A preprint analysis of climate-model computing finds that the reported footprint changes sharply depending on what the accounting boundary includes. For one 5 km Destination Earth workload, operational accounting recorded 10.3 MWh of energy and 1,501.7 kg of carbon-dioxide equivalent per simulated year. When the calculation counted active-node work only, the figures were 5.81 MWh and 847.7 kg per simulated year.
The study applies these alternative boundaries to the same workload, so the lower active-only figure reflects a narrower count rather than a second simulation with different performance. Under the paper's comparison, active-only accounting is reported as a 43.5% reduction in energy and emissions compared with operational accounting.
A large reported total, with incomplete coverage
The broader CMIP6 comparison puts the scale in a different frame. Across the eight institutions with complete data, the reported total operational carbon footprint reached 1,692 tonnes of CO2, out of 45 participating institutions. Because the complete-data set covered only that subset, the number describes the reports available for those institutions rather than a total for every center.
That total uses the study's operational convention for turning electricity into emissions. The calculation multiplies job energy in kilowatt-hours by the facility's Power Usage Effectiveness, or PUE, and by a regional grid emission factor expressed as kilograms of carbon-dioxide equivalent per kilowatt-hour. PUE accounts for the facility in the energy chain, while the emissions factor reflects the region's grid.
The case study behind the numbers
The workload was a 5 km simulation configured to use 475 compute nodes and reach a throughput of 1 simulated year per day, or 1 SYPD. To estimate operational power, the analysis used an empirical average draw of 900 watts per node, measured across more than 2,000 baseline runs. On that basis, a simulated year came to 10.3 MWh and 1,501.7 kg of carbon-dioxide equivalent.
Active-only accounting used a 390-watt idle-node baseline and a 510-watt active-node draw. It assigned 5.81 MWh and 847.7 kg of carbon-dioxide equivalent to a simulated year, both lower than the operational figures. The lower result should not be read as evidence that the workload itself consumed less energy; it reflects a narrower boundary that leaves baseline idle power outside the active calculation.
Hardware has a footprint too
Life-cycle accounting widened the boundary beyond electricity used during the run. The embodied-carbon allocation used a reported 22,614 tonnes of carbon-dioxide equivalent for a 7,600-node system, assuming a five-year hardware lifecycle and 80% utilization. Allocated to the workload, manufacturing emissions added 967.4 kg of carbon-dioxide equivalent per simulated year. The total life-cycle footprint reached 2,469.1 kg per simulated year, or 1.64 times the operational carbon figure.
The same accounting exercise put Total Cost of Ownership, or TCO, at approximately €0.76 per node-hour. The workload used 11,400 node-hours per simulated year, producing a reported cost of €8,664 per simulated year. The cost measure sits alongside the carbon totals, allowing energy, emissions and financial costs to be compared together.
A proposed reporting standard
For CMIP7, the framework prioritizes dynamic active-only energy measured through hardware counters or scheduler logs. The purpose is to separate power drawn by the workload from the cluster's baseline idle power, giving the accounting a clear distinction between activity and availability.
A case study, not a universal yardstick
The detailed workload results come from the specified 5 km configuration, while the CMIP6 total is based on complete reports from eight of 45 institutions. The study therefore maps how accounting choices affect these reported figures; it does not establish a single universal footprint for every climate-model workload or center.
The analysis is a preprint, and its front matter identifies arXiv version 2 dated 28 August 2026. It reports EuroHPC JU and European Union funding through ESiWACE3 under grant agreement No. 101093054, along with support from the European Commission's DestinE initiative.
Paper data and sources
Original title: Energy and CO2 Footprint of Climate Model Intercomparison Projects
Authors: Sergi Palomas, Pablo Aparici, Gladys Utrera, Mario Acosta
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text