Google Launches Compute Satellite with SpaceX, Eyes Future in Orbit

In a significant step forward for space technology, Google has partnered with SpaceX to launch its first orbital compute satellite from California. This marks the debut of Google's advanced Tensor Processing Unit (TPU) chips in outer space, setting the stage for future projects in space computing.

Testing the TPU in Space

The satellite, developed by Planet Labs, is designed to assess the performance of Google's TPU under the unique conditions of space. This includes verifying the chip's ability to operate with a continuous power supply while maintaining temperature control. Travis Beals, a Google executive overseeing Project Suncatcher, expressed the importance of conducting these tests in a real space environment.

“While ground tests provide valuable data, nothing compares to the actual conditions in space,” said Beals. Once operational, the satellite will utilize its TPU in 15-minute increments to mitigate stress on its power management systems. The initial version is based on a standard design, but upcoming demonstrations scheduled for next year will feature more specialized satellites capable of handling larger computational tasks.

Vision for the Future of Space Computing

Google’s ambitions extend well beyond this initial launch. The company foresees a groundbreaking orbital data center that consists of 81 satellites operating in tandem. Such infrastructure aims to significantly enhance processing capabilities for future AI applications.

“The interconnectivity and speed of communication between TPUs are critical for managing complex workloads,” Beals said. “We're focusing on not just today’s computing demands but also on future needs that may arise in the next five years.” However, to realize this vision, the development of cost-effective rockets capable of achieving numerous flights will be essential.

Understanding Launch Costs and Future Projections

On the same day, Google unveiled a peer-reviewed white paper detailing its research on orbital data centers, contributing to the discourse on space computing. Published in Joule, the paper emphasizes access to space and speculates on future pricing trends for rocket launches.

The analysis indicates a favorable trajectory in launch costs, estimating that SpaceX could reduce prices to approximately $200 per kilogram by 2035. This reduction is anchored in the concept of a “learning curve,” which suggests a yearly decrease of around 20% since the inception of SpaceX’s Falcon 1 rocket.

To achieve these lower launch costs, the researchers believe that SpaceX's Starship could need to deliver around 370,000 tons of payload to orbit. This ambitious target translates to approximately 1,800 launches over the next decade, averaging 180 flights annually, assuming each mission can transport 200 metric tons.

However, this presents a significant challenge. Starship has yet to reach such a high-frequency launch rate, having successfully completed only a handful of missions in any given year. Despite this, SpaceX’s CEO, Elon Musk, has expressed confidence that the vehicle could eventually reach an hourly launch rate as early as 2029—though this, too, has yet to be substantiated.

Ensuring Reliability in Harsh Conditions

Despite the challenges surrounding launch frequencies, Google’s recent research also provides encouraging news about its chips’ robustness against space radiation. The company revised its testing protocols after discovering that initial configurations provided better shielding than anticipated. Although this adjustment led to a slight increase in error rates within the chip's circuitry, the overall findings remain positive.

Beals highlighted that during typical inference operations, the error likelihood is quite low, estimated at about one in a million instances. Conversely, challenges still persist when considering large-scale training operations with numerous chips functioning over extended periods.

Overall, Google’s foray into space-based computing signals a pivotal moment in the race for advanced data processing capabilities off-planet. With ongoing testing and ambitious plans for orbital data centers, the future of AI in space looks promising.