Source-linked AI summary
Carbon Explorer: A Holistic Approach for Designing Carbon Aware Datacenters
Bilge Acun, Benjamin Lee, Fiodar Kazhamiaka, Kiwan Maeng, Manoj Chakkaravarthy, Udit Gupta, David Brooks, Carole-Jean Wu
TL;DR
Datacenters pursuing 24/7 carbon-free computing must address intermittent renewable supply, while existing approaches often overlook embodied carbon and regional trade-offs. Carbon Explorer evaluates renewable capacity, batteries, and carbon-aware scheduling across operational and embodied footprints, finding that carbon-optimal designs vary by location and may not provide 100% coverage.
Problem
24/7 carbon-free computing requires coordinating intermittent renewable supply with datacenter demand, but existing approaches often overlook embodied carbon and multidimensional regional trade-offs.
Method
Carbon Explorer explores renewable investments, battery storage, and carbon-aware computation shifting using geographically dependent supply and datacenter demand data.
Results
Carbon-optimal strategies vary by geography, and considering embodied carbon means 100% 24/7 operational carbon-free computing may not always be optimal.
Takeaways & Limitations
Renewable capacity, storage, and scheduling should be evaluated as a coordinated portfolio against both operational and embodied carbon.
Takeaways & Limitations
The study omits lithium-extraction impacts, electronic waste, and cooling-water use from its broader environmental assessment.
Abstract
from arXiv · showhide
Technology companies have been leading the way to a renewable energy transformation, by investing in renewable energy sources to reduce the carbon footprint of their datacenters. In addition to helping build new solar and wind farms, companies make power purchase agreements or purchase carbon offsets, rather than relying on renewable energy every hour of the day, every day of the week (24/7). Relying on renewable energy 24/7 is challenging due to the intermittent nature of wind and solar energy. Inherent variations in solar and wind energy production causes excess or lack of supply at different times. To cope with the fluctuations of renewable energy generation, multiple solutions must be applied. These include: capacity sizing with a mix of solar and wind power, energy storage options, and carbon aware workload scheduling. However, depending on the region and datacenter workload characteristics, the carbon-optimal solution varies. Existing work in this space does not give a holistic view of the trade-offs of each solution and often ignore the embodied carbon cost of the solutions. In this work, we provide a framework, Carbon Explorer, to analyze the multi-dimensional solution space by taking into account operational and embodided footprint of the solutions to help make datacenters operate on renewable energy 24/7. The solutions we analyze include capacity sizing with a mix of solar and wind power, battery storage, and carbon aware workload scheduling, which entails shifting the workloads from times when there is lack of renewable supply to times with abundant supply.
1 INTRODUCTION
24/7 carbon-free computing is difficult because renewable supply fluctuates hourly and seasonally, while Net Zero accounting can conceal continued carbon-intensive consumption. Carbon Explorer evaluates renewable capacity, storage, and workload scheduling together, including operational and embodied carbon.
- Datacenters consumed 205 TWh worldwide in 2018, making computing a significant and growing contributor to electricity demand.
- Renewable intermittency creates periods of excess supply and scarcity, complicating hourly carbon-free operation.Excess supply can force curtailment, while scarcity causes datacenters to use carbon-intensive grid energy.
- Net Zero matching can offset annual consumption with renewable credits while hourly electricity use remains as carbon-intensive as the grid during shortages.
- Carbon Explorer analyzes renewable investments, energy storage, and carbon-aware workload scheduling as complementary responses to intermittent supply.
- Batteries and scheduling improve coverage but add embodied carbon through battery and server manufacturing, so their benefits depend on workload flexibility and location.Scheduling raises coverage by 1%–22% and requires 6%–76% additional servers; 40% workload flexibility can justify those embodied costs.
2 CARBON EXPLORER
Carbon Explorer combines datacenter demand and geographically varying renewable supply to explore renewable deployment, batteries, and carbon-aware scheduling. It quantifies trade-offs between operational and embodied carbon to identify carbon-minimum designs.
- Carbon Explorer uses time-series datacenter power demand and geographically specific renewable generation as inputs.
- The framework explores varied renewable investments, energy-storage amounts, and computation shifting.
