Digest · AI clusters and inference in orbit · read after Chapter 11
導讀 · 在軌 AI 叢集與推論 · 第 11 章後閱讀
Toward a Future Space-Based, Highly Scalable AI Infrastructure System Design邁向未來太空式、高度可擴展的 AI 基礎設施系統設計
This digest explains how a close formation supports high-rate optical links and how near-field propagation limits spatial multiplexing. It separates the paper’s link evidence from the cluster-network design that remains open.讀完後，你能說明 Google 的叢集為什麼讓衛星彼此相距 100–200 m、一具 5 W 的發射器如何跨越這段距離承載約 10 Tbps、哪些鏈路預算常數決定了 5 km 的近場界限與多工距離，以及為什麼分析叢集內部網路仍需補充架構輸入。
Main result
主要結果
Suncatcher places accelerators in a close solar-powered formation to support high-rate optical links. The paper quantifies physical subsystems while leaving the cluster network for further design.
Suncatcher 將加速器放在緊密的太陽能編隊，以支援高速光學鏈路。論文量化了物理子系統，叢集網路則留待後續設計。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Capacity and rate | Qualitative scope | Google's Suncatcher paper proposes putting TPU accelerators on solar-powered satellites that fly in a tight formation. Thus, that the satellites can be wired together by optical links short enough to carry data-center bandwidth. |
| Orbital configuration | 81 · 1 km · 650 km | Its illustrative cluster is 81 satellites inside a 1 km radius at a mean altitude of 650 km in a dawn–dusk sun-synchronous orbit |
| Orbital configuration | 100 · 200 m | with neighbors whose distance oscillates between about 100 and 200 m twice per orbit |
| Capacity and rate | 10 · 24 · 400 · 9.6 · 100 · 12.8 · 75 · 800 · 1.6 | The target is about 10 Tbps per link using commercial 400G DWDM transceivers. Twenty-four channels give 9.6 Tbps on a 100 GHz grid. A 75 GHz grid gives approximately 12.8 Tbps. A bench demonstration carries 800 Gbps in one direction, or 1.6 Tbps bidirectionally [36, pp. 3–4, 8]. |
| Power and thermal conditions | 2 krad · 750 rad · 17 rad · 200 · 810 · 570–3,000 | Proton tests found Trillium TPU HBM irregularities from 2 krad(Si), about three times the estimated five-year dose of 750 rad(Si). Silent data corruption occurred about once per 17 rad [36, p. 5]. At $200/kg launch cost, the paper estimates launched power at $810/kW/yr. It compares this with US data-center power spending of $570–3,000/kW/yr [36, p. 6, Table 1]. |
| Evidence and scope | Qualitative scope | The peer-reviewed Joule version is the citation of record |
| Evidence and scope | 2026 | This digest distinguishes the Joule article from the June 2026 arXiv v2 preprint. One difference concerns the highest tested dose with zero hard failures. |
| Evidence and scope | 10 krad · 15 krad | which Joule gives as 10 krad(Si) where the preprint gives 15 krad(Si) |
| Orbital configuration | 30 m | and in a formation-maintenance budget of under 30 m/s of ΔV per year that only Joule states [36, p. 5]. |
| Latency comparison | 8 | Part III follows an open network-design question. The paper specifies per-link rate, an eight-neighbor lattice, and a ΔV budget. Terminal count, switching, routing, and the latency budget need further specification. |
| 模型與運算 | 定性範圍 | Google 的 Suncatcher 論文提議把 TPU 加速器放在以太陽能供電、緊密編隊飛行的衛星上，讓衛星之間能以夠短的光學鏈路相連，短到足以承載資料中心等級的頻寬。 |
| 軌道配置 | 81 · 1 km · 650 km · 100 · 200 m | 它的示範叢集是 81 顆衛星，位於半徑 1 km 的範圍內，平均高度 650 km 的晨昏太陽同步軌道，相鄰衛星的距離每圈兩次在約 100 與 200 m 之間振盪。 |
| 容量與速率 | 10 · 24 · 400 · 100 · 9.6 · 75 · 12.8 · 800 · 1.6 | 目標是每條鏈路「10 Tbps 量級」的總頻寬，由 24 個 DWDM 通道的商用 400G 收發器組成（100 GHz 柵格為 9.6 Tbps，75 GHz 柵格約 12.8 Tbps），並有單向 800 Gbps（雙向 1.6 Tbps）的實驗台展示作為支撐[36, pp. 3–4, 8]。 |
| 模型與運算 | 2 krad · 750 rad · 17 rad | 對 Trillium TPU 的質子測試發現 HBM 要到 2 krad(Si) 才出現異常，約為五年任務劑量 750 rad(Si) 的三倍，靜默資料損毀約每 17 rad 一次[36, p. 5]。 |
| 電力與熱條件 | 200 · 810 · 570–3,000 | 經濟分析一節則主張在每公斤 200 美元的發射價格下，每 kW 每年 810 美元的「發射電力價格」可與美國資料中心每 kW 每年 570–3,000 美元的電力支出相比[36, p. 6, Table 1]。 |
| 證據與範圍 | 2026 · 6 · 10 krad · 15 krad · 30 m | 經同儕審查的 Joule 版本是正式引用版本，它與 2026 年 6 月的 arXiv v2 預印本有幾處相異，本導讀都會標出，其中最明顯的是硬性失效計數為零的最高劑量，Joule 寫 10 krad(Si)，預印本寫 15 krad(Si)，以及只有 Joule 給出的編隊維持預算：每年 ΔV 低於 30 m/s[36, p. 5]。 |
| 證據與範圍 | 8 | 決定第三部分走向的附註是：論文給了每條鏈路的速率、一個 8 鄰居晶格與一個 ΔV 預算，每顆衛星的終端數、交換架構、路由與延遲預算則待指定，所以叢集的網路是待解的問題，後續仍需完成設計。 |
1. Where this paper sits1. 這篇論文的位置
Suncatcher reaches the training-network range through shorter geometry. This digest uses Joule as the citation of record and tracks its differences from the arXiv version.
