Carbon accounting in the Cloud: a methodology for allocating emissions across data center users

Overview

Google has made considerable efforts to reduce electricity consumption and the associated GHG emissions: by 2022 it delivered approximately three times as much computing power with the same electrical power as five years prior and uses 5.5 times less overhead energy for every unit of IT equipment energy than the industry average, yet it consumed 22 TWh of electricity in 2022. To reach its net-zero and 24/7 carbon-free energy (CFE) goals and better serve customers, it began in 2020 a substantial effort to develop granular data on hourly electricity consumption and carbon.

Machine idle power goes to whoever holds the resource allocation (dedicated owner, or a power-weighted blend of CPU/RAM/SSD/HDD allocations on shared machines). Dynamic power goes to real-time users in proportion to their CPU (GCU) usage. Energy used by shared internal services is passed on to their end users, by resource usage for large services and by internal net cost for small ones. Hourly emissions per user and cluster are CF = P (IT energy, MWh) × PUE (hourly, per cluster) × CI (hourly grid intensity of the Electricity Maps zone, kgCO2e/MWh); where no real-time data exists, IEA annual grid averages are used. Emissions are spread to SKUs in proportion to list price. Resources with equivalent busy power used to weight idle-power allocation: 1 Intel Skylake N1 CPU core = 20 GiB RAM = 1 TiB SSD = 6 TiB HDD.

On average an idle server draws about 45% of its peak power, which is about 60% of its average energy consumption; doubling utilisation from 40% to 80% raises power by only about 30%. Across all Electricity Maps zones and hours in 2022, the mean grid carbon intensity was 320.8 gCO2e/kWh with a standard deviation of 227.5 gCO2e/kWh. For Google Cloud in January-December 2023, market-based emissions were on average 63% lower than location-based emissions, with the ratio varying by region and time of use.

Scope 2 boundary exclusions: transmission and distribution losses, upstream fuel extraction and generation-facility life-cycle emissions, all Scope 1 (including refrigerant fugitive emissions), networking equipment outside data centres, and end-of-life emissions. Location-based customer metrics use hourly emission factors; the planned market-based metrics will use annual factors, to stay consistent with Google's company-wide annual Scope 2 accounting. Limits: operational Scope 2 only; the paper notes manufacturing (embodied) emissions will matter more as grids get cleaner but does not quantify them. It is a preprint.

Key points

1. On average an idle server draws about 45% of its peak power, which is about 60% of its average energy consumption; doubling utilisation from 40% to 80% raises power by only about 30%.
2. Hourly emissions are CF = P (IT energy, MWh) × PUE (hourly, per cluster) × CI (hourly grid intensity of the Electricity Maps zone, kgCO2e/MWh); where no real-time data exists, IEA annual grid averages are used.
3. Across all Electricity Maps zones and hours in 2022, the mean grid carbon intensity was 320.8 gCO2e/kWh with a standard deviation of 227.5 gCO2e/kWh.
4. For Google Cloud in January-December 2023, market-based emissions were on average 63% lower than location-based emissions, with the ratio varying by region and time of use.
5. Resources with equivalent busy power used to weight idle-power allocation: 1 Intel Skylake N1 CPU core = 20 GiB RAM = 1 TiB SSD = 6 TiB HDD.
6. Scope 2 boundary exclusions: transmission and distribution losses, upstream fuel extraction and generation-facility life-cycle emissions, all Scope 1 (including refrigerant fugitive emissions), networking equipment outside data centres, and end-of-life emissions.
7. Location-based customer metrics use hourly emission factors; the planned market-based metrics will use annual factors to stay consistent with company-wide annual Scope 2 accounting.

Using it in design

• In allocating carbon to data-centre users, give idle power to whoever holds the resource allocation and share dynamic power by actual usage; because an idle server draws about 45% of peak power, the idle share is large.
• Calculate emissions as IT energy × hourly PUE × hourly grid intensity and report location-based and market-based approaches separately (market-based about 63% lower).
• Document the Scope 2 boundary explicitly (transmission and distribution losses, Scope 1 and embodied emissions excluded) and add to the reporting plan that embodied emissions will matter more as grids get cleaner.

Who should read it

Sustainability and operations teams developing or choosing user-level carbon accounting methods for cloud and shared data centres.

Document contents

1. 1 Introduction
2. 2 Idle power allocation (dedicated and shared machines, power-weighted resource blend)
3. 3 Dynamic power allocation (GCU usage) and pass-through of shared internal services
4. 4 Hourly emissions calculation (IT energy, hourly PUE, Electricity Maps grid intensity) and spreading to SKUs
5. 5 Scope 2 boundary and market-based metrics
6. 6 Limitations and future work

Related standards

• GHG Protocol Scope 2 Guidance
• GHG Protocol Scope 3 Standard

Read the document

Publication details

Authors

Ian Schneider; Taylor Mattia

Publisher

arXiv preprint (cs.SE), Google

Year

2024

Document number

arXiv:2406.09645

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2406.09645

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation