Chapter 11 of 13 · Part III · AI in orbit
第 11 章（共 13 章）· 第三部 · 在軌 AI
Google's Suncatcher formation: the links, the cluster network and what must be redesignedGoogle 的 Suncatcher 編隊：鏈路、叢集網路與必須重新設計之處
This chapter derives the 81-satellite cluster’s 2:1 relative orbit and explains the near-field link budget for a 10 cm telescope at 100 m. It reviews the Joule 2026 design’s specified and open items, compares lattice, hexagonal and Clos networks, and assesses terrestrial designs for orbital use.本章推導 81 顆衛星叢集的 2:1 相對軌道，並說明 10 cm 望遠鏡在 100 m 距離的近場鏈路預算。接著檢視 Joule 2026 設計的指定項目與待解問題，比較格狀、六角與 Clos 網路，並評估地面訓練網路設計在軌道上的適用範圍。
Before you start
開始之前
This chapter builds on Chapters 1, 2, 8 and 10. These terms recur throughout the chapter. Each link opens its definition.
New terms in this chapter (7)本章新名詞（7）
Each link opens the definition at the term’s first appearance.
各名詞在首次出現處都有定義，點選連結即可前往。
- Formation flying編隊飛行 Several satellites controlled so that their relative positions follow a prescribed, repeating geometry. 多顆衛星受控制，使其相對位置遵循預先設定、週期重複的幾何。
- Relative orbit相對軌道 The closed path a satellite traces in the rotating frame attached to a reference satellite. 衛星在附著於參考衛星的旋轉座標系中所描出的封閉路徑。
- HCW equationsHCW 方程 The Hill–Clohessy–Wiltshire linearized equations of relative motion about a circular reference orbit. Hill–Clohessy–Wiltshire 方程，描述繞圓形參考軌道的相對運動的線性化方程。
- J₂J₂ 項 The oblateness term of Earth's gravity field, the dominant perturbation that drifts relative orbits apart. 地球重力場的扁率項，是使相對軌道漂移分離的主要攝動。
- Spatial multiplexing空間多工 Splitting one aperture into an array of smaller sub-apertures, each carrying an independent beam, at the price of range. 將一個口徑分割為多個較小子口徑的陣列，每個子口徑承載獨立光束，代價是縮短距離。
- Cluster fabric叢集網路架構 The network inside a formation cluster: terminals per satellite, switching and the topology that links them. 編隊叢集內部的網路：每顆衛星的終端數、交換方式，以及連接它們的拓樸。
- Switch satellite交換衛星 A satellite whose role in a Clos mapping is switching dedicated to switching, so that it displaces a compute node. 在 Clos 對應中專責交換的衛星，因此會取代掉一個運算節點。
11.1The proposal and its status提案與現況
Recall
回顧
Section 1.3 shows that dawn–dusk sun-synchronous orbits keep solar beta angle above the eclipse threshold. Section 10.6 places formations between constellation and single-satellite regimes, at Tbps rates and microsecond delays.
Suncatcher combines a peer-reviewed design, a flown prototype and independent network analyses. Its orbit and close formation support sustained solar access and short optical paths.
Suncatcher 同時具備同儕審查設計、已飛行原型及獨立網路分析，其軌道與緊密編隊有助於持續取得日照並縮短光學路徑。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Project Suncatcher combines a peer-reviewed system paper, a flown prototype and two independent network analyses, which together provide the basis for this chapter. |
| Result and scope | 2025, 2026 | The paper is by Agüera y Arcas, Beals, Biggs, Bloom, Fischbacher, Gromov, Köster, Pravahan and Manyika. “Toward a future space-based, highly scalable AI infrastructure system design” appears in Joule [36]. This chapter takes its quantitative claims from that version. The November 2025 arXiv preprint [108] preprint is cited where its June 2026 v2 differs from Joule. |
| Capacity and rate | 81, 1 km, 650 km, 10 | The illustrative cluster has 81 TPU satellites on a square lattice within a 1 km radius. It flies at mean altitude 650 km in a dawn–dusk sun-synchronous orbit. Free-space optical links connect neighbors at roughly 10 Tbps per link [36, pp. 2–4]. |
| Orbital configuration | 1, 650 km, 35, 1,400 km | At 650 km, a dawn–dusk plane stays sunlit for most of the year. This avoids battery sizing for a 35-minute shadow every orbit. Some eclipse remains below about 1,400 km. Altitude also trades orbital crowding, insertion ΔV and inner Van Allen belt exposure. |
| Orbital configuration | 1,000 km | which starts near 1,000 km [36, p. 4] |
| Orbital configuration | Qualitative scope | while the sun-synchronous condition makes the plane rotate once per year with the Sun. |
| 結果與適用範圍 | 定性範圍 | Project Suncatcher 是本入門中唯一同時具備同儕審查系統論文、已飛行原型機，以及兩份針對其網路的獨立分析的在軌 AI 提案，因此獨立成章。 |
| 結果與適用範圍 | 2025, 11, 2026, 6 | 該論文是 Agüera y Arcas、Beals、Biggs、Bloom、Fischbacher、Gromov、Köster、Pravahan 與 Manyika 的「Toward a future space-based, highly scalable AI infrastructure system design」，刊於《Joule》[36]，本章所有數字都以此版本為準，而 2025 年 11 月首次張貼的 arXiv 預印本[108]preprint 只在其 2026 年 6 月的 v2 內文與 Joule 有差異時才引用。 |
| 容量與速率 | 81, 1 km, 650 km, 10 | 它描述的是一個示範性叢集：81 顆衛星排成方形格狀，叢集半徑 1 km，飛在平均高度 650 km 的晨昏太陽同步軌道上，每顆衛星搭載 TPU 加速器，並以每條鏈路 10 Tbps 量級的自由空間光鏈路連接鄰居[36, pp. 2–4]。 |
| 軌道配置 | 1, 650 km, 35, 1,400 km, 1,000 km | 這個軌道選擇直接承自第 1 章，因為 650 km 的晨昏軌道面一年中大部分時間都在日照下，可節省應付每圈 35 分鐘陰影所需的電池容量，論文同時指出，1,400 km 以下仍有部分地影，而高度的選擇還要在擁擠軌道、入軌所需的 ΔV 與約 1,000 km 起的內范艾倫帶之間取捨[36, p. 4]，而太陽同步條件又讓軌道面每年隨太陽轉一圈。 |
The paper quantifies formation dynamics, optical links, radiation tolerance and power cost. The inter-satellite network architecture remains a separate design task.
