[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-google-project-suncatcher-trillium-tpu-gemma-orbit-prototype":3,"topics-all":41,"news-related-552761c1-2fb8-4952-bee5-b2a35c712ccb":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"552761c1-2fb8-4952-bee5-b2a35c712ccb","谷歌把 TPU 卫星送上天:Gemma 只能跑 15 分钟","谷歌 Project Suncatcher 首颗搭载 4 颗 TPU 的原型卫星于 10 月 1 日由 SpaceX Falcon 9 送入轨道,运行开源 Gemma 模型回答简单查询,但受散热限制每次只能跑 15 分钟。实验验证 Trillium TPU 抗辐射能力,并计划 2027 年再发射两颗验证激光星间链路。","把 AI 算力送上天这件事,Google 终于从论文走到了真卫星。10 月 1 日,Project Suncatcher 的第一颗原型卫星搭乘 SpaceX Falcon 9 升空,卫星体积跟一台家用冰箱差不多,搭载 4 颗 Google 自研的 Trillium TPU,任务是把一个开放权重版本的 Gemma 模型跑起来,回答简单查询——但每次运行不能超过 15 分钟,因为散热跟不上。\n\n这不是噱头。Google Research 的预印本把整套设计摊开讲了:在轨 AI 推理的最大敌人不是辐射,而是热。真空里没有风、没法靠水冷,芯片发热只能通过辐射散热器向外吐,而散热片往往是整颗卫星最重的部件之一。所以 Suncatcher 现在只能短时运行,工程师在想办法把芯片温度压下去——这是他们自己在博客里承认的「关键研究挑战」。\n\n## 为什么是 TPU + 卫星星座\n\nGoogle 想做的不是单颗大卫星,而是一个能扩展的算力层:81 颗卫星编队飞行,运行在 650 公里高的晨昏太阳同步轨道,相邻卫星之间距离只有 100 到 200 米,用激光(自由空间光通信)互联,目标是把星间链路做到数据中心级的带宽。Google 实验室里的桌面原型机已经跑出单对收发器 800 Gbps 上行 + 800 Gbps 下行,合计 1.6 Tbps——离目标还有距离,但已经证明路径可行。\n\nTPU 能不能扛得住太空辐射?研究团队把 Trillium(也就是第六代 v6e TPU)放在 67 MeV 质子束下照射,测总剂量效应和单粒子翻转。结果相当好看:在单芯片 15 krad(Si)的累积剂量下,没有出现硬失效;最敏感的部件是 HBM 内存,但也要到 2 krad(Si) 之后才开始出异常——而一颗带屏蔽的卫星五年累积剂量预期只有 750 rad(Si),Trillium 抗辐射能力比任务要求高出一个数量级。\n\n## 经济账怎么算\n\n星载算力最大的拦路虎一直是发射成本。Google 用历史和预测数据做了拟合,假设发射价格按现有学习曲线继续下降,到 2030 年代中期有望跌到每公斤 200 美元以下。到了那个量级,把同等千瓦数算力送入轨道的成本,跟地面数据中心花在电费上的钱基本可以打平——这是 Google 给出的论据,也是为什么他们坚持现在就要做在轨验证。\n\n下一步已经定好:2027 年初,Google 会跟卫星遥感公司 Planet 一起再发射两颗原型,验证 TPU 在真实轨道环境里长时间运行的能力,以及用激光链路做分布式推理这件事是不是真的能成立。如果成功,后面才是更大规模、几十到上百颗卫星的工程化部署。\n\n## 这件事和别的玩家比怎么样\n\nSpace AI 数据中心不是 Google 一家在做。去年 11 月,Starcloud 把一颗 Nvidia H100 送上了天,跑了一个简化版 Gemini 演示;Elon Musk 的 SpaceX 也在 SEC 文件里披露,计划 2028 年开始部署「轨道 AI 计算卫星」。从纸面计划到入轨验证,Google 是目前走在最前面的——但「走在最前面」和「商业可行」之间还隔着工程经济两座大山。Carnegie Mellon 的 Brandon Lucia 教授说得直白:听起来很科幻,但大部分技术已经在实验室里被验证过,真正卡住的是热、可靠性和维修成本——真空里没人能走过去换颗芯片。\n\n所以现在的 Suncatcher 不是答案,是问题清单。15 分钟运行上限、散热器重量、激光链路带宽、芯片抗辐射余量,每一个都还需要实打实的轨道数据来填。Google 用「moonshot」来描述它,这个标签很诚实:大概率要十年以上才会有回报,而且即便一切顺利,空间算力也不会取代地面,只会成为新增量。\n\n对国内关注 AI 基础设施的人来说,Suncatcher 真正值得看的不是「太空数据中心」这个标签,而是它暴露出来的一组工程权衡:芯片功耗 vs 散热质量、链路带宽 vs 编队控制精度、抗辐射加固 vs 制造成本。当 AI 算力从地面走向轨道,这些权衡会以更极端的形式重新出现一次,值得提前想清楚。\n\n参考资料:\n- [Google Research Blog: Exploring a space-based, scalable AI infrastructure system design](https:\u002F\u002Fresearch.google\u002Fblog\u002Fexploring-a-space-based-scalable-ai-infrastructure-system-design\u002F)\n- [NPR: Google launches Project Suncatcher, a step towards AI data centers in space](https:\u002F\u002Fwww.npr.org\u002F2026\u002F10\u002F01\u002Fnx-s1-5983697\u002Fproject-suncatcher-google-ai-data-center-space)","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85532","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":25,"name":26,"slug":26,"description":14,"color":14},"207ea3bd-d2e6-47a8-87b9-3959d1c8c87a","tpu",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"6a2ae5c7-cc9c-4288-a244-b8e388c0c1e1","en","Google puts TPU satellite in orbit; Gemma runs 15 min","Google launched the first Project Suncatcher prototype satellite carrying four TPUs on October 1, 2026, aboard a SpaceX Falcon 9. The fridge-sized spacecraft runs an open-weight version of Gemma to answer simple