It’s an experiment to see how the chips hold up in the harsh vacuum of the cosmos.
Google’s Project Suncatcher, which ultimately aims to put large AI installations in low-Earth orbit, is preparing for a milestone event. The company will be sending a teensy-tiny AI data center into space to see how it reacts to the harsh conditions, as
A satellite named MVP will be aboard a SpaceX Falcon 9 rocket set to launch on October 1, where it will be placed into orbit. This isn’t a full data center in the conventional sense. It’s four specialized computer chips called “tensor processing units” which, combined, offer the power of a single server in a data center.
Solar panels will supply just one kilowatt of power to juice the chips, which is equivalent to the energy needed to run a hair dryer. Google says this will be enough to test the hardware against the harsh conditions of space. The company plans to let the satellite answer simple AI queries for a year, though it will orbit Earth for six years before it burns up during a descent into the atmosphere.
These chips have undergone a battery of tests to simulate the void of space, including getting blasted with large amounts of radiation at Crocker Nuclear Laboratory. Radiation can seriously mess with computer chips, causing something called “bit flips.” This essentially flips the binary code from a one to a zero, or vice versa.
What‘s the solution to bit flips? Google is taking the tried-and-true PC method. It plans on restarting the chips when they act up. There’s also a proprietary cooling system in place that uses layers of conductive material to expel heat into space, though it only works for around 15 minutes before the chips have to be turned off to cool down.
The initial experiments at Crocker reportedly went well enough to take the next step, which is shooting the chips into the great beyond. However, this doesn’t mean we are particularly close to having actual AI data centers in space. James Manyika, the Senior Vice President for Research, Technology & Society at Google, urges the public to temper their expectations. “We don’t expect, to be perfectly frank, that we’ll have anything usefully operational in the next few years,” he said.
“Expanding from one satellite to a vast network of them that operate like a giant data center will take years and enormous funds,” Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University, told The New York Times. “If you do this on a large scale, there are additional engineering problems.”
Those additional engineering problems are no joke, despite people like Elon Musk pronouncing that putting AI data centers in space will be easy. There’s the radiation and that whole bit flip thing, which seems extremely finicky with four chips let alone thousands.
Cosmic rays are an issue, as they are actual physical particles traveling quickly through space that could punch through and actually melt semiconductor junctions. It’s also extremely expensive to launch this stuff into space. A single Falcon 9 launch, for instance, costs around $74 million.
Scaling up the cooling system looks to be incredibly difficult. Remember, Google’s current system can only work for 15 minutes, and that’s with the small amount of heat generated by just four chips. Standard heat dissipation does not work in space. Power must also be considered. Using a solar array for four chips is quite different from launching a massive array to power a legitimate data center.
Despite these challenges, Google plans on scaling this project up. It wants to launch two similar satellites next year, with the eventual goal of launching 80 satellites that will fly in close formation to better process AI queries. It’s also looking into building a much larger satellite, around the size of a football field. The company has not indicated how much money it is pouring into this endeavor.
