual means hide hourly swings. Without a provider token, intensity is a yearly country mean.
  • Fixed lifespan. Embodied emissions use a 4-year lifespan for every node.
  • Not an inventory. Carbon figures are workload-level estimates for decisions, not a complete organizational inventory or compliance report.

See also Limitations.