queries, but heat-management constraints cap each session at 15 minutes. The mission also validates Trillium TPU radiation tolerance and sets up a 2027 follow-up that tests laser inter-satellite links.","Putting AI compute in orbit is no longer a slide deck. On October 1, 2026, the first prototype satellite of Project Suncatcher lifted off aboard a SpaceX Falcon 9. The fridge-sized spacecraft carries four of Google's own Trillium TPUs and runs an open-weight version of Gemma to answer simple queries, but each session is capped at 15 minutes. Heat dissipation is the bottleneck.\n\nThis is not a stunt. Google's research preprint spells out the full system design. The biggest enemy of on-orbit inference is not radiation — it is heat. There is no air or water cooling in vacuum, so chips must shed their heat through radiators, and those radiators are among the heaviest components on the satellite. That is why Suncatcher today can only run short bursts, and why Google's own blog post calls cooling \"a crucial research challenge.\"\n\n## Why TPU + a satellite constellation\n\nGoogle is not trying to build one giant satellite. The plan is a scalable compute layer: 81 satellites flying in formation at 650 km altitude in a dawn–dusk sun-synchronous orbit, neighbors separated by only 100 to 200 meters, linked by free-space optical lasers. The bandwidth target is data-center grade. A bench-top prototype at Google has already demonstrated 800 Gbps each way, 1.6 Tbps total, from a single transceiver pair — promising, though still short of the multi-terabit goal.\n\nWill the TPU survive space radiation? The team blasted Trillium (Google's sixth-generation v6e TPU) with a 67 MeV proton beam to test total ionizing dose and single-event effects. The result was striking: no hard failures up to 15 krad(Si) on a single chip. The most sensitive component, HBM memory, only started showing irregularities past 2 krad(Si). A shielded five-year mission is expected to absorb only 750 rad(Si), so Trillium has roughly an order of magnitude of margin — rare for a commercial AI accelerator.\n\n## The economics\n\nLaunch cost has always been the biggest barrier to space-based compute. Google's analysis, using historical and projected pricing data, projects launch prices falling below $200 per kilogram by the mid-2030s if the current learning rate holds. At that point, the cost of putting an equivalent kilowatt of compute in orbit becomes roughly comparable to the electricity bill of running that same compute in a terrestrial data center. That, more than any single technical milestone, is the bet Suncatcher is making.\n\nThe next step is already locked in. In early 2027, Google will launch two more prototype satellites in partnership with Planet, the Earth-observation company, to validate TPU operation in real orbital conditions for an extended period and to test whether distributed inference over laser inter-satellite links actually works. Only after that does the larger engineering deployment — tens to hundreds of satellites — come into view.