Google will pause non-essential AI workloads to protect power grids, the advertising giant announced on Monday.
The web giant already does this sort of thing for non-essential workloads like processing YouTube vids, which it moves to datacenters where power is available rather than continuing to run them in places demand for energy strains the grid.
Under an agreement with Indiana Michigan Power (I&M) and the Tennessee Valley Authority (TVA), Google will use the same techniques for AI workloads.
The announcement comes as states served by the power companies brace for a heat wave that will likely strain the grid as residents use air conditioners and increase demand for energy. Amid debate about datacenters’ consumption of power and water, the last thing that the Chocolate Factory needs is folks blaming its AI Mode search function for a power outage when temperatures top 100°F (37.7°C).
Under the agreement, if energy demand surges or there's a disruption in the grid due to extreme weather, I&M and TVA can now request that Google reduce its power use by rescheduling workloads or limiting non-urgent tasks until the issue is resolved.
By dynamically adjusting how much power its bit barns are allowed to consume, a process Google calls “demand response”, the web giant argues that new datacenters can be interconnected more quickly — presumably because utilities are less concerned about them causing brown outs or outages.
"By including load flexibility in our overall energy plan, we can manage AI-driven growth even where power generation and transmission are constrained," Google wrote in a blog post on Monday.
Training and running AI models can easily consume tens or even hundreds of megawatts of power for hours, days, or weeks at a time, depending on how big or complex they are. However certain workloads don't need to run 24/7. Advancements in checkpointing mean that a model could be trained exclusively at night when grid capacity is at its greatest.
Datacenter demand-response is still a nascent technology and is only being employed at a handful of Google bit barns. Further complicating the matter, the approach is also incompatible with certain high-demand workloads, such as Search, Maps, or its cloud business, Google notes.
Google may have the ability to spin down its own machine learning workloads as it pleases, but it can't exactly pause its cloud customers' AI jobs without causing a few headaches.
... continue reading