radiators can be made as long as desirable within the shade of the solar panels, hence the designer can pracitically set arbitrarily low temperatures above the background temperature of the universe.
Yes, you can overcome this with enough radiator area. Which costs money, and adds weight and space, which costs more money.
Nobody is saying the idea of data centers in space is impossible. It's obviously very possible. But it doesn't make even the slightest bit of economic sense. Everything gets way, way harder and there's no upside.
I don't think dissipating heat would be an issue at all. The cost of launch I think is the main bottleneck, but cooling would just be a small overhead on the cost of energy. Not a fundamental problem.
Either that or your talking out of your ass.
FYI a single modern rack consumes twice the energy of the entire ISS, in a much much much much smaller package and you'll need thousands of them. You'd need 500-1000 sqm of radiator per rack and that alone would weight several tonnes...
You'll also have to actively cool down your gigantic solar panel array
"Radiators can shadow each other," this is precisely why I chose a convex shape, that was not an accident, I chose a pyramid just because its obvious that the 4 triangular sides can be kept in the shade with respect to the sun, and their area can be made arbitrarily large by increasing the height of the pyramid for a constant base. A convex shape guarantees that no part of the surface can appear in the hemispherical view of any other part of the surface.
The only size limit is technological / economical.
In practice h = 3xL where L was the square base side length, suffices to keep the temperature below 300K.
If heat conduction can't be managed with thermosiphons / heat pipes / cooling loops on the satellite, why would it be possible on earth? Think of a small scale satellite with pyramidal sats roughly h = 3L, but L could be much smaller, do you actually see any issue with heat conduction? scaling up just means placing more of the small pyramidal sats.
but you'd rarely ever need it though: it just needs to rotate at a low angular velocity of 1 rotation per year to keep facing the sun.
No need to apply at NASA, to the contrary, if you don't believe in Stefan Boltzmann law, feel free to apply for a Nobel prize with your favorite crank theory in physics.
Without eventually moving compute to space we are going to have compute infringe on the space, energy, heat dissipation rights of meatbags. Why welcome that?!?
Sure, it occurs, but what does the Stefan–Boltzmann law tell us about GPU clusters in space?
In space or vacuum radiation is the best way to dissipate heat, since it's the only way.
I believe the reason the common person assumes thermal radiation is a very poor way of shedding heat is because of 2 factoids commonly known:
1. People think they know how a vacuum flask / dewar works.
2. People understand that in earthly conditions (inside a building, or under our atmosphere) thermal radiation is insignificant compared to conduction and convection.
But they don't take into account that:
1) Vacuum flasks / dewars use a vacuum for thermal insulation. Yes and they mirror the glass (emissivity nearer to ~0) precisely because thermal radiation would occur otherwise. They try their best to eliminate thermal radiation, a system optimized to eliminate thermal radiation is not a great example of how to effectively use thermal radiation to conduct heat. The thermal radiation panels would be optimized for emissivity 1, the opposite of whats inside the vacuum flask.
2) In a building or under an atmosphere a room temperature object is in fact shedding heat very quickly by thermal radiation, but so are the walls and other room temperature objects around you, they are reheating you with their thermal radiation. The net effect is small, in these earthly conditions, but in a satellite the temperature of the environment faced by the radiating surfaces is 4K, not a temperature similar to the object you are trying to keep cool.
People take the small net effect of thermal radiation in rooms etc, and the slow heat conduction through a vacuum flasks walls as representative for thermal radiation panels facing cold empty space, which is the mistake.
I provided the calculation for the pyramidal shape: if the base of a pyramid were a square solar panel with side length L, then for a target temperature of 300K (a typical back of envelope substitute for "room temperature") the height of the pyramid would have to be about 3 times the side length of the square base. Quite reasonable.
> Sure, it occurs, but what does the Stefan–Boltzmann law tell us about GPU clusters in space?
The Stefan-Boltzmann law tells us that whatever prevents us from putting GPU clusters in space, it's not the difficulty in shedding heat by thermal radiation that is supposedly stopping us.