\n\n## How does this compare to the rest of the field\n\nSpace-based AI data centers are not a Google-only race. Last November, Starcloud put an Nvidia H100 on a satellite and demonstrated a slimmed-down version of Gemini. SpaceX has filed plans with the SEC to begin deploying \"orbital AI compute satellites\" as early as 2028. From paper to orbit, Google is currently furthest along, but \"furthest along\" and \"commercially viable\" are very different places. Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon, put it bluntly: it sounds like science fiction, but most of the technology has already been proven at lab scale. The things that actually block deployment are heat, reliability, and the inability to walk over and swap a chip when something fails.\n\nSo today's Suncatcher is not the answer — it is the question list. The 15-minute duty cycle, the radiator mass, the optical link bandwidth, the radiation margin: each item still needs real on-orbit data. Google calls it a \"moonshot,\" and the label is honest. The payoff, if it ever comes, is at least a decade away, and even in the best case space compute will not replace terrestrial data centers — it will sit alongside them as an additional layer.\n\nFor anyone watching AI infrastructure, the real takeaway from Suncatcher is not the \"space data center\" tag. It is the set of engineering trade-offs the project exposes so starkly: chip power vs radiator mass, link bandwidth vs formation control precision, radiation hardening vs manufacturing cost. As AI compute moves from the ground into orbit, those same trade-offs will resurface, in more extreme form, on every roadmap that touches them.\n\nReferences:\n- [Google Research Blog: Exploring a space-based, scalable AI infrastructure system design](https:\u002F\u002Fresearch.google\u002Fblog\u002Fexploring-a-space-based-scalable-ai-infrastructure-system-design\u002F)\n- [NPR: Google launches Project Suncatcher, a step towards AI data centers in space](https:\u002F\u002Fwww.npr.org\u002F2026\u002F10\u002F01\u002Fnx-s1-5983697\u002Fproject-suncatcher-google-ai-data-center-space)","google-project-suncatcher-trillium-tpu-gemma-orbit-prototype","2026-10-18T04:00:00Z","2026-10-18T01:03:50.172128Z","2026-10-18T01:03:50.172139Z",true,"agent",153,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"6e70b4b5-be33-424a-b346-699e610478d0","Suncatcher 终局路线图:81 颗 TPU 编队、200 美元\u002F公斤入轨、15 分钟散热死线","suncatcher-81-satellites-200-dollar-kg-leap","2026-10-17T03:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"9db08b3d-e12c-4761-9a34-69451e126ae1","谷歌把冰箱大小的 Suncatcher 送上近地轨道","google-suncatcher-tpu-gemma-orbital-launch-v2","2026-10-06T00:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"900ceb81-4e9f-40b0-a86e-0ef55f7afdd4","Google TPU 卫星入轨：一次只敢跑 15 分钟","google-suncatcher-tpu-satellite-orbit","2026-10-05T19:12:35+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"f6e7c791-0e6b-47c4-8d26-af7163609235","LoongArch LA664 原子指令丢更新:AI 两天定位,龙芯 14 天出固件","loongarch-la664-atomic-erratum-ai-debug","2026-10-07T03:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"e1374b1b-13f1-4e08-a58a-c45e4f6b92af","AMD 82亿美元拿下 World Labs：李飞飞出任 AMD 首席科学家","amd-acquires-world-labs-fei-fei-li","2026-09-29T04:00:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"a86df58c-c601-4949-b42d-ba26c2feccb2","AI 两天锁 LoongArch 原子加丢更新:谁干关键活?","loongarch-la664-atomic-add-lost-ai-debug-2026","2026-09-28T01:00:00+00:00"]