論文量化了編隊動力學、光學鏈路、輻射耐受與電力成本，星間網路架構則仍待另外設計。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The paper’s four quantitative arguments define its evidence scope, which press coverage sometimes extends. |
| Orbital configuration | 100, 200 m, 4, 3 | Next-nearest-neighbor distances vary between about 100 and 200 m. The derived formation repeats its shape twice per orbit (p. 4, Fig. 3). |
| Capacity and rate | 10, 800, 3, 8 | The paper argues that commercial DWDM transceiver pairs support roughly 10 Tbps per link at these distances. Its commercial-component bench demonstration reaches 800 Gbps in one direction (pp. 3, 8). |
| Reliability and accuracy | 2 krad, 750 rad, 17 rad, 5, 9, 8 | Trillium proton tests show HBM irregularities only after 2 krad(Si), almost three times the 750 rad(Si) five-year dose. Silent corruption occurs about once per 17 rad (pp. 5, 9). Chapter 8 uses these results. |
| Power and thermal conditions | 200, 2030s, 810, 570–3,000, 6, 9–10, 1 | At roughly $200/kg launch cost in the mid-2030s, the projected launched-power cost is $810/kW/yr for a Starlink-v2-mini-class satellite. US data-center power costs $570–3,000/kW/yr. The comparison covers power costs only (pp. 6, 9–10, Table 1). |
| Result and scope | Qualitative scope | The inter-satellite network design remains open, and Section 11.4 lists that gap item by item. |
| 結果與適用範圍 | 定性範圍 | 論文提出四個量化論證，明確界定了證據範圍，新聞報導有時會將其延伸。 |
| 結果與適用範圍 | 100, 200 m, 4, 3 | 它推導出一種編隊，其中次近鄰距離在約 100 到 200 m 之間振盪，叢集形狀每圈重複兩次（p. 4，Fig. 3）。 |
| 容量與速率 | 10, 800, 3, 8 | 它主張在這種距離下，一對商用 DWDM 收發器可以閉合每條鏈路 10 Tbps 量級的鏈路預算，並報告以現成元件在實驗台上達成單向 800 Gbps（pp. 3、8）。 |
| 模型與運算 | 2 krad, 750 rad, 17 rad, 5, 9, 8 | 它報告以質子照射 Trillium TPU，HBM 要到 2 krad(Si) 才出現異常，幾乎是 750 rad(Si) 五年任務劑量的三倍，靜默資料損毀約每 17 rad 一次（pp. 5、9），這些數字第 8 章已經用過。 |
| 電力與熱條件 | 2030, 200, 810, 570–3,000, 6, 9–10, 1 | 它還推算若 2030 年代中期發射成本降到約 $200/kg，Starlink v2 mini 等級衛星的每 kW 發射電力價格會是 $810/kW/yr，而美國資料中心電力支出為 $570–3,000/kW/yr，同時聲明這項比較僅涵蓋電力成本（pp. 6、9–10、Table 1）。 |
| 結果與適用範圍 | 定性範圍 | 衛星之間的網路設計仍待完成，第 11.4 節會逐項列出這個缺口。 |
The flown prototype tests launch, radiation and thermal behavior. Network testing belongs to a later paired-satellite mission, while detailed prototype specifications retain their press-report status.
已飛行原型測試發射、輻射與熱行為，網路測試則安排在後續雙衛星任務。原型細部規格仍維持媒體報導的證據等級。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 2026 | As of late October 2026, one prototype is in orbit, while the network claims await testing in space. |
| Result and scope | 1, 2026, 18 | Google’s 1 October 2026 blog confirms the Planet-built prototype launched that day on SpaceX Transporter-18 from Vandenberg company. Contact was established, and it was “operating as expected”. It collects launch-stress, radiation and thermal-extreme data. Link tests belong to a later mission. |
| Power and thermal conditions | 1 kW, 15–20 | Press reports describe the MVP as four Trillium TPUs, roughly one server’s compute, with about 1 kW solar power. Its chips reportedly shut down every 15–20 minutes to cool. Sources are Payload and the New York Times, relayed by Light Reading and the IEEE ComSoc blog press. Google’s posts leave these specifications open. |
| Result and scope | Qualitative scope | The three figures therefore retain their status as press reports. |
| Power and thermal conditions | 2026, 10, 50–100 | Google’s September 2026 “facts” post gives sustained launch loads of 10 g and component loads of 50–100 g company. Heat pipes and radiators passed thermal-vacuum validation, while orbital validation remains pending. Future satellites will carry “dozens of TPU chips”. |
| Planned optical test | 2027 | and that two satellites in 2027 will test the laser interconnect |
| Flight-test plan | 4, 2025, 2027 | Planet’s 4 November 2025 announcement matches this plan company. It will build two Owl-bus prototypes for launch by early 2027. They test “the ability of two spacecraft to fly in tandem with high-bandwidth cross-link communications”. |
| Result and scope | 2027 | The MVP uses ground connectivity only. The 2027 pair therefore supplies the first Suncatcher ISL measurement in space. James Manyika places useful operation beyond the next few years press, defining the planning timescale. |
| 結果與適用範圍 | 2026, 10 | 2026 年 10 月底的現況是：一顆原型機在軌，而網路方面的主張仍待太空測試。 |
| 結果與適用範圍 | 2026, 10, 1, 18 | Google 2026 年 10 月 1 日的部落格company確認與 Planet 合作打造的原型機當天搭乘 SpaceX Transporter-18 共乘任務自 Vandenberg 發射，已建立聯繫且「運作如預期」，並指出其目的是蒐集發射應力、輻射與極端溫度的資料，鏈路測試則留給後續任務。 |
| 電力與熱條件 | 1 kW, 15–20 | 被廣泛引用的這顆「MVP」衛星規格——四顆 Trillium TPU、約等於一台伺服器的算力、約 1 kW 太陽能、晶片每 15–20 分鐘必須關機散熱——來自 Payload，以及 Light Reading 與 IEEE ComSoc 部落格轉述的《紐約時報》press，Google 文章則待補充這些數字，因此本章將三項數值標為媒體報導。 |
| 模型與運算 | 2026, 9, 10, 50–100, 2027, 2025, 11, 4 | Google 自己 2026 年 9 月的「facts」文章company補充：發射負載達 10 g 持續、元件上達 50–100 g，熱管與散熱板已通過熱真空驗證，在軌驗證則待完成，後續衛星將搭載「數十顆 TPU 晶片」，而 2027 年將有兩顆衛星測試雷射互連，這與 Planet 2025 年 11 月 4 日的公告company一致：Planet 將以其 Owl 平台打造兩顆原型機，於 2027 年初前發射，測試「兩具太空船並列飛行並進行高頻寬交叉鏈路通訊的能力」。 |
| 結果與適用範圍 | 2027 | MVP 僅使用地面鏈路，因此 Suncatcher ISL 的第一次在軌量測屬於 2027 年那一對衛星，而 James Manyika 自己的說法——Google 把實用運作的時間點放在後續數年之外press——為後文的一切設定了時間尺度。 |
The independent analyses examine formation packing and network bisection. Both remain preprints, and training-step evaluation and detailed link critiques remain open.