If the base were a solar panel aimed perpendicular to sun, then the tip is facing away and all side triangles faces of the pyramid are in the shade.
I voluntarily give up heat dissipation area on 2 of the 4 triangular sides (just to make calculations easier, if we make them thermally reflective -emissivity 0-, we can't shed heat, but also don't absorb heat coming from lukewarm Earth).
The remaining 2 triangular sides will be large enough that the temperature of the triangular panels is kept below 300 K.
The panels also serve as the cold heat baths, i.e. the thermal sinks for the compute on board.
Not sure what you mean with wings, I intentionally chose a convex shape like a pyramid so that no part of the surface of the pyramid can see another part of the surface, so no self-obstruction for shedding heat etc...
If this doesn't answer your question, feel free to ask a new question so I understand what your actual question is.
The electrical power available for compute will be approximately 20% (efficiency of solar panels) times the area of the square base L ^ 2 times 1360 W / m ^ 2 .
The electrical power thus scales quadratically with the chosen side length, and thus linearly with the area of the square base.
Just look at a car. Maybe half a square meter of “radiator” is enough to dissipate hundreds of kW of heat, because it can dump it into a convenient mass of fluid. That’s way more heat than the ISS’s radiators handle, and three orders of magnitude less area.
Or do a simple experiment at home. Light a match. Hold your finger near it. Then put your finger in the flame. How much faster did the heat transfer when you made contact? Enough to go from feeling mildly warm to causing injury.
Also this assumes a flat surface on both sides. Another commenter in this thread brought up a pyramid shape which could work.
Finally, these gpus are design for earth data centers where power is limited and heat sinks are abundant. In the case of space data centers you can imagine we get better radiators or silicon that runs hotter. Crypto miners often run asics very hot.
I just don't understand why every time this topic is brought up, everyone on HN wants to die on the hill that cooling is not possible. It is?? the primary issue if you do the math is clearly the cost of launch.
its not exactly good advertisement for conductive or convective heat transfer if its really employing thermal radiation under the hood!
but do you want big tech to shit where you eat? or do you want them to go to the bathroom upstairs?
At some point I'm thinking the large resistance to the idea I am seeing in a forum populated with programmers is the salivation-inducing idea that all that datacenter hardware will eventually get sold for less and less, but if we launch them to space there won't be any cheap devalued datacenter hardware to put in their man-caves.
they could go near a Lagrange point
there are so many options
heavier boats are also slower to accelerate or decelerate compared to smaller boats, does this mean we should ban container ships? having special orbits for megastructure lanes would seem a reasonable approach.
My example is optimized not for minimal radiator surface area, but for minimal mathematical and physical knowledge required to understand feasibility.
Your numbers are different because you chose 82 C (355 K) instead of my 26 C (300 K).
Near normal operating temperatures hardware lifetime roughly doubles for every 10 deg C/K decrease in temperature (this does not hold indefinitely of course).
You still need to move the heat from the GPU to the radiator so my example of 26 deg C at the radiator just leaves a lot of room against criticism ;)
However, what do you reckon the energy balance is for launching the 1 GW datacenter components into space and assembling it?
You can prove that the lower efficiency can be managed, and they will still say the only thing they know: "Thermal radiation is not efficient".
The economics and energy balance is where I too am very skeptical, at least near term.
Quick back of envelope calculations gave me a payback time of about 10 years, so which is only a single order of magnitude off which can easily accumulate by lack of access to detailed plans.
I can not exclude they see something (or can charge themselves lower launch costs, etc.) that makes it entirely feasible, but also can't confirm its infeasible economically. For example I have no insight of what fraction of terrestrial datacenter establishment cost goes into various "frictions" like paying goverments and lawyers to gloss over all the details, paying permission taxes etc. I can see how space can become attractive in other ways.
