Google has taken a giant leap towards fulfilling its vision of launching artificial intelligence infrastructure into space by successfully launching and deploying the prototype satellite equipped with its unique AI hardware into low earth orbit. This is an element of Project Suncatcher, an innovative research project investigating whether upcoming AI computing systems could work from space by relying on easy access to solar energy.
The first prototype satellite, which was launched on October 1 on SpaceX’s newly launched Transporter-18 rideshare mission was built by partnership with Planet. Google has verified communication with the satellite and said it was functioning as intended. This mission will now provide the real-time test of how Google’s Tensor processing units, or TPUs perform under space’s harsh conditions.
The satellite is equipped with four Google TPUs [processing units,] custom-built for the type of machine-learning work Google specializes in. Instead of rushing to develop a full-fledged orbital data center, the company is using this first launch as a kind of engineering trial. Scientists are eager to discover whether their hardware can withstand the torrent of force, radiation, and heat experienced during the launch process, and still operate accurately.
Space has its own unique set of daunting hurdles compared to a typical data center. During launch, ships endure shock and vibration. It is reported that these structures experience accelerations of fifty to one hundred g0s, or 50 to 100 g0s as Google. In orbit, particles are bombard the vehicle and may cause electronics to malfunction due to radiation, performing errors such as bit flips, which can have serious implications for calculations.
Cooling is also a key challenge. Very high-powered AI computer chips produce enormous amount of heat, but traditional air-based cooling solutions cannot operate in the vacuum of space. Google’s solution involves a series of heat pipes and radiators to remove this heat from the chips and radiate it into space. Some parts of the cooling system have already been tested in Earth’s thermal vacuum chambers, but the orbital mission should be the ultimate test.
The idea of Project Suncatcher stems from the large energy demands of today’s AI support infrastructure. Google postulates that satellites in propitious orbits could stay in sunlight most of their operational cycle. The firm calculates that solar cells in the correct orbital conditions could outrun equivalent cells on the ground by as much as eight times. This could be a new source of power for up and coming AI computing structures, alleviating the load placed on Earth’s grids.
Google has visions for much more than the 4-chip prototype. In time, the next generation of satellites would be equipped with many more (even dozens) TPUs working together as a network of computers. Rather than depending solely on radio communications, the satellites would communicate through high-bandwidth laser links. Google has scheduled a trial of this technology with two satellites in 2027.
Orbital AI computing is something the tech and space worlds are starting to focus on. In the tech world, there are other players developing this concept and they’re asking the question, could satellites someday be incorporated into infrastructure scale computing? Given the increase pressures on terrestre data centers for electricity, cooling, and space.
