Google Unveils Project Suncatcher, Envisioning AI Models Running in Space

"Google Research has unveiled Project Suncatcher, a research initiative exploring how solar powered satellite constellations equipped with Tensor Processing Units TPUs could one day enable large scale artificial intelligence computation in space. The project is part of an early stage study into scalable AI infrastructure that operates beyond Earth's surface, leveraging the continuous solar energy available in orbit. According to Google, satellites operating in sun synchronous orbits can collect solar power almost continuously, up to eight times more efficiently than ground based systems."

"The proposed design envisions constellations of compact satellites linked by free space optical connections. These high speed links could distribute machine learning workloads across multiple TPUs in orbit, reducing dependence on terrestrial data centers and minimizing environmental impact. A preprint paper, Towards a future space-based, highly scalable AI infrastructure system design, details the system architecture and the technical challenges involved. Among these are maintaining high bandwidth communication between satellites, managing orbital dynamics for tight formations, and ensuring radiation tolerance for TPU hardware."

"Early laboratory experiments demonstrated optical data transmission speeds of up to 1.6 terabits per second using a single transceiver pair. The research team also modeled orbital behaviors using the Hill-Clohessy-Wiltshire equations to simulate how clusters of up to 81 satellites could maintain stable formations at altitudes around 650 km. These simulations suggest that compact satellite groupings just hundreds of meters apart, could remain stable with limited station keeping maneuvers."

Project Suncatcher explores creating solar-powered satellite constellations equipped with Tensor Processing Units to perform large-scale AI computation in orbit. Satellites in sun-synchronous orbits can collect solar power almost continuously and up to eight times more efficiently than ground-based systems. The design uses compact satellites connected by free-space optical links to share ML workloads across multiple TPUs and reduce dependence on terrestrial data centers. Laboratory experiments achieved optical transmission speeds up to 1.6 terabits per second with a single transceiver pair. Orbital simulations using the Hill-Clohessy-Wiltshire equations indicate clusters of up to 81 satellites around 650 km can hold tight formations. Radiation testing of Trillium TPU v6e showed tolerance for expected five-year low Earth orbit exposure with only minor irregularities.

Read at InfoQ