Then again if you look at the energetic cost to do a training run, it seems MW facilities would suffice. So why do we read all the noise about restarting nuclear power plants or trying to secure new power plants strictly for AI? It certainly could make sense if governments are willing to throw top dollar at searching algorithmic / mathematical breakthroughs in cryptography. Even if the compute is overpriced, you could have a lot of LLM's reasoning in space to find the breakthroughs before strategic competitors do. Its a math and logic race unfolding before our eyes, and its getting next to no coverage.
as an example my points almost instantly fell down 15 points, but over the last 11 hours it has recuperated back to just a 1 point drop.
it's not because they don't like to write an apology (which I don't ask for) that they aren't secretly happy they learnt something new in physics, and in the end thats what matters to me :)
Yeah doesn't sound particularly feasible, sorry. Glad you know all the math though!
For a 230 kW cluster: 16 x DGX (8x)B200; we arrived at a 30m x 30m solar PV area, and a 90 meter distance from the center of the solar array to the tip of the pyramid.
1 GW = 4348 x 230 kW
sqrt(4348)= ~66
so launch 4348 of the systems described in the calculation I linked, or if you insist on housing them next to each other:
the base length becomes 30 m x 66 = 1980 m = ~ 2 km. the distance from center of square solar array to the tip of the pyramid became 6 km...
any of these systems would need to be shipped and collected in orbit and then assembled together.
a very megalomaniac endeavor indeed.
The resistance to the idea is because it doesn’t make any sense. It makes everything more difficult and more expensive and there’s no benefit.
To run just one cluster (which would be generally a useless endeavor given it is just a few dozen GPUs) would be equivalent to the best we've ever done, and you wonder why you're being downvoted? Your calculations, which are correct from a scientific (but not engineering) standpoint, don't support the argument that it is possible, but rather show how hard it is. I can put the same cluster in my living room and dissipate the heat just fine, but you require a billion dollar system to do it in space.
The main benefit is that solar panels go from a complicated mess of batteries + permitting to a very stable, highly efficient energy source.
It's you who didn't answer my question :)
Would you prefer big tech to shit where we eat, or go to the bathroom upstairs?
The reason I'm talking about computers on the ground using the atmosphere for cooling is because that's how things are done right now and that's the obvious alternative to space-based computing.
Why does it matter what I prefer? I'd love to see all industry in space and Earth turned into a garden. I'm not talking about what I want. I'm talking about the economics. I'm asking why so many people are talking about putting data centers in space when doing so would be so much more difficult than putting data centers on Earth. If your argument is that it's more difficult but it's worth the extra effort so we don't "shit where we eat," great, but that's the first time I've ever seen that argument put forth. None of the actual players are making that case.
In reality, probably radiator designs would rely on fluid cooling to move heat all the way along the radiator, rather than thermal conduction. This prevents the above problem. The issue there is that we now need to design this system with its pipes and pumps in such a way that it can run reliably for years with zero maintenance. Doable? Yes. Easy or cheap? No. The reason cooling on Earth is easier is that we can transfer heat to air / water instead of having to radiate it away ourselves. Doing this basically allows us to use the entire surface of the planet as our radiator. But this is not an option in space, where we need to supply the radiator ourselves.
In terms of scaling by instead making many very small sats, I agree that this will scale well from a cooling perspective as long as you keep them far enough apart from each other. This is not as great from the perspective of many things we actually want to use a compute cluster for, which require high-bandwidth communication between GPUs.
In any case, another very big problem is the fact that space has a lot of ionizing radiation in it, which means we also have to add a lot of radiation shielding too.
Keep in mind that the on-the-ground alternative that all this extra fooling around has to compete with is just using more solar panels and making some batteries.
Radiation hardening:
While there is some state information on GPU, for ML applications the occasional bit flip isn't that critical, so Most of the GPU area can be used as efficiently as before and only the critical state information on GPU die or host CPU needs radiation hardening.
Scaling: the didactic unoptimized 30m x 30m x 90m pyramid would train a 405B model 17 days, it would have 23 TB RAM (so it can continue training larger and larger state of the art models at comparatively slower rates). Not sure what's ridiculous about it? At some point people piss on didactic examples because they want somebody to hold their hand and calculate everything for them?