AI Needs So Much Power, It's Making Yours Worse
bloomberg.com
bloomberg.com
Better power supplies can help with the distortion, but it really shouldn't matter much, except in extreme cases where things are at 99% of their limits in the power grid, on a hot day.
By the way, you can learn interesting things from analysis of the power grid. Long ago I remember a Slashdot comment about using it with covertly gathered recordings estimate the effective yield of nuclear enrichment operations in the middle east. I've been trying to find that thread but Google isn't what it used to be
[1] https://www.whiskerlabs.com/analysis-of-total-harmonic-disto...
https://www.researchgate.net/publication/228873835_Estimatio...
Previous:
https://news.ycombinator.com/item?id=42523401 - Dec 2024 (3 comments)
Maybe there's an alternative design for AC->DC conversion that can use the full cycle?
No, large power supplies used for AC/DC conversion don’t do this. They even already do power factor correction to correct the AC power factor.
The issue is that power factor can only be corrected so well (0.95 is typical best case), so if you keep adding more and more loads of the same type it will lead to excessive capacitive loading on a grid.
In comparison, industrial stuff (things that use lots of motors) have an inductive loading on the grid. Basically, if you have equally paired inductive and capacitive loads you can cancel out the “imbalance”. Industrial plants already do this with large capacitor banks.
Essentially all this means is that the power companies probably need to start forcing power factor requirements on data center customers. They’re already billed for it, but if it’s causing grid issues they can require it (or, implicitly dictate it through pricing structures).
In the unhappy case of needing to generate power using fossil fuels, grid scale equipment like combined cycle gas turbines will bury the efficiency and emissions of any typical DC class generation capacity.