Suncatcher 透過縮短幾何距離進入訓練網路的速率範圍，本導讀以 Joule 為正式引用版本，並追蹤其與 arXiv 的差異。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | This is the central case of Part III and the only orbital-compute proposal whose inter-satellite link is in the right order of magnitude for training. |
| Evidence and scope | 10 | Chapter 10 lines up the link rates the satellite literature assumes |
| Capacity and rate | 16 Mb/s · 0.5 Gbit/s · 100 · 200 · 400 · 3 · 6.4 | Related assumptions range from 16 Mb/s for satellite federated learning to 0.5 Gbit/s and ≥100 Gbps in inference preprints. Other studies use 200 Gbps terminals. Meta’s Llama 3 cluster gives each terrestrial GPU 400 Gbps. Google Cloud lists TPU v6e inter-chip capacity at 6.4 Tbps per chip company. These values describe different link and workload scopes. |
| Capacity and rate | 9.6–12.8 | Suncatcher's 9.6–12.8 Tbps per aperture is the one figure that reaches the terrestrial range. It does so by changing the geometry rather than the optics. |
| Evidence and scope | Qualitative scope | The paper exists in two versions that a careful reader must keep apart. |
| Evidence and scope | 2511.19468 · 4 · 2025 · 102678 · 1 · 2026 | The preprint, arXiv 2511.19468, appeared with Google’s 4 November 2025 blog post. Its title is “Towards a future space-based, highly scalable AI infrastructure system design.” Joule published the peer-reviewed version as article 102678 on 1 October 2026 [36]. Its title changes “Towards” to “Toward.” The first prototype launched that day. |
| Cost assumptions | 2026 | Read side by side, the June 2026 arXiv v2 text and the Joule article carry the same link budget, radiation rates and launch-cost analysis |
| Orbital configuration | 100 km | Joule adds formation-maintenance ΔV, altitude trade-offs involving eclipse and the inner Van Allen belt, and thermal discussion. It also considers joining clusters separated by over 100 km. |
| Evidence and scope | 15 · 10 krad | and it lowers the stated maximum test dose from 15 to 10 krad(Si) |
| Evidence and scope | Qualitative scope | so this digest cites Joule for every number and marks where the arXiv text differs. |
| 證據與範圍 | 定性範圍 | 這是第三部分的核心案例，也是唯一一個星間鏈路落在訓練所需數量級的在軌運算提案。 |
| 容量與速率 | 10 · 16 Mb/s · 0.5 Gbit/s · 100 · 200 · 3 · 400 · 6.4 | 第 10 章把衛星文獻假設的鏈路速率排成一列：從衛星聯邦學習的 16 Mb/s，經推論預印本的 0.5 Gbit/s 與「≥100 Gbps」，到 200 Gbps 的終端，全都低於 Meta Llama 3 叢集中單一地面 GPU 獨享的 400 Gbps，更遠低於 Google Cloud 文件所列單一 TPU v6e 晶片間互連的 6.4 Tbps company。 |
| 容量與速率 | 9.6–12.8 | Suncatcher 每個孔徑 9.6–12.8 Tbps 是唯一進入地面範圍的數字，而它靠的是縮短鏈路幾何，光學元件則使用商用技術。 |
| 證據與範圍 | 定性範圍 | 這篇論文有兩個版本，細心的讀者必須把它們分開。 |
| 證據與範圍 | 2511.19468 · 2025 · 11 · 4 | 預印本 arXiv 2511.19468〈Towards a future space-based, highly scalable AI infrastructure system design〉隨 Google 2025 年 11 月 4 日的部落格文章發布。 |
| 證據與範圍 | 102678 · 2026 · 10 · 1 | 經同儕審查的版本〈Toward a future space-based, highly scalable AI infrastructure system design〉則以 Joule 第 102678 號文章[36]於 2026 年 10 月 1 日上線，正是第一顆原型發射的那一天。 |
| 軌道配置 | 2026 · 6 · 100 km · 15 · 10 krad | 把 2026 年 6 月的 arXiv v2 內文與 Joule 文章並排來讀，兩者的鏈路預算、輻射速率與發射成本分析相同，而 Joule 另外加入了編隊維持的 ΔV 預算、高度在地影與內范艾倫帶之間的取捨、一段熱管理的說明，以及叢集可再以相距超過 100 km 的鏈路組成「叢集的叢集」的說法，並把所述的最高測試劑量從 15 降為 10 krad(Si)，所以本導讀的每個數字都引用 Joule，並標出 arXiv 內文相異之處。 |
2. The problem, measured2. 問題，用量測說話
Nearly continuous sunlight motivates orbital compute, which then needs high-bandwidth interconnects. Close formation reduces optical divergence enough to support multiple commercial transceiver channels.
