The consensus I’ve heard on AI data centers in space is that cooling them is going to be next to impossible, since they’re in a vacuum. And sure, the vacuum of space isn’t really cold, it just doesn’t have any temperature whatsoever. But our sun (and every other star) radiates heat into the vacuum of space. So why can’t space-based data centers lose their excess heat by radiating into space?
I went back three months in this forum and didn’t see a recent thread on the subject, but if there’s something further back, please link.
Elsewhere, I linked to a New York Times article (gift link) about Google’s plan to launch an experimental AI data center in space later this week. The article mentions how the cooling will occur.
The biggest challenge was how to cool the A.I. chips, which heat up when they perform calculations and process information. Fans, which typically help dissipate the heat, do not work in space. So the Silicon Valley company instead developed a cooling system that uses layers of conductive material to expel the heat into space.
The bottom layer is made up of Google’s A.I. chips, which sit on a green motherboard. The next layer consists of “thermal interface material,” a pale green putty that comes in sheets like Fruit Roll-Ups and connects the chips to layers of aluminum and copper, radiating heat away from the motherboard. Finally, there is a radiator panel, which projects heat into space.
The chips can operate for about 15 minutes in space before needing to be shut down so they can cool off, said Travis Beals, Google’s senior director of product management for Project Suncatcher. During that time, the chips can process short queries so Google’s A.I., known as Gemini, can respond.
So, yes, radiation is being used to dissipate heat. But that’s not enough, which is why the system will work for only a short time before being shut down to allow it to cool.
Likely it’s because the chips are generating more heat energy than can be radiated away. Lots of different materials can convert heat energy into light energy. You can detect this process easily yourself by putting your hand near a hot object, like a cast-iron skillet. The skillet is absorbing the heat from the stove and converting it into light energy. If you put your hand near the skillet, you can feel your hand getting hot even if you’re not touching the skillet. That heat you feel is the light energy radiating away from the cast-iron skillet. An infrared camera pointed at the skillet will see it as a very bright object from all the heat it is radiating away. If you turn off the stove, the skillet will remain hot for a long time. If you instead put the skillet in water or point a fan at it, it will cool off quickly. But if it’s just sitting on the stove radiating the residual heat away with light energy, it will take a long time to cool down fully. For AI chips in space, there would need to be a material in the satellite which can absorb the heat from the chips and radiate it into space faster than the chips can generate the heat.
The Sun generates less power per cubic meter than a data center does, and the Sun, while radiating heat outward into space remains hotter than a data center could tolerate.
Yeah, my brain recalls the three means of heat transfer: conduction, convection, radiation. And on the moon, excess heat could be directly conducted to the cold lunar rock, at least during the lunar night. Nothing to conduct to in the vacuum of space, so orbiting data centers’ only option is radiation.
Good point. When the sun’s light gets here, much of it is converted to heat, but it traveled the 93 million miles as light, not heat.
I expect turning a data center’s excess heat into light or other electromagnetic radiation would be even more complicated and expensive than any other alternatives.
Hoocoodanode? I’m learning a lot in this thread already.
Hey, maybe how that’s how Elon should raise the temperature of Mars to habitable levels, just cover it with AI data centers.
(Yeah, I know there are other, more unalterable reasons why human life on Mars isn’t tenable. Cold temperatures and thin atmosphere are just the start.)
Actually, a vacuum has virtually no molecules, so it eliminates heat transfer through conduction (direct physical contact) and convection. It actually inhibits the cooling process, that’s why a vacuum thermos is very effective.
Want to add some details here:
Yes, the suns power density is actually quite low. In fact, a human at rest has a higher power output per volume.
However, the sun is absolutely gigantic, so the total power production inside is as well.
The sun is also really hot. It is hot because it cannot cool down faster. In other words, even with such a low power density it is really hard for the sun to lose its energy.
An AI data center with a much higher power density will have a hard time staying cool enough using radiation only. As long as the center is small enough, the surface to volume ratio will help with that.
This doesn’t sound right to me. It’s true that our skin can feel any intense radiation as heat, and the visible wavelength photons can have an effect. However, most of what we feel as heat comes from infrared radiation, whose photons remain constant from emission to absorption. This is as true from an incandescent light bulb as it is the sun.
Other wavelengths can have an effect on skin without much heat. Ultraviolet radiation tans skin and x-rays can cause burns.
The entire electromagnetic spectrum emitted by the sun travels as photons of varying wavelengths that have varying effects on various surfaces. Saying that it all travels as light is not correct and that it is converted to heat isn’t either.
The comparison I’ve heard is that the Sun has about the same power density as a rotting compost pile. It’s not the most relevant comparison for a data center, though, since the Sun, being so large, has an extremely low ratio of surface area to volume. Any human-scale structure will have much more surface area to volume, and so would be better at radiating heat than the Sun. Still not good enough, though, without phenomenally huge radiators.
That’s just because most of the light is infrared radiation. But if you had a green laser of the same total intensity as sunlight, and shone it on your arm, it would feel as warm as sunlight.
That was what I was coming here to post; it should surprise nobody that the Sun is really hot, and if the problem is that data centers get too hot, then saying “it works for the Sun” is not a great example.
And the rate of radiation output varies with the fourth power of the radiating body’s temperature, so the Sun radiates a lot more effectively than a data center could, I think
All satellites need to reject heat. Some more than others, but it isn’t intrinsically hard. A good start is too look at the solar arrays. Their area is a guide to the power need, and to a good approximation the power that needs to be rejected.
If you are shielded from the Sun and the Earth, space is mostly very cold. Any warm body will radiate heat away with essentially no heat coming back from space. A panel shaded from the Sun pointed so it’s sides are directed toward the cold of space will radiate power. The radiated power increases dramatically with temperature, but even a tepid object will cool. This technology has been well understood for decades.
In general the area of the radiating panels required is less than that of the solar panels creating the power. Some active control of attitude is needed, but nearly all satellites already do this.
Elon’s Starlink satellites already need to reject some heat, and the new Starlink 3 satellites even more. His Starmind satellites will just be more of the same.
Where things go silly is the idea of enormous monolithic data centres. Everything goes horribly more complicated. You need large scale active heat transfer mechanisms to move the heat into the radiators. Not a new problem, the ISS does this. But it is an unwelcome and costly problem.
So don’t make the satellites any larger than the point where simple passive systems stop working. Which is, for instance, Starmind. Google are doing something similar. This limits the compute capability of each satellite, but the thinking seems to be that these will be doing the user side of the AI, with training still earthbound. Communications bandwidth is more than enough.
Whether the business case works out is the harder question. Oracle has just stalled a huge data centre project. Whether this is a blip or the first signs of industry wide contagion will be interesting to observe.
