跳到正文
DEV Community· The daily flare·· 18 天前AI 評分67

Google 首枚 Suncatcher 衞星在軌運行 AI 任務並傳回數據

Google’s first Suncatcher satellite runs an AI workload in orbit

AI 導讀

Google 的 Project Suncatcher 原型衞星已在軌道上運行 AI 任務,並把數據傳回地球。Google 表示,發射前載入衞星的 Google DeepMind Gemma 4 已完成首次成功 AI 回應;衞星搭載四枚 TPU,測試包括生成太空俳句和回報遙測資料。這些初步任務只測試系統能否生成回應並傳回地球,未能反映硬件執行更長或要求更高的計算任務時的表現。

正文

Originally published on The Daily Flare.

Google’s experimental Project Suncatcher satellite has run an AI task in orbit and sent data to Earth, Computerworld reported Nov. 3, after the craft’s Oct. 1 launch.

The prototype, named MVP, launched aboard a SpaceX Falcon 9 rocket as part of the Transporter-18 rideshare mission. Heise reported that Google brought the satellite’s AI accelerators and other onboard hardware online and conducted initial machine-learning tests after launch. The first data generated by the satellite’s onboard software and hardware reached Earth after several weeks of commissioning.

The spacecraft carries four Tensor Processing Units, or TPUs, Google’s chips for running AI models. Data Center Dynamics reported that solar panels supply up to one kilowatt of electricity to the system. The satellite is a test platform for running those processors in orbit, rather than a full-scale computing facility.

On Oct. 30, Google announced that its Gemma 4 model, developed by Google DeepMind and loaded onto the satellite before launch, had completed its first successful AI response. During a ground-station pass, engineers prompted it to write a haiku about space. The satellite sent back: “Blue marble below / Data streams in silent grace / Starlight feeds my core.”

Google also tested space-themed limericks and prompts asking the model to report telemetry about the satellite’s systems, Data Center Dynamics reported. These early tasks tested whether the onboard system could generate responses and transmit them to Earth. They do not establish how the hardware will perform on longer or more demanding computing jobs.

Reports give different estimates for how long the processors can run at a time. TechRepublic said the satellite’s available power reportedly limits operation to one TPU for about 20 minutes. Heise, citing The New York Times, reported that the accelerators can operate for 15 minutes before they must be shut down to cool. Google’s next tests will examine thermal stress and radiation exposure, both of which can affect the processors in orbit.

Heat management is a particular engineering concern because there is no air in space to carry heat away. Before launch, Google said it was working on approaches including heat pipes and heat sinks. The orbital tests will show how the equipment performs under actual space conditions. Heise also cited challenges for the broader concept, including the cost of transporting equipment and limits on data transmission back to Earth.

Google plans to test two interconnected satellites in 2027, according to Data Center Dynamics. TechRepublic reported that the planned spacecraft are expected to demonstrate optical links for direct satellite-to-satellite communication and have greater power-generation capabilities. Google’s longer-term vision, as reported by Data Center Dynamics, is a cluster of as many as 81 satellites sharing computing work.

The economics remain a major obstacle. Google’s global head of energy and power, Amanda Peterson Corio, said space data centers are “incredibly, incredibly expensive, up to 100 times more expensive than building on Earth,” Data Center Dynamics reported. She added: “Let’s explore all the terrestrial options before we go extraterrestrial.”

Google previously estimated that launch costs would need to fall substantially for orbital computing to compete with terrestrial facilities. The company said a launch price of $200 per kilogram to low Earth orbit could make launch costs, spread over a spacecraft’s lifetime, roughly comparable with data-center energy costs per kilowatt. Google speculated that this level might be reached by 2035 if SpaceX’s Starship begins flying and conducts 180 launches a year. That comparison remains a projection, and the cost of terrestrial data-center power could also change by then.

來源:DEV Community · dev.to