獨立分析各自研究編隊堆積與網路對切頻寬，兩者仍屬預印本，訓練步驟評估與鏈路細部評論則待補充。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Two independent analyses frame the rest of the chapter. |
| Orbital configuration | 100 m, 1 km, 367, 81 | MIT’s Pénot and Balakrishnan treat the cluster as packing and switching problems. Circular relative orbits fit 367 satellites, compared with 81, at the same 100 m minimum spacing and 1 km envelope. Mapping Clos onto feasible line-of-sight links requires at least three switching layers [102, Figs. 4, 6, Tab. 4] preprint. |
| Latency comparison | 8,000 | Van Berkel compares torus and Clos bisection and latency at 8,000 nodes. Laser-wired tori lose two to three bisection orders relative to terrestrial Clos [101, Tab. 3] preprint. |
| Result and scope | Qualitative scope | Both analyses are preprints and leave training-step simulation outside their scope. |
| Review status | 2026 | As of October 2026, a published critique of the link budget and formation dynamics remains to be identified. |
| Result and scope | Qualitative scope | This chapter’s analysis therefore uses the paper’s equations and those two follow-ups. |
| 結果與適用範圍 | 定性範圍 | 兩份獨立分析構成本章其餘部分的框架。 |
| 軌道配置 | 100 m, 1 km, 367, 81 | MIT 的 Pénot 與 Balakrishnan 把叢集視為堆積與交換問題，證明在同樣的 100 m 最小間距與 1 km 包絡下，若相對軌道改為圓形可容納 367 顆衛星，正方配置則為 81 顆，而 Clos 架構只有在三層以上交換層時才能對應到可行的視線鏈路上[102, Figs. 4, 6, Tab. 4]preprint。 |
| 結果與適用範圍 | 8,000 | van Berkel 計算 8,000 節點在軌叢集的環面與 Clos 架構的對分頻寬與延遲，結論是以雷射鏈路佈線的環面比地面 Clos 少了二到三個數量級的對分頻寬[101, Tab. 3]preprint。 |
| 結果與適用範圍 | 定性範圍 | 兩份分析都屬預印本，訓練步驟模擬則在研究範圍之外。 |
| 結果與適用範圍 | 2026, 10 | 截至 2026 年 10 月，鏈路預算與編隊動力學的論文評論仍待補充，因此本章以原文方程式與這兩篇後續研究展開分析。 |
11.2Flying in formation: HCW, the 2:1 ellipse and J₂編隊飛行：HCW、2:1 橢圓與 J₂
Recall
回顧
Section 1.1 derives a 97.73 min period at 650 km from Kepler’s relation. Section 1.3 introduces J₂-driven nodal precession.
Free orbital motion permits a cluster with separations of roughly a hundred meters. Formation flying controls satellites so their relative positions follow a prescribed, repeating geometry. Each path in a reference satellite’s frame is its relative orbit. Suncatcher starts with Hill–Clohessy–Wiltshire (HCW) equations about a circular reference orbit [36, p. 7]. These linearized relative-motion equations originated in rendezvous guidance (Clohessy & Wiltshire, J. Aerospace Sci. 27(9), 1960).
The coordinates are \(x\) radially, \(y\) along track and \(z\) across track. Mean reference-orbit motion is \(n = 2\pi/T\). Burnett and Schaub give the unforced equations as follows [103] preprint.
一群相距一百公尺的衛星之所以可能，是因為軌道力學的一項性質，可由自由軌道運動形成。編隊飛行是控制多顆衛星使其相對位置遵循預先設定、週期重複之幾何的做法，而每顆衛星在參考衛星座標系中遵循的幾何稱為它的相對軌道。Suncatcher 論文自稱以之為解析起點的[36, p. 7]，是 Hill–Clohessy–Wiltshire（HCW）方程，即繞圓形參考軌道之相對運動的線性化方程，最初為會合導引而發表（Clohessy & Wiltshire，J. Aerospace Sci. 27(9)，1960）。以 \(x\) 表徑向偏移、\(y\) 表沿軌偏移、\(z\) 表橫軌偏移、\(n = 2\pi/T\) 表參考軌道的平均運動，依 Burnett 與 Schaub[103]preprint 的寫法，自由運動的的方程為
\[ \ddot x - 2n\dot y - 3n^2 x = 0,\qquad \ddot y + 2n\dot x = 0,\qquad \ddot z + n^2 z = 0 . \]
The first two equations couple radial and along-track motion through the Coriolis terms \(2n\dot y\) and \(2n\dot x\). In comparison, the cross-track motion is an independent oscillation at the orbital frequency. The along-track equation integrates once to \(\dot y + 2nx = c\). The constant \(c\) is the drift rate. A radial offset with zero compensating along-track velocity changes the satellite’s orbital period relative to the reference.
Separation therefore grows linearly with time. Bounded motion requires \(\dot y_0 = -2n x_0\). The in-plane solution is then \(x(t) = A_0\cos(nt+\alpha)\) and \(y(t) = -2A_0\sin(nt+\alpha) + y_{\rm off}\) [103] preprint. This closed ellipse has along-track semi-axis \(2A_0\), twice radial semi-axis \(A_0\). Suncatcher describes motion “±R prograde, ±R/2 in altitude” relative to S0. Apoapsis occurs at three twelfths and periapsis at nine twelfths of an orbit [36, pp. 4–5].
The equations supply the factor two.
Pénot and Balakrishnan therefore describe the same orbits as ellipses of eccentricity \(\sqrt{3}/2\), since \(b/a = \sqrt{1 - 3/4} = 1/2\) [102] preprint.
前兩式透過柯氏項 \(2n\dot y\) 與 \(2n\dot x\) 耦合徑向與沿軌運動，橫軌運動則是以軌道頻率振盪的獨立項。沿軌方程積分一次得 \(\dot y + 2nx = c\)，常數 \(c\) 就是漂移率：具有徑向偏移、補償性沿軌速度為零的衛星，其軌道週期與參考軌道略有差異，會隨時間線性分離。因此有界運動需要初始條件 \(\dot y_0 = -2n x_0\)，在此條件下面內解為 \(x(t) = A_0\cos(nt+\alpha)\) 與 \(y(t) = -2A_0\sin(nt+\alpha) + y_{\rm off}\)[103]preprint，是一個封閉橢圓，其沿軌半軸 \(2A_0\) 恰為徑向半軸 \(A_0\) 的兩倍。這正是 Suncatcher 論文所述的 2:1 橢圓：每顆外圍衛星相對中央衛星 S0 做「順行方向 ±R、高度方向 ±R/2」的運動，在十二分之三圈時到達遠地點、十二分之九圈時到達近地點[36, pp. 4–5]。這個因子 2 來自方程本身，所以 Pénot 與 Balakrishnan 能把同樣的軌道描述為離心率 \(\sqrt{3}/2\) 的橢圓，因為 \(b/a = \sqrt{1 - 3/4} = 1/2\)[102]preprint。
The HCW solution rotates and stretches the square phase pattern. This gives the formation a repeated shape cycle and makes direct and diagonal link distances vary predictably.
