To quench AI's thirst, the way we build, operate datacenters needs to change

Generative models like GPT-4, Midjourney, and others have many understandably concerned about the technology's potential to disrupt society, but it's quenching AI's seemingly insatiable thirst that has researchers at the University of California Riverside and the University of Texas Arlington attention. In a paper released [ …

[Duncan Macdonald]

Look North !!

If the datacenters are located in cold locations (Alaska, North Canada, Iceland etc) then cooling becomes trivial. (Iceland also has the advantage of cheap geothermal electricity.)

(Unfortunately there is no icon for freezing !!!)

[Yet Another Anonymous coward]

Re: Look North !!

And what's better you have reliable power, good communications infrastructure and low latency links to make centers of population.

[Dimmer]

Re: Look North !!

Here is some real numbers:

Using outside air at 30deg @ 1200cfm requires 32 gal of water a day to keep the humidity up to %35 and it replaces 3 tons of cooling.

Cisco ASA 9k minimum operating requirement is %7 humidity. Moisture = heat transfer. So not humidifying is not an option.

Question, when you evaporate water for cooling does it not become water later? Another words, it rains down wind from the dc or at least seeds the clouds?

Or, evaporate sea water and capture the humidity as it condenses again and use the water for farming.

[Yet Another Anonymous coward]

Re: Look North !!

>Question, when you evaporate water for cooling does it not become water later? Another words, it rains down wind from the dc or at least seeds the clouds?

But that's in a different zoning district so doesn't count.

[Michael Wojcik]

Re: Look North !!

The problem is data centers in daft locations like Phoenix, or elsewhere in the Great American Desert (as it used to be known), using up scarce freshwater resources. If that water then precipitates down over wetter parts of the country, or the ocean, that's still a big problem.

Using seawater would avoid this, but 1) you'd have to build data centers near the sea, rather than in Arizona¹ or Texas² or other terrible choices; and 2) evaporating seawater leaves behind all sorts of tricky stuff to deal with, particularly salt.

As for "captur[ing] the humidity": That would require either a great deal of additional energy input to move the heat somewhere else, or magical engineering. That steam will condense when it gets high enough in the atmosphere to dump its heat through radiation. How big a cooling tower do you think you can build? Again, that water will return to earth somewhere far away. (And, no, a bunch of data centers along the Pacific coast of the US will not put enough steam into the atmosphere to make a difference to the drought.)

¹Seriously, Arizona. What, was the Sahara not available?

²Where you can have simultaneous water and energy crises. Yay!

[mpi]

Same story as with datacenter energy consumption.

As long as our society doesn't seem to have too many problems with the fact that flotillas of 2t SUVs burn ungodly amounts of fossile fuel to transport people in singles and pairs, when public transport extension would be a much more energy efficient system, datacenter energy consumption is probably the least of our energy efficiency problems.

And as long as our society doesn't seem to have a problem with wasting 50,000 liters of drinking water to produce 1kg of beef, when the same amount could produce many more edible calories in plants and insects and poultry, datacenter water consumption is probably the least of our water concerns.

Or to put this another way: I'm worried more about the giant dumpster full of old tires burning, than the smoke of my neighbors scented candle.

[Lomskij]

A bit of nitpicking

Can you please stop posting invalid comparisons like this:

"For reference, Submer and LiquidStack, two immersion-cooling vendors, often tout PUE ratings of less than 1.05 — making them far more efficient than typical air-cooled datacenters which usually come in at 1.4-1.5."

PUE of 1.05 is efficiency of an immersion tank in isolation, while data centre PUE of 1.50 is the efficiency of the whole data centre, including power transmission losses, CRAH units, chillers, etc.

If you're comparing PUE (or pPUE to be precise) of a Submer immersion tank of 1.05 on average against a typical air-cooled OCP rack at 1.10, then your total data centre PUE will drop from 1.50 to 1.45, as power loss and chiller consumption is not affected - if your servers need 10MW, it needs to be transmitted and then vented into the air regardless of the heat collection method.

Back to the topic - lots of countries are now banning the use of adiabatic coolers outright, try to get a permission to build an evaporative cooled data centre in the UK and you'll see what I mean. Even some states in the US are doing that - increasingly large number of new data centres are either being built further north, or adopting new solutions like hybrid chillers with free cooling etc. I would take these AI water consumption figures with a big grain of salt.