Feasibility of In-orbit Compute

在轨计算的 "可行性"

Space is an unusual computing environment, both in terms of establishing computing resources there, and operating and maintaining them. We thus discuss several related aspects.

太空是一个非同寻常的计算环境，无论是在其中建立计算资源，还是运营和维护它们，都极具挑战。因此，我们在此讨论几个相关的方面。

Weight and volume: We compute a commodity server’s weight and volume relative to that of a Starlink satellite. We use a high-end server, the HPE ProLiant DL325 Gen10 [28]: 64 cores that clock 2.43.35 GHz, up to 2 TB memory, and 15.6 kg weight. Compared to the latest Starlink satellites launched, the weight is 6% of a satellite’s weight, and the volume is 1%. These are significant costs, but not prohibitive.

重量与体积

我们计算了一台商用服务器相对于一颗星链（Starlink）卫星的重量和体积。我们选用了一款高端服务器——HPE ProLiant DL325 Gen10 [28]：它拥有64个主频为 2.4 - 3.35 GHz的核心，最高支持 2 TB内存，重量为 15.6 kg。与最新发射的星链卫星相比，该服务器的重量占卫星总重的 6%，体积占 1%。这些成本虽然可观，但并非高到无法接受。

Radiation hardening: The HPE Spaceborne Computer [29] aboard the International Space Station (ISS; 408 km) is commodity hardware without any specialized hardware frame. Instead, it uses software hardening. Thus, in LEO, especially for orbits below the inner Van Allen radiation belt (outwards from 643 km), it is likely that commodity hardware is sufficient, although this is not yet a fully settled question.

抗辐射加固

部署在国际空间站（ISS；高度 408 km）上的HPE Spaceborne Computer [29] 采用的是商用硬件，没有任何专门的硬件框架，而是使用软件加固技术。因此，在LEO轨道，特别是低于内范艾伦辐射带（从 643 km向外延伸）的轨道上，商用硬件可能已足够，尽管这个问题尚未有完全定论。

Power: Based on rough estimates of a Starlink satellite’s solar array size and solar efficiency numbers for the ISS, the average solar power output available is estimated to be around 1.5 kW [49]. (Satellites use batteries for continuous operation, given that substantial orbital time is spent in the Earth’s shadow.) The HPE server operating at 225 W (350 W) would consume 15% (23%) of this power. This is quite large, but if necessary, lower wattage servers could be used, or a larger solar array and battery could be built-in to support this additional power requirement. This would, however, come at the cost of additional payload weight and volume.

It is also unclear how the addition of compute skews power usage over time, e.g., due to spikes in communication demands coinciding with spikes in compute demands. If a satellite’s power use fluctuates more due to this, it may create additional challenges in power management beyond the average output over time.

Another related problem is the increased heat generation. Heat is harder to dissipate without an atmosphere, so additional radiators, or thermoelectric harvesters [54] may be necessary. However, as noted above, such additional components will increase the mass per satellite, and hence may result in fewer satellites per launch.

功耗

根据对星链卫星太阳能电池阵列尺寸的粗略估计以及国际空间站的太阳能效率数据，其平均太阳能输出功率估计约为 1.5 kW [49]。（考虑到卫星有相当一部分轨道时间在地球阴影区度过，它们使用电池来维持持续运行。）一台在 225 W (350 W) 功率下运行的HPE服务器将消耗掉其中 15% (23%) 的电力。这个比例相当大，但如有必要，可以选用功耗更低的服务器，或者通过内置更大的太阳能电池阵列和电池来支持这部分额外的电力需求。然而，这将以增加有效载荷的重量和体积为代价。

目前还不清楚增加 计算负载会如何使功耗随时间发生变化 ，例如，通信需求峰值与计算需求峰值同时出现的情况。如果卫星的功耗因此波动更大，除了平均输出功率外，还可能给电源管理带来额外的挑战。

另一个相关问题是 热量产生的增加。在没有大气的情况下，散热更为困难 ，因此可能需要额外的散热器或热电收集器[54]。然而，如上所述，这些额外组件会增加单颗卫星的质量，从而可能导致每次发射的卫星数量减少。

Life-cycle: Starlink satellites will have a life of ∼5 years [44]. This is a bit longer than the typical data center server life of 3 years [33]. Of course, if a satellite-server malfunctions before its expected life, unlike in a data center, it would not be replaced immediately. However, operators continually replenish their satellite fleet, and maintain backup satellites per orbit. Thus, even with a substantial fraction of servers failing, a large LEO constellation could continue to provide valuable in-orbit computing resources.

星链卫星的寿命约为5年[44]。这比数据中心服务器3年的典型寿命要长一些 [33]。当然，如果卫星服务器在其预期寿命前发生故障，与数据中心不同，它无法被立即更换。但是，运营商会持续补充其卫星舰队，并在每个轨道上维护备用卫星。因此，即使有相当一部分服务器失效，一个大型LEO星座仍能继续提供有价值的在轨计算资源。

replenish 补充; fleet 舰队.

Cost: The relative cost of adding compute to a satellite is already accounted for in terms of weight and volume, as the server is much cheaper than the cost of launching its weight. But in absolute terms, how does in-orbit compute compare to terrestrial compute in terms of cost? Based on the per-kilogram launch cost for the Falcon 9 rockets used for Starlink launches [31], and the 15.6 kg server weight, the cost of launching the server is ∼42,000 USD. The perserver total cost of ownership for a data center is estimated to be roughly 5,000 USD per year [33]. If we assume the satellite-server is also used for only 3 years instead of 5, then over 3 years, a coarse estimate for a satellite-server would be roughly 3× as expensive as a data center server. This low figure may be a bit deceiving: many of the data center costs are incurred for the scale (physical sites, power infrastructure, IT and support staff, etc.) but the benefits of scale are not priced here.

成本

将计算能力添加到卫星上的相对成本已经通过重量和体积体现出来，因为服务器本身的价格远低于发射同等重量的成本。但从绝对值来看，在轨计算的成本与地面计算相比如何？根据用于星链发射的猎鹰9号（Falcon 9）火箭的每公斤发射成本[31]和服务器 15.6 kg的重量，发射这台服务器的成本约为 42,000 美元。数据中心每台服务器的总拥有成本估计约为每年 5,000 美元[33]。如果我们假设卫星服务器也只使用3年而非5年，那么在3年内，对一台卫星服务器的粗略估计成本将大约是数据中心服务器的3倍。

这个较低的数字可能有些误导性：数据中心的许多成本是为规模而产生的（物理站点、电力基础设施、IT和支持人员等），但规模带来的效益并未在此计入价格。

Summary: While there are surely other factors that would need careful accounting for a more precise feasibility study, our analysis is cautiously encouraging. A server’s power-draw could be a substantial burden, and the cost of compute is several times that of terrestrial facilities. Nonetheless, if operators and customers perceive enough value in the use cases, these barriers may not be insurmountable.

总结

尽管更精确的可行性研究肯定还需要对其他因素进行仔细核算，但我们的分析结果是审慎乐观的。服务器的功耗可能是一个巨大的负担，并且其计算成本是地面设施的数倍。尽管如此，如果运营商和客户认为这些应用场景具有足够的价值，这些障碍可能并非不可逾越。