Comment by phs318u
17 hours ago
But what do you do with all the waste energy? All those MW and GW have to end up somewhere and radiation into a vacuum is the hardest way to dump heat.
17 hours ago
But what do you do with all the waste energy? All those MW and GW have to end up somewhere and radiation into a vacuum is the hardest way to dump heat.
At 70-80C (working temp of silicon chips) 1m2 radiates 700-800W, i.e. the heat of 1 GPU like H200 without any need for any cooling equipment beside the radiator itself( and may be some dumb heatpiping) . To acquire that energy you'd need 3-4m2 of solar panels. So a datacenter would be a large field of solar panels with a smaller field of heat radiators in their shadow.
To the commenter below: yes, exactly, this is where my thinking on that started at the cryptocurrency boom - https://news.ycombinator.com/item?id=26289423 - as you don't need close connection between mining GPUs. For AI you'd need to cluster several together while still overall scheme is the same.
>what the equilibrium temperature of a black planar surface is at a given distance from the sun.
it is 120C at the Earth orbit. So you do need to have some reflection, either back through the solar panels, or the radiators to have a reflective back toward the solar panels in the shadow of which they are to be located.
You can probably (I haven't verified this) omit separate radiators and just use the back of the solar panels. Effectively you're describing mounting each H200 to the back of a 4 m^2 solar array at which point I suspect the equilibrium temperature will fall within an acceptable range. In fact the H200 and electricity are both entirely irrelevant here - the core question is what the equilibrium temperature of a black planar surface is at a given distance from the sun.
Would it be feasible to put several JWST-like stirling engines somewhere in the mix to use up some of that heat and turn it into some kind of useful energy? ....
Perhaps running pumps that move around coolant passing over the cubes of GPUs? ..
That would be extra weight/cost into orbit though...
Also, don't solar panels have reduced efficiency when they're hot? And having anything hot surely increases failure rates.. with metals getting closer to melting points...?
1 reply →