- Carbon-aware scheduling shifts computation toward periods of abundant renewable energy but requires additional server capacity for deferred work.
- Carbon Explorer models operational and embodied carbon to navigate trade-offs among renewable generation, batteries, and scheduling.
- Renewable energy is constrained by geographic availability, while batteries and additional servers introduce manufacturing-related carbon overheads.
3 OPERATIONAL GRID INPUTS: DEMAND AND SUPPLY CHARACTERISTICS
Carbon Explorer characterizes datacenter demand and regional renewable supply using hourly grid and production-datacenter data. The inputs reveal substantial geographic and temporal variability, curtailment, and the gap between annual Net Zero accounting and hourly carbon intensity.
- The framework uses hourly datacenter demand and local grid-generation data for Meta datacenter locations, with nearly six gigawatts of renewable investment.
- The EIA Hourly Grid Monitor supplies hourly generation statistics collected from balancing authorities across the lower 48 states.
- Characterizing Datacenter Power Demand: Meta datacenters exhibit diurnal load patterns and event-driven peaks; average CPU utilization swings about 20% at Meta and 15% between maximum and minimum at Google.
- Characterizing Datacenter Power Supply: Renewable profiles vary by geography: BPAT is primarily wind, DUK primarily solar, and PACE uses a mix of wind and solar.
- Characterizing Datacenter Power Supply: For BPAT, the best ten days provide approximately 2.5 times the average renewable energy, while the worst days provide very little.
- Characterizing Datacenter Power Supply: Increasing renewable capacity intensifies supply-demand mismatch and curtailment, while Net Zero credit matching can leave hourly carbon intensity near the grid level during shortages.
- Characterizing Datacenter Power Supply: Energy storage and demand response address intermittent supply by providing carbon-free energy during scarcity or shifting demand toward abundant renewable periods.
4 DATACENTER DESIGN: STRATEGIES FOR CARBON FREE COMPUTING
Carbon Explorer analyzes complementary renewable investments, battery storage, and carbon-aware scheduling to achieve 24/7 carbon-free computing while accounting for operational and embodied carbon. Its analyses show that regional renewable profiles create diminishing returns and different storage and server-capacity requirements.
- Strategies for Carbon-Free Computing: Carbon Explorer models renewable investments, energy storage, and carbon-aware scheduling together, including their operational and embodied carbon trade-offs.The framework uses regional demand and renewable-supply data, battery models, and production datacenter traces.
- Renewable Energy: Hourly renewable coverage reaches only 46% and 51% in two regions under Meta’s existing renewable investments, despite monthly or annual Net-Zero alignment.Investment profiles generally follow local grid preferences, except Oregon, where Meta emphasizes solar despite a wind-oriented grid.
- Renewable Energy: More than 5× greater renewable investment is needed to increase Oregon coverage from 95% to 99.9% than from 0% to 95%.The flattening coverage curves indicate diminishing marginal returns, motivating complementary storage and scheduling solutions.
- Renewable Energy: Assuming average daily wind and solar output would underestimate the investment needed for 100% coverage by an order of magnitude.Carbon Explorer therefore matches fine-grained hourly renewable supply against datacenter demand when determining investments.
- Battery Storage: Mixed solar and wind regions require less battery capacity because complementary generation reduces day-to-day variability; Utah reaches 24/7 coverage with around five hours of battery capacity.By contrast, North Carolina’s solar-only datacenter requires 14 hours of battery-based compute, while wind-dominated Oregon also has high requirements.
- Carbon-Aware Scheduling: Achieving 24/7 carbon-free computation can require 19% to over 100% additional server capacity, creating a trade-off between operational and embodied carbon.Turbo Boosting existing servers is identified as an alternative that can increase compute throughput without increasing capital costs and embodied carbon.
5 CARBON MINIMIZATION: HOLISTIC DESIGN EXPLORATION
Carbon Explorer explores datacenter designs by jointly minimizing operational and embodied carbon across renewable generation, batteries, and carbon-aware scheduling. The carbon-optimal design varies by region and may not provide 100% coverage when embodied costs outweigh operational savings.