近乎持續的日照提供在軌運算的動機，運算接著需要高頻寬互連，緊密編隊則降低光學發散，支援多個商用收發器通道。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Power and thermal conditions | Qualitative scope | The problem the authors set themselves is power. |
| Orbital configuration | 8 | A dawn–dusk sun-synchronous orbit gives panels nearly continuous sunlight. The paper estimates annual yield up to eight times that of a mid-latitude ground panel [36, p. 2]. Pénot and Balakrishnan separately estimate approximately five times preprint. Each multiplier requires its own operating assumptions. |
| Capacity and rate | 1 · 1–100 | Orbital power creates a network-design challenge for large-model training. Accelerators need terabit-per-second connections. Conventional optical ISLs span thousands of kilometers. The paper describes about 1 µW received power and 1–100 Gbps rates at those ranges [36, p. 3]. |
| Power and thermal conditions | 1 | The authors' key observation is that received optical power falls as 1/d² once a beam has left its near field |
| Power and thermal conditions | Qualitative scope | so moving the satellites from thousands of kilometers apart to hundreds of meters raises received power by many orders of magnitude |
| Power and thermal conditions | 5 W · −20 · 10 · 1 | A commercial 5 W amplifier can supply a full DWDM channel set at short range. Receivers collectively need hundreds of microwatts. Each channel needs about −20 dBm, or 10 µW, compared with about 1 µW received over a long-range link [36, pp. 3, 8]. |
| Orbital configuration | Qualitative scope | Everything else in the design follows from accepting the cost of that move, which is that the satellites must fly in close formation for years. |
| 證據與範圍 | 定性範圍 | 作者給自己設定的問題是電力。 |
| 證據與範圍 | 8 | 晨昏太陽同步軌道上的太陽能板幾乎持續照到太陽，論文估計其年產量最多可達中緯度地面太陽能板的 8 倍[36, p. 2]。 |
| 證據與範圍 | 定性範圍 | 其他作者對這個倍數有相異說法，Pénot 與 Balakrishnan 說「約五倍」preprint，所以引用時一定要附上背後的假設。 |
| 容量與速率 | 1 · 1–100 | 把運算放到電力所在之處，接著就造出一個網路問題，因為訓練大型模型需要以每秒數 Tb 鏈路相連的加速器，而傳統的光學星間鏈路是為數千公里設計的，論文指出那種距離下的接收功率約 1 µW、速率為 1–100 Gbps[36, p. 3]。 |
| 電力與熱條件 | 1 · 5 W · −20 · 10 | 作者的關鍵觀察是：光束一旦離開近場，接收光功率就隨 1/d² 下降，所以把衛星間距從數千公里縮到數百公尺，可以把接收功率提高許多個數量級，於是一具 5 W 的商用放大器就能餵飽一整組 DWDM 通道，這些接收器合計需要數百微瓦，即每通道約 −20 dBm、10 µW，，長程鏈路則僅送達 1 µW[36, pp. 3, 8]。 |
| 證據與範圍 | 定性範圍 | 設計中的其他一切，都來自接受這一步的代價：衛星必須以緊密編隊飛行好幾年。 |
3. The design3. 設計
The bounded HCW solution creates a rotating and stretching lattice. Shape tuning limits perturbation drift, while coherent optics and spatial multiplexing use the short link distances.
有界 HCW 解形成旋轉並拉伸的晶格，形狀微調限制攝動漂移，相干光學與空間多工則利用較短鏈路距離。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 81 · 1 km · 650 km | The illustrative formation has 81 satellites in a planar lattice at mean altitude 650 km, within cluster radius R = 1 km. The arXiv version uses a square layout to simplify description. Joule expects a commercial design to use a hexagonal lattice, potentially with many more satellites [36, pp. 3–4, Fig. 2]. |
| Orbital configuration | 2:1 · 100 · 200 m | In the cluster-center rotating frame, each outer satellite follows a closed ellipse with a 2:1 along-track-to-radial axis ratio. This is the bounded Hill–Clohessy–Wiltshire solution. The cluster changes shape twice per orbit, with neighbor distances varying between about 100 and 200 m [36, p. 4, Fig. 3]. |
| Evidence and scope | Qualitative scope | For a fixed minimum separation the satellite count grows as N ∼ R² |
| Orbital configuration | 2:1.0037 · 3 m | Tuning the axis ratio to 2:1.0037 limits J₂-induced drift below 3 m/s per year per kilometer from the reference orbit. The paper describes this as a simplified numerical estimate. |
| Orbital configuration | 10 | The methods use SciPy’s eighth-order DOP853 integrator for perturbed dynamics. Centimeter accuracy over a 10⁷ m orbit needs about nine correct digits [36, pp. 5, 7–8]. |
| Evidence and scope | 1.55 µm | Commercial DWDM transceivers operate at 1.55 µm. Coherent detection rejects neighboring-beam light. Spatial multiplexing divides one aperture into smaller parallel beams, trading range for capacity [36, pp. 3, 8]. |
| Power and thermal conditions | 5 W · 105.1 · 3 | The budget uses 5 W from an erbium-doped fiber amplifier, antenna gains of 105.1 dB at each end and 3 dB of other losses [36, p. 8]. |
| 軌道配置 | 81 · 650 km · 1 km | 編隊是一個由 81 顆衛星組成、全部位於軌道面內的平面晶格，在平均高度 650 km、叢集半徑 R = 1 km 之內。 |
| 證據與範圍 | 定性範圍 | arXiv 內文說它採正方配置，六角形則留待商用設計，「主要是為了方便描述」，Joule 內文則預期商用設計會改用六角晶格，衛星數也可能多得多[36, pp. 3–4, Fig. 2]。 |
| 證據與範圍 | 2:1 · 100 · 200 m | 在隨叢集中心旋轉的座標系中，每顆外圍衛星描出一個沿軌對徑向軸比 2:1 的封閉橢圓，也就是 Hill–Clohessy–Wiltshire 方程式的有界解，所以整個叢集每圈改變兩次形狀，相鄰距離在約 100 與 200 m 之間擺盪[36, p. 4, Fig. 3]。 |
| 證據與範圍 | 定性範圍 | 在最小間距固定時，衛星數依 N ∼ R² 成長。 |
| 證據與範圍 | 2:1.0037 · 3 m · 10 | 把軸比調成 2:1.0037，可以把地球扁率（J₂）造成的漂移壓在每公里離參考軌道距離每年 3 m/s 以下，論文稱這個數字來自簡化的數值計算，而方法一節則以 SciPy 的八階 DOP853 積分器積分受擾動態，因為在 10⁷ m 的軌道上要達到公分級精度需要約九位正確數字[36, pp. 5, 7–8]。 |
| 證據與範圍 | 1.55 µm | 鏈路採用 1.55 µm 波段的商用 DWDM 收發器，搭配相干偵測讓接收器能排除相鄰光束的光，以及空間多工，即把一個孔徑分成數個較小的平行光束，以距離換取容量[36, pp. 3, 8]。 |
| 電力與熱條件 | 5 W · 105.1 · 3 | 鏈路預算使用摻鉺光纖放大器的 5 W、兩端各 105.1 dB 的天線增益，以及 3 dB 的其他損耗[36, p. 8]。 |
Near-field length and imaging resolution set how aperture subdivision trades reach for parallel channels. The optical constants retain the paper’s best-case assumptions.
