Key facts
- Google launched its first advanced AI chip, a Tensor Processing Unit, into space on a SpaceX rocket.
- The satellite will test the TPU's ability to function in space, requiring continuous power and cooling.
- Google envisions a network of 81 satellites flying in close formation for parallel processing.
- Google's research estimates Starship needs 1,800 launches over ten years to achieve launch prices near $200 per kilogram.
- The company's chips are expected to handle large inference workloads in orbit for a five-year lifespan.
- Google is a significant investor in SpaceX.
Google has launched its first advanced AI chip into space, a Tensor Processing Unit (TPU) developed as a competitor to Nvidia's GPUs, aboard a SpaceX rocket from California. This mission, managed by Google executive Travis Beals, aims to prove the chip's functionality in the harsh space environment, including its ability to withstand radiation and manage power and thermal demands.
The satellite, built on a platform by Planet Labs, will operate its TPU in 15-minute bursts. Google is also working with Planet Labs on a more purpose-built dual-satellite demo for advanced compute, expected to fly next year, which will test laser communication links between satellites.
Google's long-term vision, dubbed Project Suncatcher, involves creating large-scale compute clusters in orbit, envisioning a network of 81 satellites flying in close formation. This ambitious project is contingent on significant advancements in space launch capabilities and cost reductions. Google's research, published in Joule, analyzes the necessary infrastructure, suggesting that SpaceX's Starship vehicle would need to complete approximately 1,800 launches over the next decade to achieve a target launch price of $200 per kilogram. This projection assumes Starship can carry 200 metric tons per mission and fly significantly more frequently than its current rate.
Despite the challenges, Google's research indicates its chips are resilient enough for space, with a low error rate for inference operations. However, large-scale training runs might still pose challenges.