Google is practicing what it preaches, handing control of one of the most vital components of its data center operation over to its machine-learning algorithms during the past few months.
DeepMind, the Google subsidiary that is responsible for much of its advanced artificial intelligence research, announced Friday that Google has saved 30 percent on its energy bills by improving the efficiency of its cooling systems. “This first-of-its-kind cloud-based control system is now safely delivering energy savings in multiple Google data centers,” Google said in a blog post.
Big cloud companies love to talk about how earth-friendly they are with their data center designs and blueprints, and while that’s true in many cases the real motivation is money: electricity usage can be the biggest expense on a data center cost sheet. Google Cloud has data centers in 17 regions around the world, and multiple availability zones within many of those regions, which is a lot of hardware to keep cool.
DeepMind and Google’s data center teams developed an algorithm that can predict future energy consumption in its data centers based on past usage trends. When implemented, the algorithm manages the cooling configurations — something usually done by a team of experienced humans — in response to changes in demand for computing power, and therefore energy.
As it is perfected, this work should allow Google to get a little more return on the massive investments it is making in data center hardware as it tries to keep up with Amazon Web Services and Microsoft Azure in the cloud market. In just the past quarter, Google spent $5.5 billion on capital projects, and most of that likely went toward data centers.
And it’s another selling point for Google’s AI technology, which it has made a core part of its pitch to cloud computing customers. After turning over control of its bread-and-butter search engine to AI algorithms in 2015, the company is now putting another extremely important part of its business model in the hands of the machine.
(Editor’s note: This post was updated to correct the percentage improvement obtained by this project.]