近場長度與成像解析度決定了孔徑分割如何以傳輸距離換取平行通道，光學常數則保留論文的最佳情況假設。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | The constants that follow from this budget are the same in the arXiv v2 and Joule texts. |
| Orbital configuration | 5 km · 10 cm · 5 cm · 1.55 µm | A symmetric confocal system places the beam waist midway between transceivers. Its near-field distance is L = πa²/λ. A 10 cm beam, with a = 5 cm at 1.55 µm, gives about 5 km. This best-case estimate excludes beam-quality loss, aperture truncation, and pointing jitter. |
| Evidence and scope | Qualitative scope | and inside that range the minimum aperture for one link grows only as the square root of distance |
| Evidence and scope | Qualitative scope | so more independent links fit into a fixed area as the distance falls [36, p. 8]. |
| Evidence and scope | −20 · 0.24 mW · 24 · 1 · 71 · 196 · 16 · 1.39 · 2 | Receivers need about −20 dBm per channel, or 0.24 mW for 24 channels. Figure 1 uses about 71 photons per bit for on-off keying and 196 for PM-16QAM. The shot-noise Shannon limit is 2 ln 2, about 1.39 photons per bit [36, p. 8]. |
| Power and thermal conditions | 10 cm · 2×2 · 5 cm · 1.25 km · 4×4 · 2.5 cm · 0.32 km · 1 | A 2×2 array divides the 10 cm aperture into 5 cm systems and gives 1.25 km range. A 4×4 array uses 2.5 cm systems and gives 0.32 km. Figure 1 attributes these limits to sub-aperture imaging resolution. |
| Orbital configuration | 24 · 300 km | while 24-way DWDM through the undivided aperture closes the budget out to about 300 km [36, p. 3, Fig. 1]. |
| Orbital configuration | 5 km · 1.0 | At 5 km, limiting beam wander to one tenth of the aperture radius requires about 1.0 µrad pointing accuracy. The authors cite demonstrated commercial-terminal capability at this precision [36, p. 8]. |
| Evidence and scope | 100–200 m · 4×4 | The 100–200 m spacing therefore lies far inside the near field and inside the 4×4 range, as Figure A shows. |
| Evidence and scope | 5 | The radiation and economics halves of the design are summarised in Section 5. |
| 證據與範圍 | 定性範圍 | 由這份預算推出的常數，在 arXiv v2 與 Joule 內文中完全相同。 |
| 軌道配置 | 1.55 µm · 10 cm · 5 cm · 5 km | 在光束腰位於兩具收發器正中央的對稱共焦系統中，近場鏈路距離為 L = πa²/λ，對 1.55 µm、直徑 10 cm（a = 5 cm）的光束約為 5 km，這是忽略光束品質、孔徑截斷與指向抖動的最佳情況。 |
| 證據與範圍 | 定性範圍 | 在這個範圍內，單一鏈路所需的最小孔徑只隨距離的平方根增加，所以距離縮短時，固定面積內能放進更多獨立鏈路[36, p. 8]。 |
| 證據與範圍 | −20 · 24 · 0.24 mW | 收發器每通道約需 −20 dBm，24 通道合計 0.24 mW。 |
| 證據與範圍 | 1 · 71 · 16 · 196 · 2 · 1.39 | 圖 1 各條線背後的光子預算為：開關鍵控每位元約 71 個光子、PM-16QAM 約 196 個，散粒雜訊下的 Shannon 極限 2 ln 2 約 1.39 個[36, p. 8]。 |
| 軌道配置 | 10 cm · 5 cm · 2×2 · 1.25 km | 把 10 cm 孔徑分成 5 cm 系統的 2×2 陣列，可用到 1.25 km。 |
| 軌道配置 | 2.5 cm · 4×4 · 0.32 km | 分成 2.5 cm 系統的 4×4 陣列，可用到 0.32 km。 |
| 軌道配置 | 1 · 24 · 300 km | 圖 1 的圖說指出這些距離受限於各子孔徑的成像解析度，，接收功率仍有餘裕，至於完整孔徑以 24 路 DWDM 可讓預算成立到約 300 km[36, p. 3, Fig. 1]。 |
| 軌道配置 | 5 km · 1.0 | 要在 5 km 處把光束漂移壓在孔徑半徑的十分之一，需要約 1.0 µrad 的指向精度，作者認為這在商用終端已展示的能力範圍內[36, p. 8]。 |
| 證據與範圍 | 100–200 m · 4×4 | 因此如圖 A 所示，100–200 m 的間距遠在近場之內，也在 4×4 的距離之內。 |
| 證據與範圍 | 5 | 設計中輻射與經濟的部分摘要在第 5 節。 |
Original figure note原始圖說出處
Drawn for this primer from Agüera y Arcas et al., Joule 2026, pp. 3–4, 8, Figs. 1–3.4. Evaluation setup4. 實驗設定
The evidence combines orbital simulation, control budgeting, bench optics and ground radiation tests. Flight network performance remains a later validation task.
