Google has sent one of its Tensor Processing Units into space for the first time as part of Project Suncatcher, a long-term effort to determine whether large AI computing clusters could eventually operate in orbit. The prototype satellite launched from California aboard a SpaceX rocket and will test whether Google’s AI chip can reliably run workloads in space.
The satellite was built by Planet Labs and is designed to provide about one kilowatt of continuous power while cooling the TPU and running a series of AI models. Google said the mission will test conditions that cannot be fully recreated on the ground.
Google Begins Testing AI Chips In Orbit
Once the satellite completes commissioning, its TPU will run in 15-minute intervals to reduce pressure on the spacecraft’s power and thermal systems. The current spacecraft uses a standard Planet Labs platform rather than hardware designed specifically for large-scale computing.
Google plans a more advanced demonstration next year involving two satellites built more specifically for AI workloads. Those spacecraft will attempt to communicate and work together through a laser link, providing an early test of the distributed computing architecture Google wants to develop under Project Suncatcher.
The company ultimately envisions orbital data centers made from networks of 81 satellites flying in close formation. Google executive Travis Beals said high bandwidth and low latency between TPUs will be important if future systems are expected to process multi-rack AI workloads in parallel.
Project Suncatcher Looks Beyond Current AI Workloads
Google describes Suncatcher as a long-term research project rather than a near-term deployment plan. The company is designing around computing requirements it expects several years from now, when AI workloads may be larger and more distributed.
Google also published an updated, peer-reviewed analysis of orbital data centers in the journal Joule. The research examines technical requirements including launch costs, radiation exposure, power, cooling, and the economics of moving large amounts of computing equipment into orbit.
The study assumes launch costs could fall toward $200 per kilogram by 2035 if SpaceX continues reducing costs at roughly the historical rate identified by Google’s researchers. Reaching that level could require Starship to carry about 370,000 metric tons into orbit across roughly 1,800 launches over 10 years, assuming 200 metric tons per flight.
Radiation Tests Show Limits For Large Training Runs
Google also repeated radiation testing on its TPU hardware after determining that an earlier test configuration provided more shielding than the chips would receive in orbit. The updated tests produced somewhat more errors in the processors’ logic circuitry.
Even so, Google believes the chips could support large inference workloads during a satellite’s expected five-year operating life. Beals estimated the error rate for typical inference operations at about one in a million.
The same error rate could become more significant for large training jobs involving thousands of chips operating continuously for months. Google’s Project Suncatcher overview therefore presents orbital AI computing as a research problem that still depends on improvements in launch systems, satellite hardware, networking, and workload design.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.