Ummm... This chart is TERRIFYING. Have the data centers that undergird AI already triggered this much of a shift in energy consumption? ::shudder:: https://www.washingtonpost.com/business/2024/03/07/ai-data-centers-power/
@zephoria more likely that the majority of new demand comes from the electrification transformation of the economy - toward electric vehicles, electric stoves and ovens, etc. Yes, some of this is from data centres, but electrification has long been predicted to require 2.5x - 3x the current electric generation to replace fossil fuel generated uses. c.f. Saul Griffith's work on this
@mpesce @zephoria not very likely as that process is far more gradual. Also I think most of us haven’t come to grips with just how extravagant the demand is from GPUs of the new class. See: https://infosec.exchange/@dymaxion/112060242911196648
@zephoria
And here I was hoping like a fool that this is electrification of heating and transportation driving demand
@zephoria What’s even more terrifying the amount of high energy AI algorithms will be used to solve well known / easy to solve problems.
@zephoria I am pretty sure this is wrong - maybe missing data in earlier sources, maybe a change in the measurement/collection methodologies? I'm planning to dig into the underlying data sources soon, but a national doubling of energy consumption in 12 months just seems outside the realm of reasonableness to me. Just - how would the supply networks even handle that? We have trouble with holiday weekends and 3-day heat waves!
@playinprogress @zephoria Ahh - I can read it that way, too. That would be less unreasonable, but still wild! I need to dig more...
@playinprogress @zephoria Finally had a moment to dig into this!
Some key points I've learned:
- Yes, this shows change in growth, not total demand. Nice catch!
- The rate does double, but only to 1.2% 10-year growth
- This does reverse a ≈10-year trend, but over 30 years, it's actually still historically low!
- The NERC report attributes almost all of the forecasted growth to two areas: EV charging and electrification of heating in traditionally oil/gas heat in regions.
@zephoria
Assuming the excess over previous years was entirely ML, that corresponds to about 400,000 H100-class GPUs (and associated hardware), if my math is right, which is within what Nvidia is shipping.
@zephoria my whole thing with AI is pretty much 'Havent you guys seen at least the first 2 Terminator films?' I know its a simple argument against it but I feel its all I need lol AI sucks, I hate it . It makes me nauseous when I look at it.
@zephoria This conversation with @jgkoomey is a good counterpoint:
https://heatmap.news/podcast/shift-key-episode-10-ai-data-centers