證據結合軌道模擬、控制預算、光學實驗台與地面輻射測試，在軌網路效能仍待後續驗證。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | The paper combines four kinds of evidence, and network flight testing remains outside all four. |
| Evidence and scope | Qualitative scope | The orbital claims come from numerical integration of the perturbed relative motion |
| Evidence and scope | Qualitative scope | which is simulation |
| Orbital configuration | Qualitative scope | Formation-maintenance modeling includes position and velocity uncertainty, additional accelerations, differential drag, minimum thruster impulse, and thrust jitter. It considers calm and stormy space weather. |
| Evidence and scope | Qualitative scope | which is also simulation |
| Capacity and rate | 800 · 1.6 | Link claims combine budget analysis with a bench demonstrator using commercial components. The short free-space path carries 800 Gbps one way, or 1.6 Tbps bidirectionally. |
| Evidence and scope | Qualitative scope | which is a bench measurement [36, pp. 3, 7–8]. |
| Latency comparison | 10 · 9.6–12.8 · 6.7 µs · 2 km | The per-link target is about 10 Tbps, with DWDM channel calculations giving 9.6–12.8 Tbps. A full latency model and propagation figures require further specification. The calculation table gives about 6.7 µs for light to cross a 2 km cluster. |
| Latency comparison | 8 · 100 km | The topology uses a square lattice with an 8-neighborhood. Workloads needing more chips could join clusters over distances beyond 100 km. Terminal count, switching, routing, hop count and latency remain unspecified [36, pp. 3–4, Fig. 2]. |
| Model and compute | 67 | Radiation tests expose a Google v6e Trillium TPU and AMD host to 67 MeV protons at UC Davis’s Crocker Nuclear Laboratory. They measure total-dose effects and single-event rates per chip [36, pp. 5, 9]. Link outages, satellite loss, and thermal throttling require separate failure models. |
| Evidence and scope | 18 · 1 · 2026 | Flight data remain outside the paper. The Planet-built prototype launched later, on SpaceX Transporter-18 on 1 October 2026, with ground connectivity only company. |
| 證據與範圍 | 定性範圍 | 論文結合了四種證據，網路在軌測試列於這四種證據的範圍之外。 |
| 證據與範圍 | 定性範圍 | 軌道的宣稱來自受擾相對運動的數值積分，屬於模擬。 |
| 證據與範圍 | 定性範圍 | 編隊維持預算來自詳細的數值模擬，假設了位置與速度的誤差範圍、模型範圍外的加速度、差異阻力、推進器最小衝量與推力抖動，並各自考慮平靜與風暴的太空天氣，同樣屬於模擬。 |
| 容量與速率 | 800 · 1.6 | 鏈路的宣稱則來自鏈路預算分析，加上一具以現成元件搭建的實驗台規模展示器，在短程自由空間路徑上達到單向 800 Gbps（雙向 1.6 Tbps），屬於實驗台量測[36, pp. 3, 7–8]。 |
| 容量與速率 | 10 · 9.6–12.8 | 因此 ISL 速率是每條鏈路 10 Tbps 量級的設計目標，9.6–12.8 Tbps 由 DWDM 通道數算出。 |
| 軌道配置 | 2 km · 6.7 µs | 論文的延遲模型與傳播延遲數值待補充，雖然光跨越一個 2 km 的叢集只需約 6.7 µs（見計算表）。 |
| 軌道配置 | 8 · 100 km | 拓樸是帶有 8 鄰居的正方晶格，工作負載需要更多晶片時，叢集還可以用相距超過 100 km 的鏈路組成「叢集的叢集」，而每顆衛星的終端數、交換與路由待指定、跳數待指定，延遲預算則待指定[36, pp. 3–4, Fig. 2]。 |
| 模型與運算 | 67 | 失效模型只有輻射測試：一顆 Google v6e Trillium TPU 連同其 AMD 主機在 UC Davis Crocker 核子實驗室以 67 MeV 質子照射，得到每晶片的總劑量與單事件速率[36, pp. 5, 9]，而鏈路中斷、衛星損失與熱降頻都列於模型範圍之外。 |
| 證據與範圍 | 2026 · 10 · 1 · 18 | 論文飛行資料待補充，因為第一顆原型（與 Planet 合作打造，2026 年 10 月 1 日搭 SpaceX Transporter-18 發射）是在論文之後才飛，而且僅有地面鏈路 company。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| The target is about 10 Tbps per link. Twenty-four 400G PM-16QAM DWDM channels give 9.6 Tbps bidirectionally on a 100 GHz grid. A 75 GHz grid gives approximately 12.8 Tbps per aperture [36, pp. 3, 8].每條鏈路「10 Tbps 量級」：24 個 400G PM-16QAM DWDM 通道在 100 GHz 柵格為雙向 9.6 Tbps，75 GHz 柵格每孔徑約 12.8 Tbps[36, pp. 3, 8]。 | A design calculation. Google's blog says "tens of terabits per second" company, which requires separate evidence. Use the paper’s documented figure.設計計算。Google 部落格寫「每秒數十 Tb」company，仍需獨立證據，應採用論文記錄的數字。 |
| Bench link of 800 Gbps unidirectional, 1.6 Tbps bidirectional [36, p. 3].實驗台鏈路單向 800 Gbps、雙向 1.6 Tbps[36, p. 3]。 | A bench-scale demonstrator of off-the-shelf components over a short free-space path, with motion and pointing-loop control excluded. The path length and the DWDM channel count of the demo are unspecified.以現成元件搭建的實驗台規模展示器、短程自由空間路徑、衛星保持靜止，指向迴路則在展示範圍之外。展示的路徑長度與所用的 DWDM 通道數都待說明。 |