HCW 解使方形相位圖樣旋轉並拉伸，形成重複的形狀週期，直接與對角鏈路的距離也隨之規律改變。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 100–200 m | The lattice itself follows from a small piece of linear algebra that explains its 100–200 m oscillation beyond the paper’s stated result. |
| Result and scope | 0 | Write each satellite's ellipse by the vector \((p,q) = (A_0\cos\alpha,\ A_0\sin\alpha)\). |
| Result and scope | -2 | The HCW solution maps \((p,q)\) to \((x,y) = (p\cos nt - q\sin nt,\ -2(p\sin nt + q\cos nt))\). It rotates by \(nt\) and stretches the along-track direction by a factor of two. |
| Result and scope | 2s, 0, 4, 2 | If the \((p,q)\) pattern is a square lattice of spacing \(s\), the physical cluster at any instant is that square rotated and stretched. Thus, it is a rectangle with spacings \(s\) and \(2s\) at \(nt = 0\), a rhombus at \(nt = \pi/4\). A rectangle again at \(nt = \pi/2\) with the roles of the axes exchanged. |
| Orbital configuration | 0, 1, 3, 2, 100, 100 m, 200 m | A pattern vector \((s,0)\) has length \(s\sqrt{1 + 3\sin^2 nt}\). At \(s = 100\) m, direct neighbors reach 100 m twice per orbit and 200 m twice per orbit. Because \(\sin^2\) has period \(\pi\), the shape repeats twice per orbit. These results match the paper’s next-nearest-neighbor distances and shape cycle [36, p. 4, Fig. 3]. |
| Result and scope | 5, 3, 2, 141, 283 m | The diagonal vector \((s,s)\) has length \(s\sqrt{5 + 3\sin 2nt}\), ranging from 141 to 283 m. At rectangular phases, along-track spacing is 2\(s\). Pénot and Balakrishnan call this “2\(R_{\min}\) spacing needed along \(y\)”, the square lattice’s first packing inefficiency preprint. |
| Result and scope | 11.1 | The calc table below collects these results, and Panel 11.1 lets you watch them. |
| 結果與適用範圍 | 100–200 m | 以下以線性代數推導格子，說明論文所述的 100–200 m 間距振盪。 |
| 結果與適用範圍 | 0 | 以向量 \((p,q) = (A_0\cos\alpha,\ A_0\sin\alpha)\) 表示每顆衛星的橢圓。 |
| 結果與適用範圍 | -2 | HCW 解把 \((p,q)\) 映射到位置 \((x,y) = (p\cos nt - q\sin nt,\ -2(p\sin nt + q\cos nt))\)，也就是先旋轉角度 \(nt\)、再沿軌拉伸兩倍。 |
| 結果與適用範圍 | 0, 2s, 4, 2 | 若 \((p,q)\) 圖樣是間距 \(s\) 的正方格，則任一瞬間的實體叢集就是該正方格的旋轉與拉伸，因此在 \(nt = 0\) 時是間距 \(s\) 與 \(2s\) 的矩形、在 \(nt = \pi/4\) 時是菱形、在 \(nt = \pi/2\) 時又是矩形但兩軸角色互換。 |
| 結果與適用範圍 | 0, 1, 3, 2, 100, 100 m, 200 m | 格向量 \((s,0)\) 的長度為 \(s\sqrt{1 + 3\sin^2 nt}\)，所以圖樣中相距 \(s = 100\) m 的直接鄰居每圈有兩次相距 100 m、兩次相距 200 m，而由於 \(\sin^2\) 的週期為 \(\pi\)，叢集形狀每圈重複兩次，兩者都與論文對次近鄰距離與形狀週期的敘述完全一致[36, p. 4, Fig. 3]。 |
| 結果與適用範圍 | 5, 3, 2, 141, 283 m | 對角向量 \((s,s)\) 的長度為 \(s\sqrt{5 + 3\sin 2nt}\)，在 141 到 283 m 之間，而矩形相位時沿軌 2\(s\) 的間距，正是 Pénot 與 Balakrishnan 指出方形格設計第一個堆積低效之處的「沿 \(y\) 需要的 2\(R_{\min}\) 間距」preprint。 |
| 結果與適用範圍 | 11.1 | 下面的計算表彙整這些結果，面板 11.1 則讓你實際觀看。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Mean motion at 650 km650 km 的平均運動 | \(n = 2\pi/T,\ T = 97.73\ \text{min}\) | 1.07×10⁻³ rad/scomputed. T from the Chapter 1 table [5]計算值。T 取自第 1 章表格[5] |
| Bounded-motion condition for the outer satellite, \(x_0 = 500\) m最外圈衛星的有界運動條件，\(x_0 = 500\) m | \(\dot y_0 = -2n x_0\) | −1.07 m/scomputed from the HCW drift constant由 HCW 漂移常數計算 |
| Axis ratio of the relative orbit相對軌道的軸比 | \(2A_0 : A_0\) | 2 : 1HCW solution. Paper: ±R along track, ±R/2 radial, tuned to 2 : 1.0037 for J₂HCW 解。論文：沿軌 ±R、徑向 ±R/2，為 J₂ 調為 2 : 1.0037 |
| Direct-neighbor separation, lattice spacing \(s = 100\) m直接鄰居間距，格距 \(s = 100\) m | \(d = s\sqrt{1+3\sin^2 nt}\) | 100–200 mcomputed from the HCW map. Paper p. 4 states 100–200 m由 HCW 映射計算。論文 p. 4 述為 100–200 m |
| Shape cycles per orbit每圈形狀週期數 | \(\sin^2 nt\) has period之週期為 \(\pi\) | 2computed. Paper p. 4 states two cycles per orbit計算值。論文 p. 4 述為每圈兩次 |
| Diagonal-neighbor separation對角鄰居間距 | \(d = s\sqrt{5+3\sin 2nt}\) | 141–283 mcomputed from the HCW map由 HCW 映射計算 |
| Neighbours within 200 m on an undeformed 100 m square lattice初始形狀的 100 m 正方格內 200 m 以內的鄰居數 | \(4\times s + 4\times s\sqrt2\ (+4\times 2s)\) | 8 (12 at ≤ 200 m)computed. The paper's Fig. 2 "8-neighborhood"計算值。論文 Fig. 2 的「8 鄰域」 |
| Relative speed of the outer satellite, \(A_0 = 500\) m最外圈衛星的相對速度，\(A_0 = 500\) m | \(v = 2nA_0\) … \(nA_0\) | 1.07 … 0.54 m/scomputed from \(\dot x,\dot y\) of the ellipse由橢圓的 \(\dot x,\dot y\) 計算 |
| Relative speed between direct neighbors直接鄰居之間的相對速度 | \(sn\sqrt{\sin^2 nt+4\cos^2 nt}\) | 0.11–0.21 m/scomputed. S = 100 m計算值。s = 100 m |
| Doppler shift at 0.21 m/s, 1,550 nm0.21 m/s、1,550 nm 的多普勒頻移 | \(0.645\ \text{GHz/(km/s)} \times 2.1\times10^{-4}\) | ≈ 0.14 MHzcomputed. Coefficient from Section 2.4計算值。係數取自第 2.4 節 |
| J₂ residual drift for R = 1 kmR = 1 km 的 J₂ 殘餘漂移 | \(< 3\ \text{m/s/yr/km} \times 1\ \text{km}\) | < 3 m/s per yearpaper p. 5 bound. Joule budgets formation maintenance at under 30 m/s per year (p. 5)論文 p. 5 的上界。Joule 把編隊維持預算定為每年低於 30 m/s（p. 5） |
Neighbor relative speeds are fractions of a meter per second. Adjacent-plane Starlink links instead have the 0.27 km/s range rate derived in Section 2.4. Formation Doppler is about 0.14 MHz, three orders below coherent-receiver tolerance, and the point-ahead angle vanishes. The Section 2.4 pointing problem becomes holding a microradian beam against thermal body flexing, with small tracking demands. Every direct link doubles in length and contracts twice per orbit. Section 11.3 shows how spatial multiplexing converts that distance cycle into rate variation.
這些數字有兩個對網路重要的後果。相對速度極小，鄰居之間低於每秒一公尺，相較於第 2.4 節導出的相鄰軌道面 Starlink 鏈路 0.27 km/s 距離變率，約 0.14 MHz 的多普勒頻移比相干接收機能容忍的低三個數量級，前置指向角也可忽略。第 2.4 節的指向問題變成在太空船本體因熱而撓曲時把光束維持在一微弧度內，追蹤目標運動的負擔則較小。距離則隨相位改變，每條直接鏈路每圈兩次拉長兩倍再縮回，第 11.3 節會說明這透過空間多工直接影響鏈路速率。
Earth’s oblateness perturbs the ideal relative orbits. Suncatcher tunes their shape and budgets control effort, while propellant, collision avoidance and the flight control law remain open.
