Google's ambitious Project Suncatcher aims to revolutionize AI by placing powerful Tensor Processing Units (TPUs) in space. This groundbreaking initiative, detailed in their research blog, envisions a network of interconnected satellites harnessing the Sun's energy to scale machine learning capabilities. With solar panels up to 8 times more productive than on Earth, these satellites will operate in a 'dawn-dusk sun-synchronous low Earth orbit', reducing the need for traditional power generation methods. This innovative approach could make space the ideal location for AI computation.
The satellites will communicate via free-space optical links, enabling high-bandwidth, low-latency connections. To match the performance of Earth's data centers, these links must achieve speeds of tens of terabits per second, with satellites flying in close formation. Despite the challenges, Google's initial analysis suggests that space-based ML compute is feasible, both physically and economically.
Google's TPUs, known as Trillium, have already undergone radiation testing, demonstrating remarkable resilience. While the High Bandwidth Memory (HBM) subsystems were the most sensitive, they only exhibited irregularities after a cumulative dose of 2krad(Si), far exceeding the expected five-year mission dose. This indicates that Trillium TPUs are well-suited for the harsh conditions of space.
Furthermore, Google predicts that launch costs will significantly decrease, falling below $200/kg by the mid-2030s. This development could make space-based data centers economically viable, comparable to their terrestrial counterparts in terms of energy costs. However, Google acknowledges the need to overcome engineering hurdles, such as thermal management, high-bandwidth ground communications, and on-orbit system reliability.
To address these challenges, Google is collaborating with Planet to launch prototype satellites by early 2027. These satellites will test the performance of models and TPU hardware in space and validate the use of optical inter-satellite links for distributed ML tasks. For more insights, visit their research paper on space-based AI infrastructure.
This project raises intriguing questions about the future of AI and space exploration, inviting discussion and debate among experts and enthusiasts alike.
