Moving inference jobs rather than speeding them up: 38% less CO₂ on 600,000 simulated jobs

A simulator backed by 140 cloud regions and three years of carbon intensity traces shows that half of the existing sites are installed on medium to high carbon networks.

On September 13, Dayuan Chen and Ziliang Zong published CATS, a simulator designed to quantify what scheduling can actually subtract from the emissions of AI data centers. The work is based first of all on a new dataset: 140 cloud regions, in service or planned, from eight major providers, matched with the carbon intensity traces of 145 network zones in five-minute increments, from 2022 to 2024.

This corpus already yields a result of its own: half of today's sites are located on medium- to high-carbon networks. Since this intensity varies by a factor of more than 5.5 depending on the area and time of day, the place and time of execution of an inference weighs much more than intuition suggests.

The simulator then profiles six inference tasks on several types of GPUs, taking into account daily cycles, geographical mixes and service commitments. On a 24-hour trace with 600,000 tasks at a fleet utilization rate of 0.37, geographical displacement subtracts 38.4% of CO₂ compared to a strategy favouring speed, and time lag 16%, at the cost of 3.27% of commitment overruns. The authors draw two distinct recommendations: to set up future sites on low-carbon networks, and to apply carbon-sensitive scheduling to the fleet already built.