Google 's Artificial Intelligence Team DeepMind succeeded in improving the cooling efficiency of the Google data center cooling system by 40% in a few months

It was a startup to do artificial intelligence development "Deep Mind"We are continuing artificial intelligence research independently based on the UK based in London after being acquired by Google. The artificial intelligence technology developed by DeepMind is an algorithm that self-learns and improves games "DQN"Software that breaks the world champion of Go"AlphaGo"Whenever it is announced from time to time, I can make the public a noise every time, but it became clear that it demonstrates great power in the energy reduction of Google's data center.

Google Deep Mind
https://deepmind.com/blog


Supporting various Google services such as Google search, Gmail, YouTube,Data centers Google has around the worldis. Since making this data center efficient will lead to improved quality of service for Google, we continue to constantly update our data center at the system level. For example, it is said that efforts are being made to replace servers used in data centers with highly efficient ones, and to cover electricity with renewable energy.

Among them, "cooling efficiency" is very important in talking about the quality of the data center. Cooling data centers that generate a lot of heat is essential for the stable operation of the data center. And, the energy required to cool the data center occupies the majority of the energy used in the data center, reducing the energy for cooling leads to lowering the environmental burden. For that reason, Google has long worked on reducing energy by increasing cooling efficiency. For companies that use the Google data center, such as cloud services, if Google's data center energy savings are realized, Google seems to focus on reducing energy to help indirectly protect the global environment.


However, the cooling system of the data center has characteristics that it is difficult to apply desk theory and human intuition etc. since the equipment making up the data center is too complicated and affects each other nonlinearly. Furthermore, because each data center in the world has its own environment and architecture, there are also difficulties that it is impossible to apply a cooling model optimized for a data center to other data centers. As a result, efficiency improvement of the cooling system of the data center has been limited to constructing a framework at a very general level, making efficiency of the entire cooling system of the data center a challenging task.

In order to promote cooling efficiency of such data center, Google said that it has taken machine learning from two years ago. It is trying to improve efficiency by setting up a model that predicts data center usage by machine learning.

Deep Mind, a specialized team of artificial intelligence, joined the data center team for several months joining this Google data center cooling system efficiency project. The DeepMind team analyzes data such as data center air temperature, electric power, coolant flow speed in the cooling pump, etc. which had already been collected by thousands of engineers by deep learning which puts it in a multilayered neural network and increases efficiency We have created a framework. As a result, we succeeded in constantly reducing the power consumption of the cooling system by 40%. This corresponds to a 15% improvement in the energy usage power usage effectiveness (PUE).

In the graph of power usage efficiency, it can be confirmed that the power usage efficiency has dramatically decreased when ML (machine learning) is turned ON.


Deep Mind, which has done a ridiculous job of Google's data center team to reduce the cooling efficiency that has been improved by machine learning over the course of two years by as much as 40% in just a few months, is the best choice among complex changes Since the method of machine learning to find the condition is versatile, it will take several months from now to increase the efficiency of the power generation facility, reduce the energy of the semiconductor part, reduce the water consumption, improve the throughput, We are planning to tackle the efficiency of the part.