地球扁率會擾動理想相對軌道。Suncatcher 調整其形狀並編列控制預算，推進劑、避撞與飛行控制律則仍待指定。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 1 | J₂, Earth’s gravity-field oblateness term, is the main perturbation beyond Keplerian motion. Chapter 1 introduces its nodal precession. Here, it acts differently on satellites with slightly different radii and inclinations. HCW ellipses closed in the two-body model therefore drift open over weeks. |
| Result and scope | 79, 2001 | The classical remedy matches mean-element drift rates across relative orbits. Schaub and Alfriend describe these J₂-invariant formations (Celestial Mechanics and Dynamical Astronomy 79, 2001). |
| Orbital configuration | 2:1, 2:1.0037, 3 m | Suncatcher uses a numerical analogue, tuning axis ratio from 2:1 to 2:1.0037. Reported J₂ drift stays below 3 m/s yearly per kilometer of maximum distance from the reference orbit. The paper attributes this estimate to a simplified numerical calculation. |
| Result and scope | Qualitative scope | while its methods integrate the perturbed equations with an eighth-order Runge–Kutta (DOP853) to about nine digits [36, pp. 5, 7–8]. |
| Result and scope | 20%, 30 m | Joule adds a maintenance budget beyond arXiv v2. Numerical modeling includes plausible position and velocity errors, unmodeled accelerations, differential drag, thruster minimum impulse and thrust jitter. It considers calm and stormy space weather. Under 30 m/s yearly ΔV keeps each neighbor distance within 20% of its intended value. |
| Power and thermal conditions | Qualitative scope | Solar storms sharply increase daily ΔV. Much station-keeping can then use a small tilt of the thrust that replaces drag-lost energy [36, p. 5]. |
| Orbital configuration | 11.1, 11.4 | An aggregate maintenance budget therefore exists, while Panel 11.1 excludes it. Per-satellite propellant mass, collision-avoidance allowance and flight-control law remain open in Section 11.4. The supplement sketches control through backpropagation across orbit integration in a framework such as JAX. |
| 結果與適用範圍 | 1 | 真實軌道包含克卜勒運動以外的攝動，最重要的修正是 J₂，即地球重力場的扁率項，第 1 章已見它造成節點進動，此處它對半徑與傾角略有差異的衛星產生差異作用，使得在二體模型中完美閉合的 HCW 橢圓在數週內漂開。 |
| 結果與適用範圍 | 79, 2001, 2:1, 2:1.0037, 3 m | 經典解法是選擇平均根數漂移率相匹配的相對軌道，即 Schaub 與 Alfriend 的 J₂ 恆定編隊（Celestial Mechanics and Dynamical Astronomy 79，2001），而 Suncatcher 論文採用數值上的類似做法：把軸比從 2:1 微調為 2:1.0037，並報告這能把 J₂ 引起的漂移壓在每公里最大距離每年 3 m/s 以下，論文稱這個數字來自簡化的數值計算，而其方法一節以八階 Runge–Kutta（DOP853）積分受攝方程至約九位有效數字[36, pp. 5, 7–8]。 |
| 結果與適用範圍 | 30 m, 20% | Joule 內文接著補上 相對於 arXiv v2 新增的預算：在位置與速度的合理誤差範圍、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等條件下，並各自考慮平靜與風暴的太空天氣，詳細的數值模擬顯示，每年低於 30 m/s 的 ΔV 就能讓每個鄰居距離維持在預定值的 20% 以內。 |
| 結果與適用範圍 | 定性範圍 | 而在太陽風暴期間每日 ΔV 急遽升高時，大部分位置保持可以併入補償阻力能量損失的推力，只需略微調整其方向[36, p. 5]。 |
| 結果與適用範圍 | 11.1 | 因此編隊有一個總量的維持預算（列於面板 11.1 範圍之外），每顆衛星的推進劑質量、避撞預留與飛行控制律則待指定，後者在補充說明中只被概述為在 JAX 之類的框架中穿過軌道積分做反向傳播，所以第 11.4 節把這三項列為待確定。 |
Original figure note原始圖說出處
Drawn for this primer from the HCW solution (Burnett & Schaub, arXiv 2108.09608) and Agüera y Arcas et al., Joule 2026, pp. 4–5, Figs. 2–3.Stop and think停下來想一想
Move the panel to the rhombic phase and read both diagonal separations. One neighbor pair is closest while the other is farthest. Assume four terminals pointing to the four nearest neighbors. Determine how often those neighbor identities change per orbit. Explain the effect of Section 2.4’s 2–30 s acquisition after each repointing.
把面板拉到菱形相位，讀出兩個對角間距。一對對角鄰居最近時，另一對正好最遠。若一顆衛星只有四個雷射終端並指向四個最近的鄰居，這四個鄰居的身分每圈會換幾次，而第 2.4 節的 2–30 秒建鏈時間，對每次都必須重新指向的鏈路意味著什麼？
11.3Links at 100 m: near field, DWDM and spatial multiplexing100 m 的鏈路：近場、DWDM 與空間多工
Recall
回顧
Section 2.2 gives the far-field budget \(P_R = P_T G_T G_R (\lambda/4\pi d)^2 L\) with \(G \approx (\pi D/\lambda)^2\). Its 1 W, 10 cm, 2,000 km example yields 1.58 µW. The near field falls outside this formula’s valid range.
第 2.2 節寫出遠場預算 \(P_R = P_T G_T G_R (\lambda/4\pi d)^2 L\)，其中 \(G \approx (\pi D/\lambda)^2\)，以 1 W、10 cm、2,000 km 的例子算得 1.58 µW，並介紹近場為此公式適用範圍之外的距離範圍。
Short optical paths provide the received-power margin for parallel DWDM channels. The reported bench result supports part of the target, with longer-range and flight validation still open.
