How Developers Can Lower AI’s Climate Impact

Quantifying AI’s carbon footprint isn’t straightforward. While energy consumption during AI training is well-recognized, energy used during deployment is less understood. Most AI companies do not disclose energy consumption metrics, which hinders efforts to measure and reduce their impact. However, several tools are emerging to fill this gap:

• Software Carbon Intensity (SCI) specification: Developed by the Green Software Foundation, this framework provides a way to estimate and track a project’s carbon emissions baseline over time.
• Green Algorithms Project: This calculator helps estimate total emissions for AI projects.
• Cloud Provider Carbon Tools: Amazon Web Services (AWS), Google Cloud Platform, and Microsoft Azure offer carbon-tracking tools specific to their platforms.
• Stanford University’s ML Framework: In collaboration with industry experts, Stanford researchers have published a lightweight framework to report energy use, compute resources, and carbon impacts for machine learning systems.

While measurement is the first step toward addressing AI’s environmental impact, making significant progress requires targeting the entire AI lifecycle.

Tackling the AI Lifecycle: Rethink, Reuse, Reduce

To make AI development more sustainable, developers must rethink how they build, train, and deploy models. Addressing each stage of the AI lifecycle can lead to substantial gains in both efficiency and sustainability.

A joint study by Google and the University of California, Berkeley demonstrated how AI’s energy consumption can be slashed up to 100x by adopting four key practices:

1. Efficient ML Architectures: Sparse models, for example, are more efficient than dense models in certain tasks.
2. Specialized Hardware: Using ML-optimized hardware like tensor processing units (TPUs) or modern GPUs, instead of general-purpose processors, reduces energy needs.
3. Cloud Computing: Utilizing the cloud for training and deployment improves data center efficiency and access to clean energy sources.
4. Energy-efficient Data Centers: Choosing cloud locations that use clean energy can cut CO2 emissions up to 1,000x.

Beyond these strategies, the core approach for sustainability should focus on three actions: rethink, reuse, and reduce. By rethinking AI workflows, reusing pre-trained models, and reducing redundant or energy-heavy processes, developers can create more value with fewer resources.