| Neighbour distance 100–200 m, shape cycling twice per orbit. J₂ drift <3 m/s/yr per km with a 2:1.0037 axis ratio [36, p. 4, Fig. 3].相鄰距離 100–200 m，形狀每圈循環兩次。軸比 2:1.0037 時 J₂ 漂移每公里每年 <3 m/s[36, p. 4, Fig. 3]。 | Simulated dynamics. Joule requires neighbor-distance drift below 20% of intended spacing. Formation maintenance needs under 30 m/s ΔV per year in calm and stormy space weather [36, p. 5]. Joule adds this budget beyond arXiv v2’s description of modest delta-v requirements.模擬的動態。Joule 另外指出，要在平靜與風暴的太空天氣下都把鄰居距離的漂移壓在預定間距的 20% 以下，編隊維持每年需要的 ΔV 低於 30 m/s[36, p. 5]。arXiv v2 內文將此項留待量化，只說「適度的 delta-v 需求」。 |
| HBM irregularities start at 2 krad(Si), about three times the estimated 750 rad(Si) five-year dose. The estimate uses approximately 150 rad(Si)/yr behind 10 mm Al. The single-chip test found zero hard TID failures through its maximum dose. Joule gives that maximum as 10 krad(Si), while arXiv v2 gives 15 krad(Si) [36, p. 5].HBM 從 2 krad(Si) 起出現異常，約為 10 mm 鋁屏蔽下每年約 150 rad(Si)、五年 750 rad(Si) 劑量的 3 倍。在單一晶片的最高測試劑量內，可歸因於 TID 的硬性失效計數為零，該劑量 Joule 寫 10 krad(Si)，arXiv v2 寫 15 krad(Si)[36, p. 5]。 | The dose ceiling comes from one chip, with different values in the two versions. Joule reports 10 krad(Si), while arXiv v2 Section 2.3 reports 15 krad(Si) [108, §2.3]. Cite Joule for the 10 krad(Si) ceiling. Identify the preprint when using 15 krad(Si).上限僅來自一顆晶片，而兩個已發表版本對它的說法相異：Joule 的結果一節寫 10 krad(Si)，arXiv v2 的 §2.3 寫 15 krad(Si)[108, §2.3]，所以引用 Joule 時寫 10 krad(Si)，用 15 時要註明出自預印本。 |
| Silent data corruption ≈1 per 17 rad (14.4–20 rad by workload), about 1 failure per 3 million inferences at 1 inference/s. SEE effect on training "requires further study" [36, pp. 5, 9].靜默資料損毀約每 17 rad 一次（依工作負載 14.4–20 rad），在每秒一次推論下約每 300 萬次推論一次失效。SEE 對訓練的影響「需要進一步研究」[36, pp. 5, 9]。 | Ground proton-beam test. HBM uncorrectable-ECC errors average one per 44 rad across 203 events. TPU system crashes, or SEFIs, average one per 5 krad per chip [36, p. 9]. Host CPU and RAM SEFIs average one per 450 rad and 400 rad, respectively [36, p. 6]. Both published versions give these rates.地面質子束。HBM 超出 ECC 更正能力的錯誤約每 44 rad 一次（203 次事件的平均），TPU 系統當機（SEFI）約每晶片每 5 krad 一次[36, p. 9]，主機的 CPU 與 RAM 則各自約每 450 與每 400 rad 發生一次 SEFI[36, p. 6]。兩個版本的速率相同。 |
| Launched-power price $14,700/kW/yr at $3,600/kg and $810/kW/yr at $200/kg for a 575 kg, ≈28 kW satellite over 5 years, against $570–3,000/kW/yr terrestrial power spend [36, pp. 6, 9–10, Table 1].一顆 575 kg、約 28 kW、五年壽命的衛星，在每公斤 3,600 美元時發射電力價格為每 kW 每年 14,700 美元，在每公斤 200 美元時為 810 美元，對照地面每 kW 每年 570–3,000 美元的電力支出[36, pp. 6, 9–10, Table 1]。 | The comparison covers power cost. A complete economic assessment needs additional cost inputs. The $200/kg estimate assumes a ≈20% learning rate. It requires about 1,800 further Starship launches by roughly 2035, averaging about 180 launches per year [36, pp. 6, 9].只比較電力成本，完整經濟可行性仍需其他成本輸入。每公斤 200 美元取決於約 20% 的學習率持續，需要在約 2035 年前再發射約 1,800 次 Starship，即每年約 180 次[36, pp. 6, 9]。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| How does the DWDM design reach 9.6 Tbps?DWDM 設計如何達到 9.6 Tbps？ | \( 24 \times 400\ \text{Gb/s} \) | 9.6 Tb/scomputed. Channel count and rate from Joule p. 8計算值。通道數與速率來自 Joule p. 8 |
| How long does light take to cross a 2 km cluster?光跨越 2 km 的叢集要多久？ | \( \dfrac{2\ \text{km}}{3\times10^{5}\ \text{km/s}} \) | ≈ 6.7 µscomputed. 2R with R = 1 km from Joule p. 4. Propagation only計算值。2R，R = 1 km 來自 Joule p. 4。僅傳播延遲 |
| How many chips' worth of ICI does one aperture carry?一個孔徑承載相當於幾顆晶片的 ICI？ | \( \dfrac{9.6\text{–}12.8\ \text{Tb/s}}{6.4\ \text{Tb/s}} \) | ≈ 1.5–2 chipscomputed. Aperture rate from Joule p. 8, v6e ICI from Google Cloud documentation (company)計算值。孔徑速率來自 Joule p. 8，v6e ICI 來自 Google Cloud 文件（公司資料） |
| How many silent corruptions does one chip see per year?一顆晶片每年遭遇幾次靜默損毀？ | \( \dfrac{150\ \text{rad/yr}}{17\ \text{rad/event}} \) | ≈ 9 per chip-yearcomputed. Both inputs from Joule p. 5計算值。兩個輸入都來自 Joule p. 5 |