短光學路徑提供了平行 DWDM 通道所需的接收功率餘裕。實驗台結果支持部分目標，較長距離與在軌驗證仍待完成。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The networking idea in Suncatcher is a single inversion of the usual inter-satellite link problem. |
| Result and scope | 1, 2, 1–100, 5,000 km, 10 cm, 95 m, 1.6 | Long-range links in Section 2.3 experience \(1/d^2\) divergence loss over thousands of kilometers. The paper lists existing optical ISLs at 1–100 Gbps. At 5,000 km, a 10 cm aperture produces at least a 95 m spot radius. Received power is at most about 1.6 µW [36, pp. 3, 8]. |
| Orbital configuration | 5,000 km, 10, 8 | Shrinking the range from 5,000 km to a few hundred meters cuts the divergence loss by a factor of order \(10^8\) |
| Power and thermal conditions | 24, 400, 16, 100 | The authors use this power margin for 24 simultaneous DWDM channels. Commercial 400G PM-16QAM coherent transceivers operate on the 100 GHz ITU grid. |
| Capacity and rate | 9.6 | which is half the C-band, give 9.6 Tbps per aperture in the paper's accounting |
| Capacity and rate | 75, 12.8, 8 | and a 75 GHz grid would give about 12.8 Tbps (p. 8). |
| Power and thermal conditions | 5 W, 105.1, 10 cm, 80%, 2, 106.1, 3, 8 | The budget uses a 5 W erbium-doped fiber amplifier, 105.1 dB antenna gains at each end and 3 dB other losses. A 10 cm aperture’s ideal \((\pi D/\lambda)^2\) gain is 106.1 dB, so the used gain implies about 80% efficiency. Coherent detection rejects adjacent-beam crosstalk (pp. 3, 8). |
| Capacity and rate | 10, 800, 1.6, 3 | The bench result supporting the 10 Tbps target is 800 Gbps per direction, or 1.6 Tbps bidirectional. It uses commercial components on a short free-space path (p. 3). |
| Result and scope | 100 m | A peer-reviewed free-space demonstration at 100 m to a few kilometers remains pending. |
| Model and compute | Qualitative scope | The transceiver model and bench range are unspecified. |
| Capacity and rate | 1.6 | The 1.6 Tbps result therefore describes a company bench test reported in a peer-reviewed paper. |
| 結果與適用範圍 | 定性範圍 | Suncatcher 的網路構想是把一般星間鏈路問題做一次反轉。 |
| 容量與速率 | 1, 2, 1–100, 10 cm, 5,000 km, 95 m, 1.6 | 第 2.3 節的每條長程雷射鏈路都要在數千公里上對抗 \(1/d^2\) 的發散損耗，所以論文指出現有光學星間鏈路僅 1–100 Gbps，而 10 cm 口徑在 5,000 km 的鏈路會產生至少 95 m 半徑的光斑、最多約 1.6 µW 的接收功率[36, pp. 3, 8]。 |
| 容量與速率 | 5,000 km, 10, 8, 400, 16, 100, 24, 9.6, 75, 12.8 | 把距離從 5,000 km 縮到幾百公尺，發散損耗減少約 \(10^8\) 倍，而作者利用功率餘裕同時承載許多通道：商用 400G PM-16QAM 相干收發器在 100 GHz ITU 網格（即 C 頻帶的一半）上的 24 個 DWDM 通道，在論文的計算方式下每個口徑可達 9.6 Tbps，75 GHz 網格則約 12.8 Tbps（p. 8）。 |
| 電力與熱條件 | 5 W, 105.1, 10 cm, 80%, 2, 106.1, 3, 8 | 背後的鏈路預算以 5 W 摻鉺光纖放大器為發射機、兩端各 105.1 dB 的天線增益（對 10 cm 口徑而言意味著約 80% 的口徑效率，因為理想的 \((\pi D/\lambda)^2\) 為 106.1 dB）與 3 dB 其他損耗，並依靠相干檢測來排除相鄰光束的串擾（pp. 3、8）。 |
| 容量與速率 | 10, 800, 1.6, 3 | 支撐 10 Tbps 目標的實驗台示範是每方向 800 Gbps、雙向 1.6 Tbps，以現成元件在短自由空間路徑上完成（p. 3）。 |
| 結果與適用範圍 | 100 m | 100 m 到數公里的同儕審查自由空間示範仍待完成，收發器型號與實驗台距離也待補充。 |
| 容量與速率 | 1.6 | 因此 1.6 Tbps 屬於同儕審查論文所報告的公司實驗台結果。 |
Near-field propagation keeps most aperture-contained power on a receiver of comparable size. The stated boundary and sensitivity figures depend on the paper’s best-case optical assumptions.
近場傳播讓口徑內的大部分功率落在尺寸相近的接收器上，所列邊界與靈敏度則取決於論文的最佳光學假設。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Near-field power | Qualitative scope | Near-field propagation explains why power is "abundant" when \(d\) is small. The point deserves care because it decides which distances the multiplexing scheme can use. |
| Power and thermal conditions | 2 | A collimated beam of radius \(a\) begins significant divergence beyond a characteristic distance of order \(a^2/\lambda\). Within this near field, described in Section 2.2, essentially all aperture-transmitted power reaches an equal-sized receiver. Applying the far-field formula here can predict more received than transmitted power, revealing its invalid scope. |
| Result and scope | Qualitative scope | The paper states that boundary for a symmetric confocal system |
| Orbital configuration | 2, 5 km, 5 cm, 1.55 µm | The confocal beam waist lies midway between transceivers, giving \(L = \pi a^2/\lambda\). A 5 cm radius at 1.55 µm gives about 5 km. This best case excludes beam quality, aperture truncation and pointing jitter. |
| Power and thermal conditions | −20, 0.24 mW, 24, 71, 196, 1.39, 16 | Both arXiv v2 and Joule use roughly −20 dBm required power per channel, or 0.24 mW across 24 channels. Sensitivities are about 71 photons/bit for OOK, 196 for PM-16QAM and 1.39 at the shot-noise Shannon limit. |
| Result and scope | 1.0, 5 km | and a pointing accuracy of about 1.0 µrad to keep beam wander within a tenth of the aperture radius at 5 km [36, p. 8]. |
| Power and thermal conditions | 100–200 m | Formation spacing of 100–200 m lies about twenty-five to fifty times inside the near-field distance. This geometry provides the large received-power margin. |
| 結果與適用範圍 | 100 m | 近場傳播解釋了為何 100 m 處功率「充裕」，此時 \(d\) 很小，這一點需要小心處理，因為它決定了多工方案能使用哪些距離。 |
| 結果與適用範圍 | 2 | 半徑 \(a\) 的準直光束只有超過約 \(a^2/\lambda\) 量級的特徵距離後才開始發散，而在該距離以內——即第 2.2 節的近場——幾乎所有口徑內的發射功率都會落在同尺寸的接收器上，所以遠場公式失去意義，若硬套用會預測出比發射功率更多的接收功率。 |
| 軌道配置 | 2, 1.55 µm, 5 cm, 5 km | 論文以對稱共焦系統（光束腰位於兩具收發器正中央）給出這個邊界：\(L = \pi a^2/\lambda\)，對 1.55 µm、光束半徑 5 cm 約為 5 km，這是忽略光束品質、口徑截斷與指向抖動的最佳情況。 |
| 軌道配置 | −20, 24, 0.24 mW, 16, 71, 196, 1.39, 5 km, 1.0 | arXiv v2 與 Joule 內文給出的相關常數也完全相同：每通道所需接收功率約 −20 dBm（24 通道共 0.24 mW）、OOK、PM-16QAM 與散粒雜訊下 Shannon 極限的靈敏度約為每位元 71、196 與 1.39 個光子，以及要在 5 km 處把光束漂移維持在口徑半徑十分之一以內所需約 1.0 µrad 的指向精度[36, p. 8]。 |
| 結果與適用範圍 | 100–200 m | 因此編隊的 100–200 m 位於近場之內二十五到五十倍處，這正是「功率充裕」這個說法的全部依據。 |
Spatial multiplexing trades reach for parallel beams by dividing an aperture into smaller systems. The available rate therefore changes with link distance and imaging resolution.
