How to Do AI Sustainably

Enterprises are increasingly concerned with how they can take a responsible approach to AI, ensuring their models maximize business value for stakeholders by delivering results that are accurate, unbiased and compliant. But AI models must also be sustainable to be responsible.

At the Gartner® IT Symposium/Xpo™ in October 2022, Daryl Plummer, distinguished VP analyst and Gartner Fellow, shared several top strategic predictions for 2023 and beyond, including this one:

“By 2025, without sustainable artificial intelligence (AI) practices, AI will consume more energy than the human workforce, significantly offsetting carbon zero gains.”[1]

Enterprises can use their AI models to pursue a variety of sustainability use cases, but they must also design their AI models to be as clean and efficient as possible. To be net-positive in their emissions goals, businesses must both use AI for sustainability and use AI in a sustainable way.

Achieving sustainability goals with AI

AI models can help enterprises turn data into predictive insights, thus allowing them to make informed decisions on key business questions. From a sustainability perspective, these insights can help businesses identify and remove waste within their operations.

Dynamic workload scheduling

AI models can help enterprises identify workloads that aren’t time-sensitive and automatically schedule them for off-peak hours when there’s less competition for available energy. This will not only decrease the organization’s overall power costs, but also position them to make the most of available renewable energy sources. Using renewable energy across more workloads will help the organization shrink its carbon footprint.

Green building practices

Any time an organization plans a new facility or updates operations in an existing facility, they’ll want to better understand how energy and space consumption will impact their carbon footprint. Well-conceived AI models can help them accurately predict airflow and emissions under specific scenarios and identify potential inefficiencies that could lead to waste. This will allow the business to design and operate facilities in a way that effectively balances sustainability alongside other top priorities, including costs and business value.

Green supply chains

AI-driven insights can help organizations better understand their Scope 3 emissions, which are emissions driven by their supply chain rather than their own direct operations. By modeling the emissions impact of working with specific partners, businesses can identify the suppliers that offer the best combination of reliability and sustainability benefits.

Building greener AI models

AI models can use significant amounts of energy and drive higher emissions—that is, unless enterprises make the extra effort to design and manage them efficiently. The simple reason AI models can be so energy-hungry is that they must process massive volumes of data or run numerous iterations on the data to ensure accuracy and statistical significance in model outcomes.

All computational processes have a carbon footprint. Since AI workloads have significantly higher compute needs than conventional IT workloads, it stands to reason that their carbon footprint might also be larger. Also, AI workloads need to be retrained frequently over time to maintain accuracy. This means the carbon footprint of specific models may only continue to grow as time passes.

For all these reasons, it’s especially important for data science teams to make their AI models as efficient as possible.

Responsible data preparation

For data science teams, the initial steps toward more efficient AI may be in focusing their perspective on the data sets they gather, and improving the way data sets are represented to models. In an era when more real-time data is available from more sources than ever before, the potential for inefficiency and waste is undeniable. A good way to limit the carbon footprint of AI models is to identify ways to utilize only the most relevant data needed with minimal compromise on model quality. There are several ways to do this:

• Determine which data attributes relate more to target variables of interest and reduce dimensionality of input feature space
• Improve effectiveness by identifying new derived features that effectively represent multiple input attributes
• Incorporate transfer learning from pretrained models where applicable
• Develop a strategy to remove redundant word embeddings in natural language models

Responsible developer mindset

Data science teams must also make their AI models efficient from a computational standpoint. Teams must judiciously design experiments that enable them to test their initial ideas in a modular and compute-efficient manner. While performing proofs of concept in formative stages of model development, even failing fast should be achieved in a sustainable way.

Running brute-force models that attempt to address the entire problem at once may lead to inefficiency, with the model spending many GPU cycles to perform relatively simple checks. Instead, teams could take a modular approach to model development and split up the problem into appropriate chunks, with the goal of “right-sizing” each chunk for optimized efficiency. As they identify subsets of the model that are performing well, they can build on that success while also making iterative improvements on lower-performing subsets. With such a modular approach, teams could even benefit from “model cascading”, a technique that utilizes a collection of models of varying capacities and complexities to output predictions in an accurate yet sustainable way.

Teams should challenge themselves to take a sustainable approach to model building. They could even host hackathons with the goal of improving the efficiency of existing ML products and reducing carbon footprint. To identify outside-the-box efficiency improvements, it may even help to ask developers how they would design, train, retrain and check efficacy of the model if they were paying all compute expenses out of their own pocket.

Displaying a custom real-time tracker of carbon footprint reduction would encourage individuals and teams to do more and innovate further to do AI in a sustainable way. Tracking specific efficiency improvements in terms of trees saved or cars taken off the road could help make things fun, bringing everyone together in pursuit of common sustainability goals.

Incorporate sustainability metrics into your success criteria

Accuracy and sustainability are typically tracked as separate metrics by organizations and teams. You could achieve accuracy using inefficient data analysis practices or models, but that would mean not meeting your sustainability goals. In contrast, being efficient doesn’t necessarily mean sacrificing accuracy. It’s entirely possible to get the accuracy you expect using efficient models and data sets; it just requires you to balance these two seemingly disparate metrics when designing your AI models. As responsible data citizens, we must all start making sustainability metrics part of the key performance indicators (KPIs) that define the success of our AI initiatives.

Work with the right digital infrastructure partner to support sustainable AI

Since AI workloads require so much data processing and compute capacity, it’s especially important that you work with a digital infrastructure partner that has a strong sustainability foundation to support them. Equinix was the first company in the data center industry to commit to becoming climate-neutral by 2030. In pursuit of that goal, we’ve worked to maximize renewable energy use in our data centers. During FY2021, we achieved 95% renewable energy coverage across our global operations, our highest annual rate so far.

We’re also working to make our data centers cleaner and more efficient by testing and deploying new sustainable technologies such as high-density liquid cooling. Compared to traditional air-cooling methods, liquid cooling is better suited to support AI and other compute-intensive technologies.

Learn more about Equinix’s climate commitments, and how we’re innovating to meet those commitments. Access our interactive sustainability report today.

[1] “Gartner Unveils Top Predictions for IT Organizations and Users in 2023 and Beyond”, Gartner Press Release, October 18, 2022. GARTNER and IT SYMPOSIUM/XPO are registered trademarks and service marks of Gartner, Inc. and/or its affiliates in the U.S. and internationally and are used herein with permission. All rights reserved.