| Does Table 1's $810/kW/yr follow from its inputs?Table 1 的每 kW 每年 810 美元應如何由其輸入推導？ | \( \dfrac{200\ \$/\text{kg} \times 575\ \text{kg}}{28\ \text{kW} \times 5\ \text{yr}} \) | ≈ $820/kW/yrcomputed. Matches the paper's $810 within rounding of the 28 kW estimate計算值。在 28 kW 估計值的捨入範圍內與論文的 810 美元相符 |
6. Limits and open questions6. 範圍與待解問題
- The cluster network remains to be specified. The paper specifies per-link rate, Fig. 2’s 8-neighborhood and possible connections between clusters over distances beyond 100 km. Optical-terminal count, switching architecture, wider topology, hop count, latency and bisection bandwidth remain unspecified. Chapter 11 and the two follow-up preprints examine these open choices. Google’s Palomar optical circuit switch serves as an analogy for parallel spatial streams [36, p. 3].叢集網路仍待指定。論文只給了每條鏈路的速率、Fig. 2 的 8 鄰居，以及叢集可再以相距超過 100 km 的鏈路組成「叢集的叢集」的說法，其餘架構輸入待指定：每顆衛星的光學終端數待指定、交換架構待指定、8 鄰居之外的拓樸待指定、跳數待指定、延遲預算待指定，對分頻寬因此也待指定，這正是第 11 章與兩篇後續預印本存在的原因。Google 的 Palomar 光路交換器只是作為平行空間光束的類比出現，其定位是類比[36, p. 3]。
- ICI capacity changes the comparison. The paper describes TPU inter-chip interconnect as hundreds of Gbps per chip [36, p. 3]. Google Cloud lists v6e capacity at 800 GB/s bidirectionally, or 6.4 Tbps per chip company. One aperture therefore carries roughly one or two chips’ ICI capacity. A satellite with dozens of TPU chips faces a fabric boundary similar to terrestrial designs that keep tensor parallelism within hosts.ICI 的比較被低估了。論文說 TPU 晶片間互連為「每晶片數百 Gbps」[36, p. 3]，而 Google Cloud 的 v6e 文件列出每晶片雙向 800 GB/s，即 6.4 Tbps company，所以一個孔徑大約只承載一到兩顆晶片的 ICI，而搭載「數十顆 TPU 晶片」的衛星會遇到與地面設計相同的層級邊界，後者的解法是把張量平行留在主機之內。
- Link distances use best-case constants. The 5 km, 1.25 km, and 0.32 km ranges exclude beam-quality loss, aperture truncation, and pointing jitter [36, p. 8]. Flight hardware must account for these losses. Figure A shows a large distance margin at 100–200 m spacing, while rate-versus-range models must retain those assumptions.最佳情況下的鏈路常數。5 km 的近場長度與 1.25 km、0.32 km 的多工距離都是忽略光束品質、孔徑截斷與指向抖動的共焦最佳情況[36, p. 8]，所以飛行鏈路能達到的會略少一些。如圖 A 所示，100–200 m 的間距留有很大餘裕，但任何建立在這些數字上的速率對距離模型，都會承襲它們的樂觀。
- ΔV and thermal evidence have distinct scopes. Joule budgets formation maintenance below 30 m/s per year. During solar storms, tilted drag-compensation thrust can supply much of the station-keeping ΔV [36, p. 5]. Propellant mass per satellite and collision-avoidance allowance need further specification. Supplemental notes sketch a control approach using backpropagation through ODEs. The HBM limit is near 90 °C, and radiator emission near 50 °C decreases about 1.2% per kelvin [36, pp. 2, 7]. Google’s September 2026 post reports thermal-vacuum validation of heat pipes and radiators, with orbital validation pending company.只有總量的 ΔV，也在軌熱資料待補充。Joule 內文把編隊維持的預算定為每年低於 30 m/s，並指出太陽風暴期間，大部分軌道維持 ΔV 可以靠調整補償阻力損失的推力方向來提供[36, p. 5]，但每顆衛星的推進劑質量待指定、避撞預留待指定，除了補充說明中概述的「穿過 ODE 反向傳播」方法之外飛行控制律亦待指定。熱的方面，論文把 HBM 的上限定在約 90 °C，並指出 50 °C 附近的輻射器每低 1 K，輻射量約減少 1.2%[36, pp. 2, 7]，而 Google 自己 2026 年 9 月的文章說熱管與輻射器已在熱真空中驗證，但「仍待在軌驗證」company。
- The prototype leaves network testing for the next mission. The specifications widely quoted for the MVP, four Trillium TPUs, about 1 kW of solar and chips powering down every 15–20 minutes, are press-reported (Payload. The New York Times via the IEEE ComSoc blog) and awaiting confirmation in Google’s posts. The first inter-satellite link test is the two-satellite Planet mission planned for 2027, which Planet and Google announced but the paper itself leaves unspecified.原型的網路測試留待後續任務。廣為引用的 MVP 規格，四顆 Trillium TPU、約 1 kW 太陽能、晶片每 15–20 分鐘斷電散熱，都是媒體報導（Payload。經 IEEE ComSoc 部落格轉述的紐約時報），Google 文章的確認待補充，而第一次星間鏈路測試是計畫在 2027 年進行的 Planet 雙衛星任務，這是 Planet 與 Google 的公告內容，論文本身列於論文範圍之外。
