谷歌近日成功将首颗搭载张量处理器（TPU）芯片的轨道计算卫星送入太空1。这颗由Planet Labs建造的卫星将在轨道上验证谷歌先进芯片在空间环境中的运行能力，其中TPU芯片设计为在15分钟内突发运行，以避免在长期工作中产生超负荷1。
谷歌的Suncatcher项目计划建立一个包含81颗卫星的轨道网络，用于并行处理和太空数据中心应用1。根据谷歌发布的白皮书分析，要实现成本有效的轨道数据中心规模化部署，SpaceX的Starship火箭需要完成约1,800次发射，平均每年180次，周期约10年1。谷歌预测到2035年，SpaceX火箭的发射成本可降至约每千克200美元1。
从技术角度看，TPU芯片在太空中的逻辑电路错误率约为百万分之一，足以支撑五年卫星寿命内的推理工作负载1。
Google has successfully deployed its first orbital compute satellite equipped with tensor processing units (TPUs) to space, marking a significant step toward the company's vision of distributed cloud computing infrastructure in orbit 1. The satellite, built by Planet Labs, carries Google's advanced chips designed to validate their performance in the space environment 1.
The mission supports Google's Suncatcher project, which aims to establish an orbital network of 81 satellites for parallel processing 1. According to a white paper released by Google, the company estimates that SpaceX's Starship rocket will need to complete approximately 1,800 launches over the next decade—an average of 180 launches annually—to achieve cost-effective scaling of orbital data centers 1. Google projects that by 2035, SpaceX rocket prices could decline to around $200 per kilogram, a threshold necessary for the economics of space-based computing to become viable 1.
The TPU chips deployed on the satellite are engineered to operate in short bursts lasting 15 minutes to prevent overload 1. Early testing indicates the processors experience logical circuit error rates of approximately one per million, sufficient to sustain inference workloads across a five-year satellite lifespan 1.
评论
还没有评论，欢迎留下第一条。