The Future of AI: Harnessing the Sun for Limitless Innovation
Artificial intelligence (AI) stands at the forefront of technological advancement, holding the potential to revolutionize our existence by unlocking new scientific breakthroughs and addressing some of the most pressing issues facing humanity. But where do we begin to unlock its full capabilities?
Look no further than the Sun, our solar system's most powerful energy source, which provides more energy than 100 trillion times the total electricity produced by humans. In the right orbit, solar panels can be up to eight times more efficient than those on Earth, producing energy almost constantly, which drastically reduces the need for bulky batteries. As we consider future possibilities, it seems that space could indeed be the optimal environment for scaling AI computational resources.
To make this vision a reality, we are proud to introduce our research initiative known as Project Suncatcher. This ambitious project imagines networks of compact, solar-powered satellites equipped with Google Tensor Processing Units (TPUs) and linked via free-space optical communications. This innovative design not only presents immense scalability potential but also seeks to lessen the strain on Earth's natural resources.
Today, we have shared our initial findings in a preprint paper titled "Towards a Future Space-Based, Highly Scalable AI Infrastructure System Design" (read more here). This document outlines our progress in addressing foundational challenges in this audacious effort, encompassing areas like high-speed communications between satellites, orbital mechanics, and the effects of radiation on computing systems. By concentrating on a modular framework of interconnected satellites, we are setting the stage for a groundbreaking space-based AI infrastructure that could redefine modern computing.
Project Suncatcher is emblematic of Google’s adventurous spirit, reflecting our history of undertaking bold moonshot projects that seek to solve complex scientific and engineering dilemmas. There are bound to be uncertainties, reminiscent of how we began our journey toward building large-scale quantum computers over a decade ago, long before it was deemed feasible, or how we pioneered the idea of autonomous vehicles more than 15 years ago, ultimately leading to the launch of Waymo, which now serves millions around the world.
Key System Design Challenges
Our proposed vision outlines a network of interconnected satellites likely to orbit in a dawn-dusk sun-synchronous pattern, which allows for near-constant exposure to sunlight. This orbital positioning maximizes solar energy absorption and minimizes the reliance on heavy onboard batteries. However, various technical challenges must be addressed to ensure the viability of this system:
- Establishing Data Center-Scale Inter-Satellite Links
The demands of large-scale machine learning (ML) applications necessitate distributing workloads across numerous accelerators connected by high-bandwidth, low-latency links. To rival terrestrial data center performance, satellite connections must support data rates in excess of tens of terabits per second. Our analysis shows that this can be achieved using multi-channel Dense Wavelength-Division Multiplexing (DWDM) transceivers combined with spatial multiplexing.
However, achieving such immense bandwidth requires received power levels that are thousands of times higher than what is typical for long-range communications. Since the received power decreases with the square of the distance, we can mitigate this challenge by positioning the satellites in close proximity to one another, minimizing spacing to just kilometers. This strategy effectively optimizes the link budget, which accurately accounts for signal power losses in the communication system. Our team has initiated testing of this concept through a bench-scale model that successfully achieved an impressive 800 Gbps transmission each way, totaling 1.6 Tbps using a single pair of transceivers.
- Managing Dense Satellite Formations
The successful operation of high-bandwidth inter-satellite links necessitates that our satellites maintain a much tighter formation than any existing systems. To explore the orbital dynamics of such a constellation, we have created both numerical and analytical models, beginning with the Hill-Clohessy-Wiltshire equations—which describe the movement of satellites relative to a circular orbit—complemented by a JAX-based model for further refinements that account for additional perturbations.
At the altitude chosen for our constellation, the irregularity of Earth's gravitational field and potential atmospheric drag become significant non-Keplerian factors influencing satellite dynamics. Our modeling indicates that for a configuration of 81 satellites situated at an average altitude of 650 km, with the cluster radius set at 1 km, the distances between satellites can vary between approximately 100-200 meters due to Earth's gravitational influence. The results suggest that, with satellites only hundreds of meters apart, only modest adjustments will be required to maintain their positions within the sun-synchronous orbit we desire.
- TPUs and Radiation Resistance
In order to ensure that our ML accelerators can function effectively in space, they must endure the harsh conditions present in low-Earth orbit. We have subjected Google’s v6e Cloud TPU, known as Trillium, to testing under a 67MeV proton beam to assess its resilience against total ionizing dose (TID) and single-event effects (SEEs).
The findings were encouraging: while the High Bandwidth Memory (HBM) subsystems displayed some susceptibility, irregularities were only noted after accumulating a dose of 2 krad(Si)—nearly three times the anticipated (shielded) five-year mission dose of 750 rad(Si). No significant failures were observed due to TID up to the maximum tested dose of 15 krad(Si) on a single chip, indicating that Trillium TPUs are surprisingly robust against radiation, making them suitable for space applications.
- Evaluating Economic Feasibility and Launch Costs
Traditionally, prohibitive launch expenses have hindered the advancement of large-scale space systems. However, our evaluation of historical and predicted launch cost data hints that these costs may decline to under $200 per kilogram by the mid-2030s. At such a price, the operating expenses of a space-based data center could become comparable to the reported energy costs of similar terrestrial data centers on a per-kilowatt/year basis. For further insights, please refer to our preprint paper (read it here).
Looking Ahead
Our initial investigations indicate that the fundamental principles supporting space-based ML computations are both physically viable and economically feasible. However, significant engineering challenges lie ahead, including thermal management, establishing high-bandwidth communications with the ground, and ensuring the reliability of on-orbit systems.
As we map the future, our next significant step is a learning mission in collaboration with Planet, set to launch two prototype satellites by early 2027. This experiment will evaluate the performance of our models and TPU hardware in space while verifying the effectiveness of optical inter-satellite links for distributed machine learning applications.
In the long run, large-scale constellations could radically evolve into a novel satellite architecture, integrating advanced computing designs tailored for the unique conditions of space along with mechanical systems that harmonize solar energy capture, computation, and thermal management. Just as the rise of sophisticated system-on-chip technologies fueled advancements in modern smartphones, the integration and scale of our design will redefine the possible limits of space-based technologies.
Acknowledgments
The paper "Towards a Future Space-Based, Highly Scalable AI Infrastructure System Design" was compiled by a talented team including Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, and James Manyika.
We extend our gratitude to Amaan Pirani for his pivotal contributions to cost modeling and overall feasibility assessments, Marcin Kowalczyk for independently validating our calculations, Thomas Zurbuchen for his insights into system and architecture concepts, and Kenny Vassigh and Jerry Chiu for their expert advice on system and thermal design.
