6 ms·

I recall current data center electricity consumption to be 2-3% of total electrical use. So at this scale, these are ultimately PR moves/engineering experiments

I recall current data center electricity consumption to be 2-3% of total electrical use. So at this scale, these are ultimately PR moves/engineering experiments that don't actually save money as a whole. However, with LLMs coming, I could see data center energy usage jump 10x over the next decade. At that point, re-using data center heat will be be a real industry with real profits. Imagine data center heated apartment heating, vertical farms. I don't know how hotter regions are going to cope with the heat though. These AI datacenters could end up being entirely concentrated in the coldest regions.

ggm 4y ago
You're conflating 2% at world energy totals with $cost to anyone able to harvest the megawatts of heat energy close by. At scale, this happens worldwide in shopping malls already and the loadshedding demand management in HVAC and chiller room is immensely valuable. Sure. It's only 2%. Two percent of Musks net worth would fix homelessness in San Francisco applied annually. [edit: Actually, it's 2x too much] 2% of cost savings on inputs to a product could translate to a 50% increase in net profit. "oh well. its only 2% so why bother..."
- screye 4y ago
  > net worth would fix homelessness in San Francisco I know this wasn't the point of your analogy. But the ill-conceived idea that money can fix homelessness in San Francisco is exactly the reason why we got here in the first place. Richest city in the world needs more money to fix simple problems, says blameless local govt.
  - ggm 4y ago
    My point is that the "2% is not useful" statement, applies the wrong logic to the problem at hand. in fact, 2% of global energy is a significant single-change outcome, irrespective. I used the homelessness problem because it has a $cost at scale which people see as intractable. The wisdom of fixing it by throwing money at it, rather than throwing money at lawyers to overturn the stupid planning issues is a different question. Actually, throwing $1.5b to the homeless would probably work pretty well.
    - mrDmrTmrJ 4y ago
      Nope - let's do the math: Cost of a new affordable housing unit in SF ~$1.17M. [1] Number of homeless people $1.5B could house ~1,282. Number of homeless people in SF 7,754 people in 2022 [2]. So $1.5B could only solve ~16.5% of SF's housing problem or leave 84% unsolved. Which shows homelessness really is a "housing cost" problem in SF :) [1] https://www.sfchronicle.com/sf/article/It-now-costs-more-than-1-2-million-to-build-a-17463355.php https://www.sfchronicle.com/sf/article/It-now-costs-more-tha... [2]https://sfgov.org/scorecards/safety-net/homeless-population https://sfgov.org/scorecards/safety-net/homeless-population
      - ggm 4y ago
        "applied annually" plus, solving 16% of the shortage of housing supply would be bigger than any other intervention at scale: it would probably force downward pricing in rent and sales overall. https://sfstandard.com/public-health/homelessness/ending-homelessness-in-san-francisco-will-cost-1-4-billion-city-says/ https://sfstandard.com/public-health/homelessness/ending-hom... ...Ending unsheltered homelessness would require vast increases in housing, shelter and prevention services, according to the report from the city’s Department of Homelessness and Supportive Housing and homelessness consultancy firm Focus Strategies. That would cost the city another $1.45 billion over three years. And it would take over $410 million each year after that to keep the programs running. ...
      - jayparth 4y ago
        So it’d be mostly solved over a 5 year period….?!?’
canadianfella 4y ago
[dead]
candiddevmike 4y ago
Wonder if it will only be a 5x increase if those LLMs and other workloads run on ARM or RISC =D
web3-is-a-scam 4y ago
Will it though? Do we have any emerging industries re-using datacenter heat from Bitcoin mining operations?
eldaisfish 4y ago
Hotter regions don't pose a significant challenge. The specific heat of water is pretty high and swimming pools can store significant energy in their water. I don't remember the exact number but raising the temperature of an olympic sized swimming pool by a degree celsius takes megawatt hours of energy. that's a lot and a datacenter is perfect for this, even in hot climates.
nl 4y ago
> However, with LLMs coming, I could see data center energy usage jump 10x over the next decade. Just a note: we are probably right now at peak inefficiency for LLMs. There are obvious improvements available but they will only be explored by organizations which have different time vs resources trade-off calculations compared to OpenAI and Google. Right now OpenAi and Google are racing to ship products. A working training framework is more important than saving $X (where X is a very large number) because capturing marketshare is seen as vital and they have no shortage of capital. But the massive and rapid progress seen by projects like llama-cpp show how much performance they are leaving on the table. (And yes this also applies at training time).
yieldcrv 4y ago
“good enough” LLMs are going to go client side in like a month and be on flagship smartphones in a year others are going to go onprem I think this might already be near peak data center LLM, the use case for the most genius cloud heavy LLMs is going to be small once it gets baked into OS layer
csomar 4y ago
Energy transfer will be a thing. Hotter climate regions export their sun. Colder climate regions export their compute.

