Artificial intelligence (AI) boasts ample use cases that promise to advance sustainability efforts and help avoid more than 1.5-degree C of global temperature rise, but the carbon intensity of AI workloads must be also factored into the equation, according to new research from Microsoft and Allen Institute for AI (AI2).

The research calls on cloud providers to provide software carbon intensity information and actionable steps to allow developers and users to recognize and lower the carbon footprint of AI workloads.

This rings especially true in sustainability use cases like using AI to integrate renewable energy into an electricity grid or lowering the cost of carbon capture, Microsoft's Will Buchanan and AI2's Jesse Dodge explained in a blog post about the research.

"The technology itself needs to be sustainable," the researchers urged, citing the current lack of mature infrastructure and accurate systems to measure AI's overall carbon footprint.

"This requires interoperable measurement tooling; only once this is built can effective carbon management strategies be built," Buchanan and Dodge explained.

The research identified carbon-aware computing as a best practice in considering the environmental impacts of AI workloads. This principle is based on shifting computing resources to capitalize on sensitive differences that impact carbon intensity.

Location and time, for example, can both affect the carbon intensity of electricity grids, which varies by day, season, and hour based on shifts in demand and the type of energy sources supplying the grid, the researchers explained.

This means that users can and should make decisions that render their workloads more carbon-aware, like choosing a geographic region where more renewable sources supply its grid or running a training job at a specific time of day.

In an effort to develop a framework that measures operational carbon emissions on a granular basis, the research from Microsoft and AI2, in collaboration with Hebrew University and Carnegie Mellon University, applied the Green Software Foundation's Software Carbon Intensity (SCI) measurement system to AI workloads running on Microsoft Azure. The SCI employs a consequential carbon accounting approach to calculate the change in emissions resulting from purposeful decisions, interventions, or actions.

The research identifies multiples actions and carbon-aware strategies users can take that will lower their SCI.

Choosing the appropriate geographic location to run workloads is by far the most influential, offering up to 75% reduction in SCI. Particularly in situations when latency and cost are equivalent, the researchers urge users to prioritize carbon intensity when choosing locations.

Time of the day is also an important aspect to consider. With shorter workload training runs, Microsoft saw more than 30% reduction in SCI in multiple regions and saw up to 80% reduction in regions with high renewable energy intermittency. With longer runs, however, the reduction is less than 1.5%.

Another strategy to lower operational footprint is pausing workloads when carbon intensity is high and resuming when it's lower, which takes advantage of daily variations in carbon intensity, Buchanan and Dodge explained. They concluded shorter workloads have an SCI reduction potential of less than 10%, but larger workloads' SCI decreased about 25%.