- Holistic Design Exploration: Carbon Explorer identifies datacenter designs by jointly accounting for operational carbon and embodied carbon from renewable farms, batteries, and servers.It exhaustively searches combinations of renewable capacity, storage, and demand-response capacity using geographic energy data and datacenter demand patterns.
- Renewables Only: Renewable generation alone achieves 37% to 97% coverage, with hybrid wind-and-solar regions reaching 88% to 97%.Wind farms and complementary wind-solar assets mitigate supply variance, while primarily solar regions struggle because solar is available only during parts of the day.
- Renewables + Battery: Batteries reduce total carbon by an order of magnitude across regions and enable 100% coverage in four of thirteen regions.Optimal battery capacities range from 200MWh to 1800MWh; other regions reach optimal coverage between 82% and 99%.
- Battery Management: 80% battery depth of discharge increases cycle life by 50%, raises embodied carbon by 43%, and lowers total carbon by 5% on average.Shallower discharge requires larger batteries, demonstrating a trade-off between battery longevity, embodied carbon, and operational carbon.
- Renewables + CAS: Carbon-aware scheduling increases coverage by 1% to 21% while requiring 6% to 76% additional servers for deferred computation.Scheduling is constrained by workload flexibility and server capacity, making it insufficient for full coverage in regions with many near-zero renewable-supply days.
- Combined Solutions: Batteries and scheduling can reduce total carbon by 15% to 65%, but their effectiveness depends on region, workload delay tolerance, and battery availability.The Pareto frontier shows increasingly expensive solutions near complete coverage; batteries incur lower embodied costs than renewable-only or server-heavy scheduling designs as coverage rises.
6 DISCUSSION AND RELATED WORK
The discussion situates Carbon Explorer among work on renewable energy, storage, and carbon-aware scheduling while emphasizing broader sustainability constraints. Its parameterized models address evolving emissions data, but several environmental impacts remain outside scope.
- Renewable Energy: Prior work studies on-site renewable generation, microgrids, modular deployment, and local solar to reduce grid energy consumption.These approaches provide foundations for datacenter renewable-energy design but do not encompass the full Carbon Explorer design space described here.
- Energy Storage: Prior storage research addresses datacenter availability, renewable intermittency, battery heterogeneity, and aging through charge-discharge management.Carbon Explorer extends this line of work by quantifying storage requirements for 24/7 carbon-free computing.
- Energy Storage: Lithium-ion batteries have fallen 80% in price from 2015 to 2020 and can supply 28 MW for four hours.These operating parameters align with hyperscale datacenters provisioned for 20 to 40 MW and could reduce demand-response requirements.
- Other Considerations: Large-scale battery deployment carries environmental and health risks from lithium extraction and disposal of toxic or flammable materials.Spent batteries require proper recycling and disposal to avoid soil, water, and air contamination.
- Carbon-Aware Scheduling: Carbon-aware scheduling research uses time-series forecasts to optimize flexible jobs against electricity prices, carbon intensity, and service quality.The paper performs offline analyses to defer flexible workloads in response to renewable supply.
- Other Considerations: Carbon Explorer quantifies manufacturing and recycling carbon but omits lithium-extraction impacts, electronic waste, and cooling water usage.These broader environmental effects are explicitly left beyond the study’s scope.
- Modeling Scope: The framework uses parameterized models because operational and embodied emissions depend on evolving technologies, energy agreements, and supply-chain reporting.Its parameters are based on the best publicly available data.
7 CONCLUSION
Carbon Explorer explores carbon-optimal investments across renewable generation, storage, and carbon-aware computation shifting. It shows that optimal strategies vary geographically and that 100% 24/7 operational carbon-free computing is not always carbon-optimal when embodied emissions are included.
- Conclusion: Carbon Explorer determines carbon-optimal investments in renewable types, energy storage, and carbon-aware computation shifting using geographic supply and datacenter demand characteristics.The tool is designed to guide investment strategies toward operational and embodied carbon-footprint optimality.
- Conclusion: Carbon-optimal strategies vary by geographic location, and 100% 24/7 operational carbon-free computing may not be carbon-optimal when embodied carbon is considered.The conclusion frames the design choice as a trade-off across renewable energy, storage, and workload shifting.