空間多工將口徑分割為較小系統，以傳輸距離換取平行光束，因此可用速率隨鏈路距離與成像解析度改變。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Power and thermal conditions | 10 cm, 2×2, 5 cm, 4×4, 2.5 cm | spatial multiplexing splits one aperture into smaller sub-apertures, each carrying an independent beam. A 10 cm telescope becomes a 2×2 array of 5 cm systems or a 4×4 array of 2.5 cm systems. Each sub-beam carries its own DWDM stream [36, p. 3, Fig. 1]. |
| Cost assumptions | Qualitative scope | The price is range. |
| Result and scope | 2 | A smaller sub-aperture has a shorter near-field length, since that length scales with \(a^2\) |
| Result and scope | Qualitative scope | so the array works only out to a distance that shrinks as the square of the array order |
| Orbital configuration | 2×2, 1.25 km, 4×4, 0.32 km | The paper’s 2×2 array works to 1.25 km, while its 4×4 array works to 0.32 km. Their range ratio is about four. |
| Capacity and rate | 3, 8 | and why the text concludes that the total bandwidth available through spatial multiplexing scales inversely with distance (pp. 3, 8). |
| Power and thermal conditions | 1, 24, 10 cm, 300 km | The Figure 1 caption adds that these distances are limited by the imaging resolution of each sub-aperture with received power still sufficient. That 24-way DWDM through a single undivided 10 cm aperture reaches about 300 km (Fig. 1 caption). |
| Result and scope | Qualitative scope | In that regime the coherent receiver's inherent rejection of crosstalk from adjacent beams earns its place. |
| Result and scope | 11.2, 2 | Panel 11.2 uses the primer’s simple crosstalk rule: far-field spot diameter must stay below half the sub-aperture pitch. This is about twice as conservative as the paper’s ranges, which match each sub-aperture’s near-field length \(\pi a^2/\lambda\). |
| 結果與適用範圍 | 10 cm, 2×2, 5 cm, 4×4, 2.5 cm | 接收功率充足後，論文把它花在空間多工上，也就是把一個口徑分割為多個較小子口徑的陣列、每個子口徑承載獨立光束，使一具 10 cm 望遠鏡變成 2×2 的 5 cm 系統陣列或 4×4 的 2.5 cm 系統陣列，每束子光束各自承載一道獨立的 DWDM 資料流[36, p. 3, Fig. 1]。 |
| 結果與適用範圍 | 定性範圍 | 代價是距離。 |
| 軌道配置 | 2, 2×2, 1.25 km, 4×4, 0.32 km, 3, 8 | 較小的子口徑近場長度較短，因為該長度與 \(a^2\) 成正比，所以陣列只能在隨陣列階數平方縮小的距離內運作，這正是論文的 2×2 陣列可用到 1.25 km、4×4 陣列只能用到 0.32 km（約相差四倍）的原因，也是正文結論說空間多工可得的總頻寬與距離成反比的原因（pp. 3、8）。 |
| 軌道配置 | 1, 10 cm, 24, 300 km | Fig. 1 的圖說另外指出，這些距離受限於各子口徑的成像解析度，接收功率仍充足，而完整的單一 10 cm 口徑以 24 路 DWDM 可達約 300 km（Fig. 1 圖說）。 |
| 結果與適用範圍 | 定性範圍 | 在這個區間，相干接收機對相鄰光束串擾的固有抑制正是價值所在。 |
| 結果與適用範圍 | 11.2, 2 | 面板 11.2 採用一條刻意簡化的串擾規則——光束的遠場光斑直徑必須小於子口徑間距的一半——並標示為本入門自己的規則，因為它比論文的數字保守約兩倍，而論文的多工距離恰與各子口徑的近場長度 \(\pi a^2/\lambda\) 一致。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Aggregate rate per aperture, 100 GHz grid每口徑總速率，100 GHz 網格 | \(24 \times 400\ \text{Gb/s}\) | 9.6 Tb/spaper p. 8 arithmetic. "bidirectional" in its accounting論文 p. 8 的算式。其計算方式稱為「雙向」 |
| Aggregate rate, 75 GHz grid總速率，75 GHz 網格 | \(32 \times 400\ \text{Gb/s}\) | 12.8 Tb/scomputed. Matches paper p. 8 "≈ 12.8 Tbps"計算值。與論文 p. 8「≈ 12.8 Tbps」一致 |
| Near-field length, a = 5 cm, λ = 1,550 nm近場長度，a = 5 cm，λ = 1,550 nm | \(\pi a^2/\lambda\) | 5.1 kmcomputed. Paper p. 8 gives ≈ 5 km for the confocal best case計算值。論文 p. 8 給出共焦最佳情況約 5 km |
| Far-field received power at 5,000 km (5 W, 2×105.1 dB, 3 dB)5,000 km 的遠場接收功率（5 W、2×105.1 dB、3 dB） | \(37 + 210.2 - 3 + 20\log_{10}\frac{\lambda}{4\pi d}\) | −28 dBm = 1.6 µWcomputed. Matches paper p. 8 best case 1.6 µW計算值。與論文 p. 8 最佳情況 1.6 µW 一致 |
| Same formula at 200 m同一公式在 200 m | \(37 + 210.2 - 3 - 184.2\) | +60 dBm ≫ P_Tcomputed. Invalid: 200 m is inside the near field計算值，200 m 在近場內，超出此式適用範圍 |
| Required received power, 24 channels24 通道所需接收功率 | \(24 \times (-20\ \text{dBm})\) | 0.24 mWcomputed. Matches paper p. 8 "approximately 0.24 mW"計算值。與論文 p. 8「約 0.24 mW」一致 |
| Divergence and spot at 5,000 km, D = 10 cm發散角與 5,000 km 的光斑，D = 10 cm | \(1.22\lambda/D;\ \theta d\) | 18.9 µrad. 95 mcomputed. Paper p. 8 "at least 95 m"計算值。論文 p. 8「至少 95 m」 |
| Per-satellite rate if one aperture per lattice neighbor若每個格鄰居各一口徑的每星速率 | \(8 \times 9.6\ \text{Tb/s}\) | 76.8 Tb/scomputed. The terminal count remains unspecified計算值。終端數待指定 |
| Light time across the 2 km cluster穿越 2 km 叢集的光行時間 | \(2{,}000\ \text{m}/c\) | 6.7 µscomputed. Compare "tens of µs" in Section 10.2計算值。對照第 10.2 節的「數十 µs」 |
Original figure note原始圖說出處
Drawn for this primer from the far-field budget of Section 2.2 with inputs from Agüera y Arcas et al., Joule 2026, pp. 3, 8, and Schieler et al., ICSOS 2023.Stop and think停下來想一想
Set range to 100 m and then 200 m, the extremes reached twice per orbit. Observe the allowed beam count. If beam count sets rate, direct-link capacity drops fourfold and recovers twice per orbit. The array changes from 4×4 at 100 m to 2×2 at 200 m on a schedule known years ahead. Identify which companion collective algorithm could exploit this cycle, and which would run at its minimum capacity.
把距離設為 100 m 再設為 200 m，也就是直接鄰居鏈路每圈兩次經過的兩個極端，觀察串擾規則允許的光束數。若鏈路速率由光束數決定，格子中每條直接鏈路的容量每圈會依數年前就已知的時程兩次降為四分之一再恢復，從 100 m 的 4×4 陣列變成 200 m 的 2×2 陣列。姊妹入門中的哪一種集體通訊演算法能利用這一點，哪一種只會被拖慢到最低值？
11.4Published specifications and open design choices論文規格與待定設計
The table separates Suncatcher’s specified subsystems from open design tasks and gives their Joule sections. Each quantitative claim can therefore be checked against its source. Orbit, link budget, radiation and power each have numbers and methods. Network design still needs terminal count, switching architecture, topology beyond eight nearest neighbors, hop count and latency budget.
The paper’s only topology statement is the central satellite’s “8-neighborhood” in its Figure 2 caption [36, pp. 3–4, Fig. 2]. Full 81-satellite bisection therefore needs design details beyond Google’s published documents. The primer computes 76.8 Tbps per satellite by assuming one aperture per neighbor. This requires eight independently pointed terminals per spacecraft, with their mass and power budgets still open.
