will pause its AI during heat waves to relieve the electricity grid

Google has announced A pioneering agreement with two of the main electrical companies in the United States to reduce the consumption of their data centers during energy demand peaks. In other words, Google will pause its artificial intelligence during heat waves.

AI is hungry. It is no secret that generative artificial intelligence has a voracious energy appetite. Train and execute the huge models that drive from the summaries in the Google search engine to the youtube folded videos requires dozens or hundreds of megawatts of power continuously.

This massive consumption, concentrated in huge data centers, has become a growing concern for electricity companies, which see how energy demand triggers at a rate that the current infrastructure is not prepared to support.

An unthinkable movement. Until now, the main concern was to add enough power to the network, with solutions that go From reopening abandoned nuclear centers until Sign the largest hydroelectric agreement in history. But Google’s last movement proposes an unthinkable solution in the competitive AI industry: instead of just increasing energy supply, managing demand flexibly.

The measure, agreed with two of the main electrical companies in the United States (TVA and I & M.), arrives just when both states prepare for an intense heat wave. While millions of homes and companies light their air conditioners, putting the electricity to the limitGoogle has agreed to reduce the consumption of its less urgent tasks to avoid overloads and possible blackouts.

Demand flexibility. Google has not invented anything new. This type of Flexible response to network demand It already applies in many industries, usually to pay a lower price for the light. Google itself has used it for a long time to postpone non -essential tasks, such as the processing of YouTube videos, moving them to data centers in other time regions or executing them at night, when the demand for the network is lower.

The novelty is that, for the first time, it will apply contractual to automatic learning workloads, the heart of the AI. According to the agreement, if the demand for energy increases dangerously or there is an interruption in the network due to extreme weather conditions, operators can ask Google to reduce their consumption. Google will respond by reprogramming or limiting non -urgent tasks until the network stabilizes.

You won’t run out of Google Maps. The company has made it clear that this demand response system has its limits. Critical services that require 100% reliability and constant availability, such as the search engine, Google Maps or cloud services for clients of essential sectors such as health, will not be affected.

Flexibility will apply to tasks such as the training of new AI models thanks to advances in techniques such as Checkpointing (which allows you to save the progress of training and resume it later). A model could be trained exclusively at night, when the network capacity is greater, without losing the work already done.

What does Google win with this? In addition to relieving the overload of the network and preventing off, the clients of the network will end up reducing the light of the light thanks to this system. Including Google data centers.

Image | Pawel Czerwinski in Unspash

In Xataka | The AI has disrupted Google’s plans to be sustainable. His plan to remedy it: the “four m”