We are witnessing a historic, global paradigm shift driven by dramatic improvements in AI. As AI has evolved from predictive to generative, more businesses are taking notice, with enterprise adoption of AI more than doubling since 2017. According to McKinsey, 63% of respondents expect their organizations’ investment in AI to increase over the next three years.
Paralleling this unprecedented adoption of AI, the volume of compute is also increasing at a stunning rate. Since 2012, the amount of compute used in the largest AI training runs has grown by more than 300,000 times. Yet, as sizable computing demands grow, significant environmental implications follow.
More compute leads to greater electricity consumption, and consequent carbon emissions. A 2019 study by researchers at the University of Massachusetts Amherst estimated that the electricity consumed during the training of a transformer, a type of deep learning algorithm, can emit more than 626,000 pounds (~284 metric tons) of carbon dioxide—equal to more than 41 round-trip flights between New York City and Sydney, Australia. And that’s just training the model.
We are also facing an explosion of data storage. IDC projects that 180 zettabytes of data—or, 180 billion terabytes—will be created in 2025. The collective energy required for data storage at this scale is enormous and will be challenging to address sustainably. Depending on the conditions of data storage (e.g., hardware used, energy mix of the facility), a single terabyte of stored data can produce 2 tons of CO2 emissions annually. Now multiply that by 180 billion.
This current trajectory for intensifying AI with an ever-growing environmental footprint is simply not sustainable. We need to rethink the status quo and change our strategies and behavior.
Driving sustainable improvements with AI
While there are undoubtedly serious carbon emissions implications with the increased prominence of AI, there are also enormous opportunities. Real-time data collection combined with AI can actually help businesses quickly identify areas for operational improvement to help reduce carbon emissions at a scale.
For example, AI models can identify immediate improvement opportunities for factors influencing building efficiency, including heating, ventilation, and air conditioning (HVAC). As a complex, data-rich, multi-variable system, HVAC is well-suited to automated optimization, and improvements can lead to energy savings within just a few months. While this opportunity exists in almost any building, it’s especially useful in data centers. Several years ago, Google shared how implementing AI to improve data center cooling reduced its energy consumption by up to 40%.
AI is also proving effective for implementing carbon-aware computing. Automatically shifting computing tasks, based on the availability of renewable energy sources, can lower the carbon footprint of the activity.
Likewise, AI can help diminish the ballooning data storage problem previously mentioned. To address the sustainability concerns of large-scale data storage, Gerry McGovern, in his book World Wide Waste, recognized that up to 90% of data is unused—merely stored. AI can help determine what data is valuable, necessary, and of high enough quality to warrant storage. Superfluous data can simply be discarded, saving both cost and energy.
How to design AI projects more sustainably
To responsibly implement AI initiatives, we all need to rethink a few things and take a more proactive approach to designing AI projects.
Begin with a critical examination of the business problem you are trying to solve. Ask: Do I really need AI to solve this problem or can traditional probabilistic methods with lower computing and energy requirements suffice? Deep learning is not the solution to all problems, so it pays to be selective when making the determination.
Once you’ve clarified your business problem or use case, carefully consider the following when constructing your solution and model:
Emphasize data quality over data quantity. Smaller datasets require less energy for training and have lighter ongoing compute and storage implications, thereby producing fewer carbon emissions. Studies show that many of the parameters within a trained neural network can be pruned by as much as 99%, yielding much smaller, more sparse networks.
Consider the level of accuracy truly needed to solve for your use case. For instance, if you were to fine-tune your models for a lower accuracy intake calculation, rather than compute-intensive FP32 calculations, you can drive significant energy savings.
Leverage domain-specific models and stop re-inventing the wheel. Orchestrating an ensemble of models from existing, trained datasets can give you better outcomes. For example, if you already have a large model trained to understand language semantics, you can build a smaller, domain-specific model tailored to your needs that taps into the larger model’s knowledge base, resulting in similar outputs with much more efficiency.
Balance your hardware and software from edge to cloud. A more heterogenous AI infrastructure, with a combination of AI computing chipsets that meet specific application needs, will ensure you save energy across the board, from storage to networking to compute. While edge device SWaP (size, weight, and power) constraints require smaller, more efficient AI models, AI calculations closer to where data is generated can result in more carbon-efficient computing with lower-power devices and smaller network and data storage requirements. And, for dedicated AI hardware, using built-in accelerator technologies to increase performance per watt can yield significant energy savings. Our testing shows built-in accelerators can improve average performance per watt efficiency 3.9x on targeted workloads when compared to the same workloads running on the same platform without accelerators. (Results may vary.)
Consider open-source solutions with libraries of optimizations to help ensure you’re getting the best performance from your hardware and frameworks out of the box. In addition to open source, embracing open standards can help with repeatability and scale. For example, to avoid energy-intensive initial model training, consider using pre-trained models for greater efficiency and the potential for shared/federated learnings and improvements over time. Similarly, open APIs enable more efficient cross-architecture solutions, allowing you to build tools, frameworks, and models once and deploy everywhere with more optimal performance.
Like many sustainability-led decisions, designing your AI projects to reduce their environmental impact is not easy. Reducing your energy and carbon footprint requires work, intention, and compromise to make the most responsible choices. But as we see in other sustainability-led business choices, even seemingly small adjustments can create large, collective improvements to reduce carbon emissions and help slow the effects of climate change.
To learn more about how Intel can help you reach your sustainable computing goals, visit Intel.com/sustainability.
This content was produced by Intel. It was not written by MIT Technology Review’s editorial staff.