Google doubles down on machine learning to drive up energy efficiency of global datacentre fleet

Google Cloud is in the midst of a multi-year C-suite charm offensive that has seen it embark on a multi-pronged push to make its offerings even more endearing to enterprise IT buyers.

This work has seen it make numerous updates to the Google Cloud Platform (GCP) to bolster its functionality, security and performance to make it a more appealing place for enterprises to host their applications and workloads.

From a corporate responsibility standpoint, the environmental friendliness of the datacentre infrastructure underpinning GCP is becoming an increasingly important factor when enterprise IT buyers are weighing up which cloud provider to entrust their data too, claims Google’s vice-president of datacentres, Joe Kava.

“More and more our customers are telling us they have their own corporate mandates around sustainability, and energy use and renewable energy purchases,” he told attendees during a session at the 2019 Google Cloud Next user and developer conference in San Francisco.  

For Google, though, sustainability concerns have been front of mind for the company for more than decade, and long before Kava joined the datacentre development team more than 11 years ago.

“We started our renewable energy and zero-carbon journey in around 2007, when we decided – as a company – we were going to achieve a zero-carbon footprint. And we achieved that, primarily, through carbon offsetting,” he recalled.

“Then, in 2017, as well as celebrating our 10-year anniversary as a carbon-neutral company, we also announced we had achieved another milestone [in that] we were buying enough renewable energy to offset the entire consumption of our datacentres and overall business.”

This means for every megawatt hour of energy that gets consumed during the course of Google’s day-to-day operations, “at least” an equal amount of renewable energy is purchased to make up for it, according to Kava.  

“We are the world’s largest corporate purchaser of renewable energy... [and] the first company of our size to achieve that,” he said.

Offsetting its energy consumption in this way is all well and good, but next on the company’s sustainability agenda is finding a way to run its datacentres exclusively on renewable energy 24/7, he added.

“For us to truly take responsibility for our energy footprint and our impact on the environment and climate, we need to match our energy consumption with the renewable energy [generated] 24 hours a day, seven days a week,” he said.

That is no easy feat, given the challenges thrown up by the fact that renewable energy generation is a notoriously intermittent activity, plagued by variables that are largely out of control of the consumer.

There is also the fact that solar and wind power, for example, are not always included in the energy mix in some of the geographical regions where Google operates its datacentres.  

“The sun doesn’t shine 24 hours a day, and the wind doesn’t blow 24 hours a day, so we’re going to have to get creative with this,” said Kava.

On that point, the company has created a tool that enables it to assess how big of a gap there is, on an hourly basis, between the amount of energy its datacentres consume and how much renewable energy is being generated to cater for that.

“In parts of the world where we don’t have a lot of access to renewable energy, it’s not matching very well yet,” said Kava.

“In other parts of the world, we’re over 65% matching. We have a long way to go, but we think it is important enough to dedicate a team of resources towards. For us to take accountability for our footprint, this is what we have to do.”

To help improve the hit rate, Kava’s team is working closely with Deepmind, the Google-owned artificial intelligence (AI) company, which has led to the creation of a machine learning system that is enabling the firm to predict how much energy will be generated by a wind farm in 36 hours’ time.

“Why is that important? Because… the intermittency of renewables make it really hard for you to plan how much generation you are going to need. You know what your demand will be, more or less, but you don’t know what the output of that wind farm will be tomorrow,” said Kava.

“This machine learning system allows us to predict that with a high degree of fidelity [how much renewable energy will be generated] and now they can plan on that, put it in their base load calculations and we can better utilise the whole output of the wind farm.”

While Kava admits it is still “early days” for the system, this work has already led to a 20% improvement in the value derived by the wind farm where the technology is being tested, but he is hopeful this percentage will rise in time.