11:51
- Data centers are expensive and complicated. They take vast amounts of water, electricity, and land. But what if we didn't have to build data centers on land at all? What if we could build them in space? That's exactly what companies like Blue Origin and SpaceX are attempting to do. In fact, earlier this year, SpaceX filed an application with the FCC asking to launch a cluster of up to a million satellites to build an orbital data center. That's part of their pitch to investors ahead of their anticipated IPO later this year. But just how feasible is this really? To find out, we asked an expert.
- I'm a computer architect by training, which is to say that I design processors and memory systems and think a lot about how they get deployed. - And computer architect is the coolest name in the computer science field, I think. - I think so. - Before we launch into space, we need to understand what data centers need on the ground. Namely an immense amount of electricity hardwired in through cables connected to the grid. And data centers also generate heat. So they need to be cooled down, either with air or often using vast amounts of water routed through cooling towers.
- This warm water, when it hits those cooling towers, some of it will evaporate and bring the temperature back down. - And inside these massive buildings, each server in Iraq needs to work in tandem with the ones around it. In order to do that, they need to share data by transferring it through physical cables. Often fiber optic. - The power of data center compute is really in the networking equipment. We are able to harness massive amounts of compute across many different servers and coordinate that computation to answer questions that we wouldn't be able to answer on a single computer on, or on a single machine.
- But when you try to translate that computing power into space, things get more complicated. Satellites already in orbit are typically for one of two purposes, either data collection, like looking at weather patterns or taking satellite images or communication like SpaceX's Starlink satellites. Both of those cases are less power intensive than doing real-time computations up in space. So the satellites for an orbital data center would need to be designed differently, with even larger arrays of solar panels. - If we were to launch these data centers into space, we would be able to use these solar panels and get access to inexpensive or essentially free energy from the sun.
- In order to get predictable sun exposure, these satellites would have to maintain a particular orbit, known as a sun synchronous orbit. - If you were to put it into sun synchronous orbit, the satellite would pass over the particular location on the planet at the same time every day. And if you were to pick a particular orbit, where you are following Ladon dusk line with the point at which daytime turns at nighttime, the satellite would might be able to get nearly constant exposure to solar energy.
- But cooling is a problem too. There's no air in space, so we can't run cool air over the chips like we can on land, and we can't pipe in extra water to cool down the chips via evaporative cooling either. So satellites might have to rely on closed loop radiator panels running a fixed amount of water over the chips outside the satellite to cool down via infrared radiation and back again. But those radiator panels can be heavy, requiring more fuel to launch into space. - Figuring out the design of the satellite is gonna be increasingly important because that will determine the weight profile and the cost of the launch.
- Radiation in space is also a concern for computer engineers. If a high energy particle strikes a GPU, it can literally flip a zero to a one, or vice versa. That could mean the difference between calculating the correct answer to a problem, and getting the wrong one entirely. So for the data center to be useful, we'd need solutions to prevent or correct those errors. - I would envision three types of strategies. The first type of strategy would be essentially to detect when those errors have arisen and then correct them.
Alternatively, we could compute the same process, the same job, it's the same task, multiple times and compare the results. But as you might imagine, this is expensive because instead of having one GPU, you would now have three. The third option we would have to protect chips and computation from radiation would be just add some sort of physical shielding that encloses the servers and prevents more of those high energy particles from hitting that ship. - But there's another problem. How do we get the data from Earth to the satellites and back again? - When we think about communication from earth to one of these satellites, we are relying on radio frequency communication.
And when we think about the satellites communicating with each other, depending on how precisely these satellites are positioned, they might be able to communicate via lasers. Essentially, one satellite would send a laser signal to another satellite and they would modulate that laser to communicate data across that signal. - With lasers, data transfer speeds are pretty high, but the problem here is bandwidth. - If you wanted to send quite a bit of data from Earth into space, it might be better to simply put on a disc and launch that disc into space.
- That might mean that implementation is key. Instead of sending massive amounts of data back and forth, send the satellite up preloaded with a training data set for an AI model, - You would pose your question to the satellite. The questions might be kind of short, right? Because even if you have a massive document or database that you want to ask a question against, your question itself may only be 120 characters or something like that. So imagining a scenario where you send only a few hundred tokens up into space, that is certainly feasible and possible for AI inference. - But with increased traffic bandwidth may still be an issue, meaning it could take longer to get an answer to your query when compared to a data center on land.
Ultimately, it comes down to one question. Does the math work out large solar arrays, radio panels, and ship shielding, all add costs and weight, meaning that launch costs would also be higher? Is that likely to be offset by the energy savings from solar? It's unclear. - I think the costs of doing this are large and at the moment unknown. We need to collect data, we need to run experiments. We need to explore the design space and figure out what combination of technologies will allow us to get to orbital computing.
