炭素強度スケジューリングは、コードレベルのエネルギー削減よりも大きな気候利益をもたらす

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides empirical evidence that carbon-aware scheduling can yield orders-of-magnitude greater emission reductions than code optimizations, informing future research on sustainable computing.

🏢実務担当者:Offers actionable recommendations for data center operators and software developers to reduce carbon footprint by shifting workloads to times and locations with lower carbon intensity.

🏛政策担当者:Highlights the potential of carbon-aware computing as a demand-side flexibility measure, suggesting policies that incentivize temporal and spatial shifting of computational workloads.

Climate change is upon us and it is imperative to take action to reduce carbon emissions. Fossil fuels<br>and aviation are often held up as culprits. Computing, and particularly generative AI, is far from innocent.<br>Computing’s emissions are said to be comparable to flying, and the emission rate is now soaring due to the<br>computational costs of training large language models and the ubiquitous integration of AI into browsers and<br>search engines. Advocates such as the Green Software Foundation (GSF) have highlighted how computing is<br>part of the climate problem, and suggested ways for computing to reduce its carbon footprint. This includes<br>reducing the energy of a set of tasks (or reducing how often those tasks are undertaken). This approach has<br>had widespread interest, with people looking at CPU and GPU clock frequencies and reducing the precision<br>of arithmetic. A second approach, often overlooked, is to reduce the carbon for a given workload. This<br>centres on “shifting”. Since some electricity is greener than others, depending on the renewable-to-fossil<br>mix, it is possible to reduce emissions by shifting in time and space. The amount of carbon emissions a unit<br>of electricity is responsible for is its carbon intensity (CI), which we show to be a key factor in sustainable<br>computing.<br>This work examines reducing the carbon footprint of computational matrix operations. We examine<br>naive choices of compiler optimisations, computing languages, and numerical libraries. We compare this with<br>potential savings due to the dynamic nature of CI via careful scheduling of where and when computation<br>is undertaken. We find that careful choice of Implementation and Deployment can yield an energy saving<br>factor of 51, and that shifting when and where a Deployment is run can yield a further carbon saving factor<br>of 354x.<br>We illustrate a real case example of carbon savings of over 7750x compared to a given Baseline, and<br>provide key insights and recommendations for code developers and users, and for policy makers.

• openalex https://doi.org/10.83056/mmu.33078665.v2first seen 2026-08-19 04:34:27 · last seen 2026-08-20 04:32:30