Trying to educate them that water and power usage are the less problematic parts and not as bad as everyone thinks. It's the noise and lack of green power that are the problems.
Also can you tell us which chemicals need to be added to the outer and inner loops of the systems, and do the inner loop has the same losses or it is different.
If there's water loss into the ground, and if the water is treated with chemicals, is it required or is it in fact treated to remove these chemicals so the used water can be reused for livestock, agriculture, or non-consumable use in the cities like for watering greenery or other processes which needs water but can accept non-potable variety.
I also heard that there's a closed loop version for outer loop. Is it used widely in US, if yes, what percentage of the large datacenters use it? Because as I heard it, it's more expensive, but doesn't lose water as much.
Adding some references if possible is greatly appreciated so we can dive deeper, as the community.
Same is also valid for power usage metrics. I'm especially interested in how a datacenter full of 800VDC systems (aka Megawatt Rack) can be fed without any difficulty with our current power generation and transfer infra.
If it doesn't, we can all stand around and make fun of the people who paid terabucks for the data centers, and enjoy a bountiful harvest of surplus hardware for pennies on the dollar.
OTOH, considering the proliferation of the EVs, greenhouse gas emissions from cars are going down, alongside brake-pad related emissions since generative braking covers 90%+ of the daily braking needs.
OTOH, many datacenters are installing their own greenhouse emitting power gensets. So, with a bitter contrast, as AI is replacing jobs, it's also filling back the carbon footprint vacated by cars and more efficient modes of transport at the same time.
These GPUs won't be scrapped soon, because besides AI, they are useful for accelerating not-so-precise floating point calculations en-masse. So, no, it won't work as you imagine.
And here we are, talking about the US. Great, that's settled.
OTOH, considering the proliferation of the EVs, greenhouse gas emissions from cars are going down
In the US, most electricity comes from carbon-based sources. EVs help, certainly, but it would still be best from an ecological standpoint if people didn't have to use them every day to get to work.
OTOH, many datacenters are installing their own greenhouse emitting power gensets.
Exactly, which is why we should work for grid renewal and expansion at the same time we build these things out. If the hyperscalers insist on building their own power plants, those power plants should be required to be (a) carbon-neutral; and (b) grid-tied.
These GPUs won't be scrapped soon, because besides AI, they are useful for accelerating not-so-precise floating point calculations en-masse.
Really. What are some other applications for not-so-precise floating point calculations at the exaflops level?
Proliferation of law skirting AI datacenters is a worldwide problem. Many companies are extending this to Europe as well, so AI datacenters is not a U.S. problem only.
From what I see in Electricity Maps[0], it's not clear cut. Some of the states have well over 50% of their generation is from carbon-free sources. As a general trend, west is greener than eastern US. So, no, EVs will have an impact there. However, you need electric semis a lot as a country since rail is not used a lot in the US as well.
> Really. What are some other applications for not-so-precise floating point calculations at the exaflops level?
In short, HPC. As a more detailed, by short list:
- Weather and climate modeling.
- Drug discovery.
- Computational Fluid Dynamics.
- Material science, behavioral and process analysis.
- All kinds of large scale simulation which can be modeled with matrices.
- Any kind of problem requiring linear algebra.
I manage HPC clusters, and sometimes develop code on them, so yeah.
[0]: https://app.electricitymaps.com/map/zone/US-NW-PACE/live/fif...
(IMHO, the fact that AI models tolerate absurdly-low precision means that we're doing something wrong and wasteful in our entire computational approach to the field, but that's neither here nor there.)
Except when some rich clown from another country comes into a drought-stressed desert area where residents have already been under increasingly draconic water usage restrictions for goin' on a decade now, wantin' to build a datacenter right on a lake that's already under severe stress with record low water levels, tellin' everyone it won't have any effect on anything and it's all just rainbows and sunshine and nothin' but benefits for everyone all around (lies). But yeah, no worries. We can surely trust the billionaire wouldn't lie to us to enrich himself at our expense.