Carbon Explorer: A Holistic Framework for Designing Carbon Aware Datacenters

Overview

Carbon Explorer (ASPLOS 2023, pp. 118-132; arXiv:2201.10036) is an open-source design-space tool for reaching 24/7 hourly carbon-free operation. It is driven by hourly demand from 13 Meta US data-centre sites and hourly EIA grid wind/solar data. The tool co-sizes wind/solar investment, on-site lithium-iron-phosphate batteries and extra servers for carbon-aware scheduling of delay-tolerant work, minimising the sum of operational and embodied carbon.

Results: renewables alone show a long tail of diminishing returns; going from 95% to 99.9% hourly coverage takes more than 5x the investment needed for 0% to 95% (Oregon). Assuming average daily wind/solar output underestimates the investment for 100% coverage by an order of magnitude. Batteries cut the total footprint by an order of magnitude, and batteries plus scheduling cut it by 15-65% and make 100% coverage carbon-optimal in five regions. Nebraska, Utah and Texas were the best sites.

Once embodied carbon is counted, 100% coverage is not always the lowest-carbon design. Siting on wind or hybrid wind-solar grids matters a lot, and solar-only regions cannot get much beyond 50% hourly coverage without storage. Battery needs are about 5 hours in Utah (mixed, with existing renewables) and 14 hours in North Carolina (solar only). Results rest on Meta's US sites and the model's assumptions and should not be transferred directly to other regions.

Key points

1. Going from 95% to 99.9% hourly renewable coverage takes more than 5x the investment needed for 0% to 95% (Oregon); assuming average daily output underestimates the investment by an order of magnitude.
2. Meta's existing Net Zero renewable investments gave only 46% and 51% hourly coverage in two example regions; solar-only regions cannot get much beyond 50% without storage.
3. Embodied carbon inputs: wind 10-15 gCO2/kWh, solar 40-70 gCO2/kWh; Li-ion batteries 74-134 kgCO2 per kWh of capacity; server 744.5 kgCO2e (5-year life); 1.16x multiplier on server embodied carbon for construction.
4. LFP battery life is 3000 cycles at 100% DoD and 4500 at 80% DoD; moving to 80% DoD raised battery embodied carbon by 43% but lowered total carbon by 5% on average; tuning DoD lowers total carbon by 3-9%.
5. Battery needs for 24/7 coverage: about 5 hours of DC load in Utah and 14 hours in North Carolina (solar only); carbon-optimal battery capacities ranged from 200 MWh to 1800 MWh across sites.
6. Carbon-aware scheduling (40% of workloads flexible) raised coverage by 1% to 21% and needed 6% to 76% extra server capacity; with all workloads flexible, 24/7 needed 19% to over 100% extra servers.
7. Carbon-optimal renewables-only coverage was 37% to 97% (hybrid regions 88-97%); batteries cut total footprint by an order of magnitude, and batteries plus scheduling cut it by 15-65% and made 100% coverage optimal in five regions.
8. Grid lifecycle intensities (gCO2eq/kWh): wind 11, nuclear 12, hydro 24, solar 41, biofuels/other 230, natural gas 490, oil 650, coal 820.

Using it in design

• Add an hourly 24/7 CFE coverage calculator to site and energy design that sizes wind, solar, BESS (in hours of IT load) and server headroom together; use hourly or sub-hourly supply profiles, never daily averages.
• Store embodied-carbon defaults with source tags: wind 10-15 and solar 40-70 gCO2/kWh, Li-ion BESS 74-134 kgCO2/kWh capacity, about 745 kgCO2e per 1U server and a 1.16x construction multiplier.
• Report operational and embodied carbon together, and flag designs that chase the last few percent of coverage with very large solar/wind overbuild.
• Expose BESS depth of discharge as a parameter (default 80%), and when ranking sites warn that solar-only grids need about 14 h of storage against about 5 h for hybrid wind/solar.

Who should read it

Engineers evaluating 24/7 carbon-free energy targets in site and energy design, sustainability managers and researchers.

Publication details

Authors

Bilge Acun; Benjamin Lee; Fiodar Kazhamiaka; Kiwan Maeng; Udit Gupta; Manoj Chakkaravarthy; David Brooks; Carole-Jean Wu

Publisher

Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ACM)

Year

2023

Document number

arXiv:2201.10036

Pages

118-132

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1145/3575693.3575754

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation