谷歌的轨道计算卫星原型今天搭乘从加利福尼亚州发射的SpaceX火箭升空——这是这家科技巨头首次将其先进的芯片送入太空。

由Planet Labs制造的这颗卫星将证明，谷歌的Tensor Processing Unit（TPU，其对标英伟达GPU的产品）能够在太空中正常运行。这意味着要提供一千瓦的持续电力、为芯片散热，并运行一系列模型以检验是否会出现任何问题。

“我们在地面上进行了测试，但你知道，没有任何测试能完全等同于真实环境，”谷歌高管Travis Beals说，他负责管理Project Suncatcher项目，即该科技巨头在地球轨道上开发大规模计算集群的计划。

一旦投入使用，这颗卫星将以每15分钟为间隔启动其TPU，以避免对卫星的电源和热管理系统造成压力。这颗卫星基于Planet Labs构建的标准平台，但两家公司正在合作开发一款预计明年发射的演示版本，该版本将使用两颗专为高级计算设计的卫星，以运行更庞大的工作负载。这些未来版本将尝试通过激光通信链路进行协作。

Suncatcher并非此次SpaceX火箭上唯一的太空AI载荷，该火箭将搭载超过100种不同的有效载荷，包括Satlyt和Cowboy Space Company的任务。

将谷歌的这项计划与那些初创公司（甚至包括SpaceX本身）区分开来的，是它是一个长期项目。

正如Beals所言，这一“长期登月计划”的重点在于构建面向未来的太空基础设施和AI工作负载。该公司设想了一个由81颗卫星组成的轨道数据中心网络，这些卫星以紧密编队飞行并进行并行处理。

“当你试图运行多机架工作负载时，TPU之间的带宽和延迟确实非常关键……我们试图展望的不仅仅是当前存在的工作负载，而是五年后它们将处于何种状态，”Beals说。这主要是因为目前尚不存在能够以成本效益方式扩展轨道数据中心规模的火箭。

周四，谷歌还发布了一份经过同行评审的关于轨道数据中心的白皮书版本，这是目前关于计算如何进入太空的最严谨分析之一。该论文将发表在《Joule》杂志上。

该论文最引人注目的方面之一是谷歌如何看待太空准入问题。尽管研究人员强调他们的分析并非经济可行性研究，但它提供了一幅有趣的画面，展示了该公司如何看待火箭成本随时间推移而降低。

与所有数据中心公司一样，谷歌正指望SpaceX将其航天器送入地面以上。（谷歌也是SpaceX的主要投资者。）

论文作者认为，自发射猎鹰1号（Falcon 1）火箭以来，埃隆·马斯克旗下的火箭制造商实现了每年约20%的价格下降“学习曲线”，因此有理由预计到2035年，该公司能将发射价格降至接近每公斤200美元。

要做到这一点需要付出什么努力？基于猎鹰9号（Falcon 9）发射的有效载荷量，他们认为类似的成本降低轨迹将需要星舰（Starship）将37万吨有效载荷送入轨道。这需要它在未来10年内进行约1,800次发射，即每年180次——前提是每次任务都能飞行200公吨的有效载荷。

对于一款从未在一年内飞行超过5次的飞行器来说，这是一个巨大的要求。SpaceX预测该公司将飞行远超此数量的次数——例如，埃隆·马斯克曾表示星舰有望在2029年实现每小时飞行的频率，但马斯克说过很多事情。

至少，在谷歌最新的研究中有一个好消息：其芯片似乎能够承受太空辐射。此前，谷歌意识到芯片的配置提供的屏蔽效果比实际环境中更强，因此不得不重新在粒子加速器中对芯片进行辐射测试。这导致芯片的逻辑电路出现了略多的错误，但公司仍对其芯片能够在卫星五年的使用寿命内，在轨道上处理大规模推理工作负载充满信心。

“如果你考虑的是典型的推理操作，那么错误率非常低，对吧？比如百万分之一。”比尔斯说，“另一方面，如果要进行大规模的模型训练，比如让数千块芯片连续运行数月，那本来就是个棘手的问题。”

更正：本报道的标题最初将星舰发射的学习曲线估算值误写为1,600；实际应为1,800。

Google’s prototype of its orbital compute satellite took off today onboard a SpaceX rocket launched from California — the first time the tech giant has sent one of its advanced chips into space.

Built by Planet Labs , the satellite will prove that a Google Tensor Processing Unit, its competitor to Nvidia’s GPUs, can function in space. That means supplying a kilowatt of continuous power, cooling the chip, and running a series of models through their paces to see if anything goes wrong.

“We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing,” said Travis Beals, the Google executive managing Project Suncatcher , the tech giant’s plan to develop large-scale compute clusters in orbit around the Earth.

Once commissioned, the satellite will fire up its TPU in 15-minute bursts to avoid straining the satellite’s power and thermal management systems. This satellite is based on a standard platform built by Planet Labs, but the two companies are working on a demo expected to take flight next year that will see two satellites more purpose-built for advanced compute that can run more substantial workloads. Those future versions will attempt to collaborate via a laser communications link.

Suncatcher isn’t the only space AI payload on this SpaceX rocket, which is launching more than 100 different payloads, including missions from Satlyt and Cowboy Space Company.

What sets the Google initiative apart from those startups (and indeed from SpaceX itself) is that it’s a long-term project.

The focus of this “long-term moonshot,” as Beals puts it, is on building for the space infrastructure and AI workloads that will exist in the future. The company envisions an orbital data center that is a network of 81 satellites flying in close formation, processing in parallel.

“The bandwidth and the latency between TPUs really, really matters when you’re trying to run a multi-rack workload…we’re trying to look ahead to not just what workloads exist today, but where they will be in five years,” Beals said. That’s largely because the rockets required to scale up orbital data centers in a cost-effective way don’t yet exist.

On Thursday, Google also released a peer-reviewed version of its white paper on orbital data centers, one of the most rigorous analyses available of how compute gets to orbit. The paper will be published in Joule .

One of the paper’s most notable aspects is how Google thinks about access to space. Although the researchers stress their analysis isn’t an economic feasibility study, it offers an interesting picture of how the company sees rockets becoming cheaper over time.

Like all data center companies, Google is looking to SpaceX to get its spacecraft off the ground. (Google is also a major investor in SpaceX.)

Arguing that Elon Musk’s rocket builders have achieved a price-reducing “learning curve” of about 20% a year since they launched the Falcon 1 rocket, the authors believe it’s reasonable to expect the company to deliver launch prices close to $200 per kilogram by 2035.

What will it take to do that? Based on the amount of payload launched by the Falcon 9, they think a similar cost-reduction trajectory will require Starship to fly 370,000 tons of payload into orbit. That’s something that would take it about 1,800 launches over the next 10 years, or 180 a year — and that’s if it can fly 200 metric tons on each mission.

That’s a big ask for a vehicle that has never flown more than five times in a year. SpaceX predicts the company will be flying far more than that — Elon Musk has suggested Starship could achieve an hourly flight rate in 2029, for example, but Musk says a lot of things.

The good news, at least, in Google’s updated research, is that it seems likely that its chips will survive the radiation of space. The company had to redo tests blasting the chips in a particle accelerator after they realized the configuration of the chips provided more shielding than they would actually experience. This produced slightly more errors in the chip’s logic circuitry, but the company is still confident its chips can handle large inference workloads in orbit for the five-year lifespan of a satellite.

“The error rate is very low if you’re thinking about typical inference operations, right? Like one in a million,” Beals said. “On the other hand, it was already problematic for doing, say, some mega-scale training run where you’re going to have many thousands of chips running for months.”

Correction: The headline for this story originally misstated the learning curve estimate for Starship launches as 1,600 ; it is 1,800.