Good thing the whole idea makes no sense, because repairs are impossible, you can only vent heat in infrared, you have to use expensive rockets to get them up there, and if they’re in LEO, they’ll fall back down to earth. Oh, and there’s more likelyhood that bits get flipped from all of the high energy particles up there.
It’ll just make LLMs at scale even more infeasible.
To run one rack of GPUs they will be about 100 square meters of solar panels and 30 square meters of radiators. That’s about 130 kilowatts of compute power. To get a gigawatt multiply that by 6,000. Musk is claiming he wants to send terawatts up to there.
The numbers didnt make sense for his tunnel thing either, but thats because he has never been an engineer but a hype-man who manipulates governments and markets for proffit.
The stats you listed are close to what 1 satellite would be if memory serves.
But when he talks about sending terrawatts up there, he’s talking about having a moon base manufacturing them and launching them into orbit with a mass acceleator, so if we’re doing that… ramping up some solar/radiator manufacturing capacity isn’t as far fetched. Were fully into sci-fi land at that point.
Yeah it’s tricky, given the pushback people are showing against datacenter construction, the issues around power, the vulnerability to attack… The space data center idea is starting to become more attractive to big tech over time, not less.
Initially, everyone agreed that the numbers just don’t add up, there’s no way it could be economical. But then these terrestrial datacenters started getting a lot of bad press, everyone was concerned about the environmental impact, about the strain on the energy grid, about water usage. All of a sudden the space plan is seeming less crazy.
and if they’re in LEO, they’ll fall back down to earth. Oh, and there’s more likelyhood that bits get flipped from all of the high energy particles up there.
You’re thinking about the right things here, these are major issues! But, they can largely avoid serious problems here. Over 500 km, there’s a lot less drag and these aren’t just going to fall out of the sky. And under 1000 km the earth’s magnetic field still mostly protects them from high energy particles. They will still definitely be more exposed to high energy particles causing random bit flips, but it’s going to be in a totally manageable range.
Giant AI datacenters will truly never be an efficient or wise use of resources, at least in my opinion, it’s just a largely stupid endeavor. But for better or worse, I expect we will be seeing some of these space based data centers in the near future. I guess we’ll just have to wait and see how economically sustainable they are…
Bad press doesn’t mean they can start losing even more money on space data centers (which are going to be incredibly inefficient and costly). They already do not make a profit, they’re all deep in the red.
Drag is not the only problem (even though these would definitely be LEO, it is just even more expensive to not have them in LEO and it increases latency), nor are bit flips. The biggest issue will continue to be heat. If you haven’t been to a data center before, you may be surprised at how much heat they generate. I’ve only been to non-LLM data centers, and the heat in the warm aisles was unbearable.
And that’s without the sun heating the exterior to incredible amounts. You cannot use contact to dump heat into space.
You might also be surprised at how much e-waste data centers generate, especially LLM-geared data centers. They obliterate hardware, even with optimal cooling, and that hardware requires constant replacement. How are they going to manage that?
Also, unless they use nuclear (which would drastically increase costs and add heat) they’re going to need gigantic solar arrays, much bigger than the IIS’ array.
I think CorrectAlias meant that heat transfer by direct contact, which is how we cool things in general, is not available, making cooling a datacenter cluster in space actual science fiction.
@phutatorius@technology So, a high-efficiency thermoelectric generator, powered by #datacenter heat directly, might be a win-win. I don’t understand why the current (jk) #technology efficiency is still so low. Surely someone must be working on this.
Good thing the whole idea makes no sense, because repairs are impossible, you can only vent heat in infrared, you have to use expensive rockets to get them up there, and if they’re in LEO, they’ll fall back down to earth. Oh, and there’s more likelyhood that bits get flipped from all of the high energy particles up there.
It’ll just make LLMs at scale even more infeasible.
Just FYI the rockets won’t be expensive if this plan ends up working.
The only way this works beyond a tech demo, is if Starship works.
If Starship works as planned (big if), the marginal cost to orbit will round down to the cost of fuel to launch over the lifetime of the rocket.
Starship being fully reusable and working is going to be a mind boggling change in what access to space looks like and means.
To run one rack of GPUs they will be about 100 square meters of solar panels and 30 square meters of radiators. That’s about 130 kilowatts of compute power. To get a gigawatt multiply that by 6,000. Musk is claiming he wants to send terawatts up to there.
The numbers didnt make sense for his tunnel thing either, but thats because he has never been an engineer but a hype-man who manipulates governments and markets for proffit.
The stats you listed are close to what 1 satellite would be if memory serves.
But when he talks about sending terrawatts up there, he’s talking about having a moon base manufacturing them and launching them into orbit with a mass acceleator, so if we’re doing that… ramping up some solar/radiator manufacturing capacity isn’t as far fetched. Were fully into sci-fi land at that point.
Yeah it’s tricky, given the pushback people are showing against datacenter construction, the issues around power, the vulnerability to attack… The space data center idea is starting to become more attractive to big tech over time, not less.
Initially, everyone agreed that the numbers just don’t add up, there’s no way it could be economical. But then these terrestrial datacenters started getting a lot of bad press, everyone was concerned about the environmental impact, about the strain on the energy grid, about water usage. All of a sudden the space plan is seeming less crazy.
You’re thinking about the right things here, these are major issues! But, they can largely avoid serious problems here. Over 500 km, there’s a lot less drag and these aren’t just going to fall out of the sky. And under 1000 km the earth’s magnetic field still mostly protects them from high energy particles. They will still definitely be more exposed to high energy particles causing random bit flips, but it’s going to be in a totally manageable range.
Giant AI datacenters will truly never be an efficient or wise use of resources, at least in my opinion, it’s just a largely stupid endeavor. But for better or worse, I expect we will be seeing some of these space based data centers in the near future. I guess we’ll just have to wait and see how economically sustainable they are…
Bad press doesn’t mean they can start losing even more money on space data centers (which are going to be incredibly inefficient and costly). They already do not make a profit, they’re all deep in the red.
Drag is not the only problem (even though these would definitely be LEO, it is just even more expensive to not have them in LEO and it increases latency), nor are bit flips. The biggest issue will continue to be heat. If you haven’t been to a data center before, you may be surprised at how much heat they generate. I’ve only been to non-LLM data centers, and the heat in the warm aisles was unbearable.
And that’s without the sun heating the exterior to incredible amounts. You cannot use contact to dump heat into space.
You might also be surprised at how much e-waste data centers generate, especially LLM-geared data centers. They obliterate hardware, even with optimal cooling, and that hardware requires constant replacement. How are they going to manage that?
Also, unless they use nuclear (which would drastically increase costs and add heat) they’re going to need gigantic solar arrays, much bigger than the IIS’ array.
Yeah, when you heat something, it emits infrared. Kind of the nature of hot things.
I think CorrectAlias meant that heat transfer by direct contact, which is how we cool things in general, is not available, making cooling a datacenter cluster in space actual science fiction.
This is exactly what I meant, thanks.
@phutatorius @technology So, a high-efficiency thermoelectric generator, powered by #datacenter heat directly, might be a win-win. I don’t understand why the current (jk) #technology efficiency is still so low. Surely someone must be working on this.
https://en.wikipedia.org/wiki/Thermoelectric/_generator
you still need thermal radiators to create the temperature differential.