Skip to main content

You are viewing a project that is currently in draft state for the Standards Working Group in the Green Software Foundation. This project should not be considered finished or officially supported in any way by the Green Software Foundation or its members.

Indirect Operational Water

Indirect Operational Water

In short: indirect operational water (W_O^ind) is water used generating the electricity a software system consumes — at the power plant, not at the compute facility. It follows exactly the same shape as direct operational water — it's just measuring thirst at the power plant instead of the data centre.

Shape: energy used × water-per-unit-of-energy-generated (the Electricity Water Intensity Factor, or EWIF) = indirect operational water.

Worked: 300 kWh × 0.9 L/kWh = 270 L. (illustrative placeholder, not a real measurement — note this is the same energy figure used in the direct operational water example, just asked a different "how thirsty" question)

Precise notation:

W_O,b^ind = E × EWIF_b   (for each basis b: c = consumption, w = withdrawal)

where E is the energy consumed by the software system, in kWh, and EWIF_b is the Electricity Water Intensity Factor for the selected basis — litres of water consumed or withdrawn per kWh generated.

The EWIF principle​

The specification is deliberately principles-based here: EWIF_b should reflect the water consumption or withdrawal of the power plants generating the electricity, at their point of generation — not at the facility where the software runs, and not averaged across a broad, loosely-defined region if a more precise figure is available. The specification requires three things of EWIF_b:

  1. It matches the water accounting basis (c or w) used elsewhere in the calculation.
  2. It corresponds to the geographic region, electricity supply, and reporting period of the energy actually consumed.
  3. Its temporal resolution is disclosed — if only annual data is available, that has to be said, not silently treated as equivalent to monthly data.

Everything below that principle — which datasets to use, how to handle generation types with unusual methodological wrinkles, whether to use average or marginal intensity — is genuinely implementation detail, which is exactly why it lives here rather than in the specification. See:


Please submit any comments you have here.