New to this? Read it in simple words
- Google started moving non-urgent computer work to hours with cleaner local electricity. The system compares carbon emission forecasts with computer demand forecasts.
- Jobs like training can often wait for cleaner hours. A live answer may need to arrive in less than a second.
- Flexible AI jobs can follow the hours when more clean energy is available. This can work without making the product slower for users.
- Forecasts are never perfect. The system needs a backup plan when the cleaner hour disappears.
- Carbon emissions
- Gases, mainly carbon dioxide, that go into the air when fuel is burned.
- Training
- Teaching an AI model by showing it huge amounts of data.
- Data center
- A building full of computers that run online services and AI.
The electricity supply changes every hour. Some AI work can move to cleaner times without making the product slower for users.
Move flexible work to cleaner hours
Google started moving non-urgent computer work to hours when the local electricity supply was cleaner. The system compares forecasts for carbon emissions with forecasts for computer demand.
Not every task has the same deadline. Training, indexing, media work, and large batches can often wait. A scheduler can move this work to hours with more low-carbon electricity.
However, teams must mark which tasks can move. A live answer may need to arrive in less than a second. A nightly test may have several hours. Clear deadlines help the scheduler protect the user experience.
Forecasts are never perfect. Weather, demand, power station problems, and grid limits can all change. The system needs new forecasts and a backup plan when the cleaner hour disappears.
Sources1
The daily schedule moves tasks that can wait into cleaner hours while urgent work continues.
The schedule has real limits
Google later used the same idea across different locations. Its cloud guide says time-based scheduling can reduce environmental impact when fresh regional energy data is available.
A scheduler can also consider price, grid pressure, water use, and power limits. The difficult part is moving work without making the service worse for people.
Moving data to another region has a cost. Privacy or national rules may block the transfer. The cleaner region may have no spare capacity, and a long network path can use extra energy.
Many teams must work together. Product teams set deadlines, energy teams provide trusted data, and cloud teams protect enough capacity. A dashboard cannot help if every task is treated as urgent.
Sources2
How to report the result honestly
A strong report should show more than yearly renewable-energy purchases. It should explain which hours matched cleaner energy, how much work could move, and which regions were included.
The report should show how much energy the task used, when and where it ran, and which carbon data guided the choice. It should also name the starting point used to calculate any improvement.
AI teams can begin by sorting tasks by urgency. They can then share those deadlines with the cloud system. This lets computer demand respond to the energy system instead of using the same amount at every hour.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · April 22, 2020Our data centers now work harder when the sun shines and wind blows Google
- 2
We did not calculate the total emissions of any AI service.
Google moves flexible computer work to hours with cleaner, sometimes cheaper electricity, without slowing products for users. How we rate the mood