Artificial intelligence (AI) holds immense promise for addressing some of the world’s toughest challenges. Google Research is quite literally aiming for the stars to fully unlock AI’s potential with Project Suncatcher.
This visionary “moonshot” is focused on bringing massive-scale machine learning (ML) to orbit. This initiative envisions an interconnected constellation of solar-powered satellites, each housing Tensor Processing Unit (TPU) AI chips.
Project Suncatcher
Initially, the research addressed several intimidating engineering and physics problems. A core challenge is replicating the high-speed connectivity of a terrestrial data center. According to Google researchers, achieving comparable ML performance to ground-based facilities requires tens of terabits per second in inter-satellite links.
To beat the steep signal loss over vast distances, the team proposed a radical solution.
They envision a tightly formed design, with satellites flying just a few kilometers apart. This close proximity, combined with advanced optical technologies, is crucial to achieving the necessary bandwidth. Early bench-scale testing successfully demonstrated this concept. Researchers say it achieved a total transmission rate of 1.6 Tbps using just one pair of transceivers.
Explore Tomorrow's World From Your Inbox
Get the latest science, technology, and sustainability content delivered to your inbox.
I understand that by providing my email address, I agree to receive emails from Tomorrow's World Today. I understand that I may opt out of receiving such communications at any time.
Maintaining such a compact satellite cluster is a complicated orbital balancing act. The researchers developed sophisticated physics models to study the orbital dynamics of the system. Their findings suggest that by spacing satellites a mere few hundred meters apart, surprisingly modest maneuvers will be enough to keep the constellation stable in its solar orbit.
Surviving Space’s Harsh Conditions
Any hardware sent to space must withstand the harsh environment of low-Earth orbit. To test this, the team exposed Trillium, Google’s v6e Cloud TPU, to a high-energy proton beam simulating space radiation. The team says the results were reassuring.
They said the chips proved to be unexpectedly radiation-hardy. While the High Bandwidth Memory (HBM) was the most sensitive part, it only began showing issues after receiving a cumulative radiation dose almost three times higher than what’s expected over a five-year mission. This suggests the TPUs are well-suited for space applications.
Historically, launch costs have grounded such ambitious space ventures. However, Google’s analysis offers a hopeful economic outlook. The research team projects that if launch prices drop below $200/kg by the mid-2030s, the overall cost could become roughly on par with the energy costs of running an equivalent facility on Earth, calculated on a per-kilowatt/year basis.
Next Steps Into Orbit
With the fundamental concepts validated, the project’s next milestone is a crucial learning mission. In collaboration with Planet, Google plans to launch two prototype satellites by early 2027. This experiment will be the first test of their TPU hardware and models in space.