ggm 4y ago

You're conflating 2% at world energy totals with $cost to anyone able to harvest the megawatts of heat energy close by. At scale, this happens worldwide in shopping malls already and the loadshedding demand management in HVAC and chiller room is immensely valuable. Sure. It's only 2%. Two percent of Musks net worth would fix homelessness in San Francisco applied annually. [edit: Actually, it's 2x too much] 2% of cost savings on inputs to a product could translate to a 50% increase in net profit. "oh well. its only 2% so why bother..."

screye 4y ago
> net worth would fix homelessness in San Francisco I know this wasn't the point of your analogy. But the ill-conceived idea that money can fix homelessness in San Francisco is exactly the reason why we got here in the first place. Richest city in the world needs more money to fix simple problems, says blameless local govt.
- ggm 4y ago
  My point is that the "2% is not useful" statement, applies the wrong logic to the problem at hand. in fact, 2% of global energy is a significant single-change outcome, irrespective. I used the homelessness problem because it has a $cost at scale which people see as intractable. The wisdom of fixing it by throwing money at it, rather than throwing money at lawyers to overturn the stupid planning issues is a different question. Actually, throwing $1.5b to the homeless would probably work pretty well.
  - mrDmrTmrJ 4y ago
    Nope - let's do the math: Cost of a new affordable housing unit in SF ~$1.17M. [1] Number of homeless people $1.5B could house ~1,282. Number of homeless people in SF 7,754 people in 2022 [2]. So $1.5B could only solve ~16.5% of SF's housing problem or leave 84% unsolved. Which shows homelessness really is a "housing cost" problem in SF :) [1] https://www.sfchronicle.com/sf/article/It-now-costs-more-than-1-2-million-to-build-a-17463355.php https://www.sfchronicle.com/sf/article/It-now-costs-more-tha... [2]https://sfgov.org/scorecards/safety-net/homeless-population https://sfgov.org/scorecards/safety-net/homeless-population
    - ggm 4y ago
      "applied annually" plus, solving 16% of the shortage of housing supply would be bigger than any other intervention at scale: it would probably force downward pricing in rent and sales overall. https://sfstandard.com/public-health/homelessness/ending-homelessness-in-san-francisco-will-cost-1-4-billion-city-says/ https://sfstandard.com/public-health/homelessness/ending-hom... ...Ending unsheltered homelessness would require vast increases in housing, shelter and prevention services, according to the report from the city’s Department of Homelessness and Supportive Housing and homelessness consultancy firm Focus Strategies. That would cost the city another $1.45 billion over three years. And it would take over $410 million each year after that to keep the programs running. ...
    - jayparth 4y ago
      So it’d be mostly solved over a 5 year period….?!?’

canadianfella 4y ago

[dead]

candiddevmike 4y ago

Wonder if it will only be a 5x increase if those LLMs and other workloads run on ARM or RISC =D

web3-is-a-scam 4y ago

Will it though? Do we have any emerging industries re-using datacenter heat from Bitcoin mining operations?

eldaisfish 4y ago

Hotter regions don't pose a significant challenge. The specific heat of water is pretty high and swimming pools can store significant energy in their water. I don't remember the exact number but raising the temperature of an olympic sized swimming pool by a degree celsius takes megawatt hours of energy. that's a lot and a datacenter is perfect for this, even in hot climates.

nl 4y ago

> However, with LLMs coming, I could see data center energy usage jump 10x over the next decade. Just a note: we are probably right now at peak inefficiency for LLMs. There are obvious improvements available but they will only be explored by organizations which have different time vs resources trade-off calculations compared to OpenAI and Google. Right now OpenAi and Google are racing to ship products. A working training framework is more important than saving $X (where X is a very large number) because capturing marketshare is seen as vital and they have no shortage of capital. But the massive and rapid progress seen by projects like llama-cpp show how much performance they are leaving on the table. (And yes this also applies at training time).

yieldcrv 4y ago

“good enough” LLMs are going to go client side in like a month and be on flagship smartphones in a year others are going to go onprem I think this might already be near peak data center LLM, the use case for the most genius cloud heavy LLMs is going to be small once it gets baked into OS layer

csomar 4y ago

Energy transfer will be a thing. Hotter climate regions export their sun. Colder climate regions export their compute.