Microsoft’s water consumption surged 34 p.c to six.4 million cubic metres in 2022 and a generative AI guzzle could also be guilty.
“It is truthful to say the vast majority of the expansion is because of AI,” Shaolei Ren advised the Related Press in a report revealed Saturday, who pointed to Microsoft’s relationship with OpenAI and large funding in generative AI merchandise as proof.
Ren, a researcher on the College of California, Riverside, has been learning the affect of generative AI adoption on datacenter water consumption, and penned a paper on the topic this spring.
In its newest environmental, social, and governance (ESG) report, Microsoft mentioned the upper fee of water consumption was consistent with enterprise progress. In response to that report, water consumption elevated by a 3rd from 4.8 cubic metres of water in 2021 to six.4 million cubic metres final yr. That is in comparison with the 14 p.c improve in water consumption the software program large reported between 2020 and 2021.
Whereas Microsoft does not specify AI adoption because the formal reason behind this improve, we do know that the cloud supplier has been deploying tens of 1000’s of GPUs to energy the massive language fashions behind Bing Chat and GitHub Copilot, and others. The mega-corp can also be working carefully with OpenAI, the developer behind the massive language fashions (LLM) utilized in ChatGPT.
The Register reached out to Microsoft for touch upon its elevated water consumption. A rep advised us:
Datacenter water consumption has turn out to be some extent of concern, particularly in America the place scientists have warned altering climate patterns are prone to lead to widespread and protracted drought situations.
In response to US local weather fashions developed on the Division of Power’s Argonne Nationwide Labs, by the center of the century — only a brief 27 years from now — giant parts of the nation could possibly be in a state of persistent drought. The fashions present that these situations are prone to be adopted by transient however devastating floods.
What’s Microsoft utilizing all of that water for?
Whereas the water consumption figures highlighted in Microsoft’s ESG report are for the corporate as an entire, not simply datacenters, it is no secret that these amenities suck down loads of water.
The GPUs used to energy these fashions are usually deployed in units of eight and devour a prodigious quantity of energy in comparison with conventional datacenter infrastructure. It is commonplace for an eight GPU system to devour between 6kW and 10kW of energy beneath full load. To place that in perspective, which means a single server can devour as a lot as a typical cloud rack.
All of that warmth must go someplace, and relying on the datacenter cooling expertise, hotter techniques can translate into higher water consumption.
Inside these amenities water is utilized in quite a lot of datacenter cooling functions. Deionized water and different fluids can be utilized in direct liquid cooled (DLC) techniques. Relatively than blow air over a heatsink to chill the processors and GPUs, this tech removes warmth by passing coolant by way of chilly plates connected to hotspots all through the system.
As we have beforehand reported, this method is significantly extra environment friendly than conventional air cooling. Google truly makes use of DLC to chill its Tensor Processing Items (TPUs) utilized in AI coaching and inference workloads. Nonetheless, many fashionable GPUs techniques are nonetheless air cooled. Regardless of which expertise is used, advanced air-handling and thermal-management techniques are required to get the warmth out of the datacenter.
Some of the environment friendly methods to do that is utilizing evaporative coolers, like cooling towers. This system makes use of water to drag the warmth out of the air exiting the datacenter. As the new air causes the water to evaporate, it’s chilled again to a usable temperature. Evaporative cooling tech is sort of widespread amongst datacenter operators because it tends to make use of much less energy than different tech and in lots of climates they solely must be run through the hottest months of the yr.
This method to cooling could be problematic in sure areas the place entry to water or therapy amenities is proscribed. In datacenter hubs similar to Phoenix, Arizona, many datacenters have transitioned to different cooling applied sciences, which do not devour water immediately however do have a tendency to make use of extra energy and generate extra noise air pollution.
How a lot water are we speaking about?
Microsoft says it used about 1.6 million cubic metres extra water in 2022 than in 2021. To place that in perspective that is sufficient water to fill 640 Olympic sized swimming swimming pools; a quantity equal to greater than three billion grapefruit. And that is simply new water consumption. Microsoft’s whole water consumption for 2022 is the equal of two,560 Olympic swimming pools.
In 2021, Researchers on the College of Oxford estimated that US datacenters collectively consumed about 1.7 billion litres of water per day. Nonetheless, they be aware that measuring datacenter water consumption is commonly challenged by an absence of transparency.
Regardless of this, researchers on the College of California, Riverside and the College of Texas at Arlington just lately tried to find out simply how a lot water generative AI was utilizing.
“ChatGPT must ‘drink’ a 500ml bottle of water for a easy dialog of roughly 20-50 questions and solutions, relying on when and the place ChatGPT is deployed,” the researchers estimated in an April paper.
Nonetheless, as we have beforehand mentioned, precise water consumption related to operating LLMs, like ChatGPT, will depend on quite a lot of elements together with the thermal administration applied sciences utilized by the ability, the place these fashions are being skilled and run, and when these jobs are run.
For instance, datacenters situated in a colder climates are prone to devour much less water and may make the most of decrease ambient temperatures, particularly in comparison with these in scorching, arid desert climes. In response to the AP report, the Iowa datacenter the place Microsoft and OpenAI skilled GPT-4 solely consumes water when the temperature is above 29.3 levels Celsius.
What’s Microsoft doing to mitigate its water consumption?
In its ESG report, Microsoft claims it is utilizing water consumption metrics to information water reclamation efforts because it seeks to be internet “water constructive” by 2030.
Initiatives to succeed in that elusive “water constructive” state usually contain funding tasks to guard watersheds, restore wetlands, and enhance infrastructure in water confused areas. For instance, in Chennai, India, the place Microsoft operates a datacenter, the corporate labored with The Pure Conservancy to revive the wetlands round Lake Sembakkam.
Up to now, Microsoft says it is signed 35 million cubic metres of water reclamation tasks.
The Home windows maker additionally says it is engaged on thermal administration applied sciences to scale back the water consumption of its amenities. This features a geoexchange system at its Thermal Power Middle in Redmond, which rejects warmth to the bottom slightly than counting on cooling towers. The expertise is anticipated to scale back that facility’s water consumption by about 30,280 cubic metres — or about 12 Olympic swimming swimming pools.
Microsoft has additionally moved away from evaporative coolers at a few of its most water challenged datacenters. This spring, they agreed to transition its closing two datacenters, situated outdoors Phoenix to “zero water” cooling infrastructure.
How do the opposite cloud suppliers stack up?
Microsoft is not the one cloud supplier grappling with water consumption. As Amazon Net Companies (AWS) and Google Cloud have steadily expanded their cloud empires, they’ve introduced a number of efforts to curb their datacenter consuming drawback.
Final fall AWS, dedicated to changing into “water constructive” by 2023, which is to say, they declare they’re going to return extra water to communities than is consumed by their operations.
Google has made related commitments, however in April admitted AI adoption was making the scenario tougher. ®
