I just had "the talk" with my 9yo.
-
@Orb2069 @momo Great, a lovely article with no reference to what hardware its talking about except "AI data centers" as a label, truly authorative.
No .. Virtually all AI focussed chips, at least the ones NVIDIA makes, are either air cooled or in closed water loops.
You can look up their spec sheets, they're public. -
@Orb2069 @momo Slightly more read self-correction: Looks like you're right for most already installed capacity with older gen chips, they run air cooled chips with evaporative cleaning to keep the buildings themselves cool, thanks!
newer gen installs with newer chips are switching towards closed water loops though (the GBs I mentioned earlier) and the facilties water loop seem to only "very rarely" (.. I guess where the climate is too hot) need evaporative cooling there.
-
@Orb2069 @momo Slightly more read self-correction: Looks like you're right for most already installed capacity with older gen chips, they run air cooled chips with evaporative cleaning to keep the buildings themselves cool, thanks!
newer gen installs with newer chips are switching towards closed water loops though (the GBs I mentioned earlier) and the facilties water loop seem to only "very rarely" (.. I guess where the climate is too hot) need evaporative cooling there.
-
-
@momo they should do this to Government ministers to explain why they shouldn't throw taxpayers money at this hustler. Most of them should get it....most of them.
-
@momo Except none of it us true ... evaporative cooling is almost entirely unused in the AI industry because it doesn't work.
Every NVIDIA chip (which is like, almost the entire industry at this point) uses either air cooling with giant fan walls, or closed loop water cooling with heat exchangers (For GB200-GB300s)
Not arguing the business model of large AI labs though ...
-
I just had "the talk" with my 9yo. He came to me asking "Dad, why does AI use so much water just for generating one answer?"
So I explained to him how LLM works, how GenAI came out of that and why AI is a marketing term and absolutely wrong in that context.
Then I loaded GPT4All on my laptop and asked the thing what the result of 9+7 is. It took my poor laptop cpu 12 seconds to come up with an answer and another 12 seconds to generate the sentence with the answer. I showed him how fast kcalc (using the CPUs ALU functions) comes up with the answer for that specific question.
And then I asked a followup question, which almost take 2 minutes to compute. I showed him how the previous question and its answer became part of the prompt which made the calculation longer more complicated just to emulate a dialoge.
He noticed that my laptops CPU fan stepped up and started to roar to get rid of the core heat. Then I explained to him how AI datacenters solve all this ineffitiency by running millions of servers in parallel and how this generates so much heat that it "uses" cooling water by just boiling it until it vapourizes.
I then explained the current business model and how the companies currently burn through their cash to be the last one standing on the market when everyone else dropped out or became bankrupt. What we user pay for using this and what we actually would have to (and in the future probably will) pay for this service.
Needless to say, my kid is not a big fan now.
-
I just had "the talk" with my 9yo. He came to me asking "Dad, why does AI use so much water just for generating one answer?"
So I explained to him how LLM works, how GenAI came out of that and why AI is a marketing term and absolutely wrong in that context.
Then I loaded GPT4All on my laptop and asked the thing what the result of 9+7 is. It took my poor laptop cpu 12 seconds to come up with an answer and another 12 seconds to generate the sentence with the answer. I showed him how fast kcalc (using the CPUs ALU functions) comes up with the answer for that specific question.
And then I asked a followup question, which almost take 2 minutes to compute. I showed him how the previous question and its answer became part of the prompt which made the calculation longer more complicated just to emulate a dialoge.
He noticed that my laptops CPU fan stepped up and started to roar to get rid of the core heat. Then I explained to him how AI datacenters solve all this ineffitiency by running millions of servers in parallel and how this generates so much heat that it "uses" cooling water by just boiling it until it vapourizes.
I then explained the current business model and how the companies currently burn through their cash to be the last one standing on the market when everyone else dropped out or became bankrupt. What we user pay for using this and what we actually would have to (and in the future probably will) pay for this service.
Needless to say, my kid is not a big fan now.
@momo
"The talk" can mean different things to different people in different circumstances, and you've given us one more good meaning for it. In the United States, today is Father's Day. Happy Father's Day to you! -
@momo
"The talk" can mean different things to different people in different circumstances, and you've given us one more good meaning for it. In the United States, today is Father's Day. Happy Father's Day to you!@tom
I know, it mostly means talking to your kids about sex and how making babies works, unless they are PoC. Then the talk is about how to interact with cops and getting home back alive. -
@otte_homan @suriele @Orb2069 @momo
Yup. Whereas I have actually worked in data centres. Of course it is entirely possible to design a data centre without evaporative cooling, but… capitalism. -
I just had "the talk" with my 9yo. He came to me asking "Dad, why does AI use so much water just for generating one answer?"
So I explained to him how LLM works, how GenAI came out of that and why AI is a marketing term and absolutely wrong in that context.
Then I loaded GPT4All on my laptop and asked the thing what the result of 9+7 is. It took my poor laptop cpu 12 seconds to come up with an answer and another 12 seconds to generate the sentence with the answer. I showed him how fast kcalc (using the CPUs ALU functions) comes up with the answer for that specific question.
And then I asked a followup question, which almost take 2 minutes to compute. I showed him how the previous question and its answer became part of the prompt which made the calculation longer more complicated just to emulate a dialoge.
He noticed that my laptops CPU fan stepped up and started to roar to get rid of the core heat. Then I explained to him how AI datacenters solve all this ineffitiency by running millions of servers in parallel and how this generates so much heat that it "uses" cooling water by just boiling it until it vapourizes.
I then explained the current business model and how the companies currently burn through their cash to be the last one standing on the market when everyone else dropped out or became bankrupt. What we user pay for using this and what we actually would have to (and in the future probably will) pay for this service.
Needless to say, my kid is not a big fan now.
-
@brezelradar
Sorry aber bei den aktuellen Temperaturen stell ich mich nicht mit nem grünen Pulli vor die Kamera!
@echopapa -
J jwcph@helvede.net shared this topic