- Network assumptions require their dates and scope. Short-range Tbps links differ in operating range from Starlink’s claimed sustained ≈100 Gbps per laser company. The static-neighbor lattice leaves 2–30 s laser acquisition outside its model. Starlink’s 15-second access schedule would affect user traffic that reaches the formation through that communications constellation. Ground access remains an open design area, with radio for the pilot and optical links planned later. The paper cites NASA’s 200 Gbps TBIRD as the optical-downlink reference [36, pp. 2, 7].網路讀者應該標上日期的假設。短程 Tbps 鏈路與 Starlink 宣稱每具雷射約 100 Gbps 的持續速率 company 屬於相異等級。現代終端 2–30 秒的雷射擷取時間列於模型之外，因為晶格被當成一組固定的鄰居。Starlink 的 15 秒排程適用於編隊外的接入星座，卻會支配任何經通訊星座抵達叢集的使用者流量。而地面段仍是待解，試行用射頻，光學鏈路「最終」才需要、詳細設計仍待補充，引用的現況是 NASA 的 200 Gbps TBIRD[36, pp. 2, 7]。
7. How it connects7. 與其他內容的關聯
- Section 11.2: the HCW 2:1 ellipse, the 2:1.0037 J₂ correction and the 100–200 m oscillation, drawn and animated in Panel 11.1.第 11.2 節：HCW 2:1 橢圓、2:1.0037 的 J₂ 修正與 100–200 m 的振盪，在面板 11.1 中繪出並動畫呈現。
- Section 11.3: the near-field argument, DWDM and spatial multiplexing, with the paper's constants and a range-dependent rate panel.第 11.3 節：近場論證、DWDM 與空間多工，附論文的常數與隨距離變化的速率面板。
- Section 11.4: the table of what the paper specifies and what it leaves open. Section 11.5: cluster fabrics built on this geometry. Section 11.6: the redesign list.第 11.4 節：論文規定了什麼、留下什麼的對照表。第 11.5 節：建立在此幾何上的叢集網路架構。第 11.6 節：重新設計清單。
- Section 10.4: the architecture inputs required alongside per-aperture rate, in the torus-versus-Clos bisection comparison.第 10.4 節：在環面對 Clos 的對分頻寬比較中，說明為什麼每孔徑速率仍需配合架構輸入。
- Section 13.2 examines collectives on geometry that cycles twice per orbit. Section 13.3 develops fabric questions involving terminal counts, range, and setup delay.第 13.2 節（在每圈循環兩次的幾何上做集合通訊）與第 13.3 節（加入終端數、距離與建鏈延遲的網路架構）是建立在本文缺口上的研究方向。
- Pénot and Balakrishnan [102] repack this cluster from 81 to 367 satellites and map a Clos onto it. van Berkel [101] computes what a torus of such links loses to a Clos. SpaceMoE shows what inference assumes with links a hundred times slower. Xing et al. [85] is the only flight measurement of compute heat. TBIRD is the 200 Gbps downlink the paper cites.Pénot 與 Balakrishnan[102]把這個叢集從 81 顆重新排列成 367 顆，並在其上對映一個 Clos。van Berkel[101]計算由這類鏈路組成的環面相對 Clos 損失多少。SpaceMoE 展示在鏈路慢百倍時推論做了什麼假設。Xing 等人[85]是唯一的運算發熱飛行量測。TBIRD 是論文引用的 200 Gbps 下行鏈路。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 81 · R = 1 km · 650 km | Illustrative planar lattice in a dawn–dusk sun-synchronous orbit [36, pp. 3–4] | 晨昏太陽同步軌道上的示範平面晶格[36, pp. 3–4] | |
| 100–200 m · 2 cycles/orbit | Neighbour distance oscillation from the 2:1 relative ellipse [36, p. 4, Fig. 3] | 由 2:1 相對橢圓造成的相鄰距離振盪[36, p. 4, Fig. 3] | |
| 2:1.0037 · <3 m/s/yr/km | Axis ratio and J₂ drift residual. Formation maintenance under 30 m/s per year (Joule only) [36, p. 5] | 軸比與 J₂ 漂移殘差。編隊維持每年低於 30 m/s（僅 Joule）[36, p. 5] | |
| ≈10 Tbps · 9.6–12.8 Tbps | Per-aperture target. 24 × 400G DWDM on 100 or 75 GHz grids [36, pp. 3, 8] | 每孔徑目標。24 × 400G DWDM，100 或 75 GHz 柵格[36, pp. 3, 8] | |
| 800 Gbps / direction | Bench demonstration, off-the-shelf components, 1.6 Tbps bidirectional [36, p. 3] | 實驗台展示，現成元件，雙向 1.6 Tbps[36, p. 3] | |
| 5 W · 105.1 dB · 3 dB | EDFA transmit power, antenna gain per end, other losses [36, p. 8] | EDFA 發射功率、每端天線增益、其他損耗[36, p. 8] | |
| ≈ 5 km · 1.25 · 0.32 km | Near-field length πa²/λ. 2×2 and 4×4 multiplexing distances [36, pp. 3, 8] | 近場長度 πa²/λ。2×2 與 4×4 的多工距離[36, pp. 3, 8] | |
| 750 rad · 2 krad · 10 (15) krad | Five-year dose, HBM onset, highest dose tested with zero hard failures in Joule (arXiv v2) [36, p. 5] | 五年劑量、HBM 異常起點、Joule（arXiv v2）中保持運作時的最高測試劑量[36, p. 5] | |
| 1 / 17 rad | Silent data corruption per chip. ≈1 per 3 million inferences at 1/s [36, p. 5] | 每晶片的靜默資料損毀。每秒一次推論下約每 300 萬次一次[36, p. 5] | |
| $810/kW/yr vs $570–3,000 | Launched-power price at $200/kg vs US data-center power spend [36, p. 10, Table 1] | 每公斤 200 美元時的發射電力價格對美國資料中心電力支出[36, p. 10, Table 1] |