Suncatcher 論文區分了已指定的子系統與後續待完成的設計。下表把兩者分開，並附上 Joule 論文中每一項被指定或被承認為待確定的章節，讀者可據此對照原文核對。模式是：面向物理的子系統——軌道、鏈路預算、輻射與電力——各有數字與方法，而所有能把「每口徑速率」變成「一個網路」的項目——每顆衛星的終端數、交換架構、八個最近鄰居之外的拓樸、跳數與延遲預算——全部缺席。論文中唯一的拓樸敘述是 Figure 2 圖說中中央衛星的「8 鄰域」[36, pp. 3–4, Fig. 2]，由此可知 81 顆衛星叢集的對分頻寬需要 Google 公開文件之外的設計細節才能計算。上方計算表中每星 76.8 Tbps 是本入門對「每個鄰居各一口徑」的算術，屬於本教材另列的假設，且它需要每艘太空船有八個獨立指向的終端，其質量與電力預算待編列。
| Subsystem子系統 | Specified已指定 | Left open待指定 | Where [36]出處[36] |
|---|---|---|---|
| Orbit and formation軌道與編隊 | 650 km dawn–dusk SSO. 81 satellites, R = 1 km, square lattice. 100–200 m oscillation. 2:1.0037 ratio. DOP853 integration. Formation maintenance < 30 m/s per year (Joule only)650 km 晨昏 SSO。81 顆、R = 1 km、方形格。100–200 m 振盪。2:1.0037 軸比。DOP853 積分。編隊維持每年 < 30 m/s（僅 Joule） | Per-satellite propellant mass. Collision-avoidance manoeuvres. Flight control law (sketched only in the supplemental notes)每星推進劑質量。避撞機動。飛行控制律（僅在補充說明中概述） | pp. 2–5, 7–8. Suppl. |
| Inter-satellite link星間鏈路 | 24 × 400G DWDM, 9.6–12.8 Tbps per aperture. 5 W, 105.1 dB, 3 dB. Coherent detection. 800 Gbps bench. Multiplexing 2×2 at 1.25 km, 4×4 at 0.32 km24 × 400G DWDM，每口徑 9.6–12.8 Tbps。5 W、105.1 dB、3 dB。相干檢測。800 Gbps 實驗台。多工 2×2 在 1.25 km、4×4 在 0.32 km | Terminals per satellite. Acquisition time. Rate as a function of the oscillating range. Flight demonstration每星終端數。建鏈時間。速率隨振盪距離的函數。飛行示範 | pp. 3, 8, Fig. 1 |
| Cluster network叢集網路 | "8-neighborhood" of S0 only僅 S0 的「8 鄰域」 | Switching architecture. Topology. Hop count. Latency budget. Bisection bandwidth. Routing. Failure handling交換架構。拓樸。跳數。延遲預算。對分頻寬。路由。故障處理 | p. 4, Fig. 2 (statement limited to Fig. 2)（敘述限於 Fig. 2） |
| Compute and radiation運算與輻射 | Trillium TPUs. 750 rad(Si) over 5 yr. HBM onset 2 krad. SDC ≈ 1/17 rad. HBM UECC ≈ 1/44 rad. SEFI ≈ 1/5 kradTrillium TPU。5 年 750 rad(Si)。HBM 2 krad 起始。SDC ≈ 每 17 rad 一次。HBM UECC ≈ 每 44 rad 一次。SEFI ≈ 每 5 krad 一次 | Effect of SEEs on training ("requires further study"). Chips per satellite. ICI vs ISL tieringSEE 對訓練的影響（「需進一步研究」）。每星晶片數。ICI 與 ISL 的分層 | pp. 5–6, 9 |
| Power and economics電力與經濟 | ≈ 28 kW for a 575 kg v2-mini analogue. $3,600 → $200/kg. $810/kW/yr vs $570–3,000575 kg v2-mini 類比約 28 kW。$3,600 → $200/kg。$810/kW/yr 對 $570–3,000 | Compute hardware, ground segment and operations cost. Full economic feasibility remains open運算硬體、地面段與營運成本（完整經濟可行性仍待研究） | pp. 6, 9–10, Table 1 |
| Thermal熱控 | Heat pipes and radiators. HBM held below ≈ 90 °C. A 50 °C radiator loses ≈ 1.2% of emission per K. Validated in thermal vacuum (Google)熱管與散熱板。HBM 維持在約 90 °C 以下。50 °C 散熱板每低 1 K 輻射約少 1.2%。已於熱真空驗證（Google） | Orbital validation pending. Sustained duty cycle. Radiator area per kW在軌驗證待完成。持續工作週期。每 kW 散熱板面積 | pp. 2, 7. Google facts postGoogle facts 文章 company |
| Ground segment地面段 | Radio for the pilot. Optical "eventually". TBIRD 200 Gbps cited as state of the art試驗階段用射頻。「最終」用光學。引 TBIRD 200 Gbps 為現況技術 | Station network, visibility from a dawn–dusk orbit ("will increase latency to some ground locations"), ingest and checkpoint flows地面站網路、晨昏軌道的可見性（「會提高對某些地面位置的延遲」）、資料匯入與檢查點流量 | pp. 2, 7 |
Original figure note原始圖說出處
Drawn for this primer from Agüera y Arcas et al., Joule 2026, pp. 2–10, and the Google "facts" post of September 2026.The ground segment limits aggregate input and output even when internal links are fast. Pilot radio links and later optical downlinks therefore require their own capacity and visibility analysis.
即使內部鏈路很快，地面段仍限制總輸入與輸出。試驗用射頻鏈路與後續光學下行都需另外分析容量及可見性。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Capacity and rate | Qualitative scope | The ground segment limits the cluster’s total ingress and egress, even when its internal network has ample bandwidth. |
| Result and scope | Qualitative scope | The paper says that for a pilot radio "suffices, avoiding the challenges of overcoming atmospheric interference", that optical ground links will "eventually" be needed |
| Result and scope | Qualitative scope | and that it reserves detailed coverage for future work because commercial development is under way |
| Result and scope | Qualitative scope | and the state of the art it cites is NASA's TBIRD |
| Capacity and rate | 200, 1 m, 4.8 TB | TBIRD demonstrates 200 Gbps from a CubeSat to a one-meter adaptive-optics telescope. Its record pass transfers 4.8 TB in about five minutes [37, §I–IV][36, pp. 2, 7]. |
| Latency comparison | 2 | A dawn–dusk orbit also constrains which ground stations see the cluster and when. The authors note that it "will increase latency to some ground locations" (p. 2). |
| Capacity and rate | 9.6 | One 9.6 Tbps internal aperture is roughly fifty times faster than a TBIRD-class ground link. Such a ground path therefore restricts aggregate cluster ingress and egress. Section 13.5 studies this interface. |
| 結果與適用範圍 | 定性範圍 | 即使叢集內部頻寬充足，地面段仍限制了叢集的總輸入與輸出。 |
| 容量與速率 | 200, 1 m, 4.8 TB | 論文說試驗階段用射頻就「足夠，可避開克服大氣干擾的挑戰」，光學地面鏈路「最終」會需要，但因為業界正在開發，細節則留待後續討論，而它引用的現況技術是 NASA 的 TBIRD，該任務從一顆 CubeSat 以 200 Gbps 下傳到配備自適應光學的 1 m 望遠鏡，並在約五分鐘的一次過頂中傳輸了創紀錄的 4.8 TB[37, §I–IV][36, pp. 2, 7]。 |
| 結果與適用範圍 | 2 | 晨昏軌道也限制哪些地面站何時能看見叢集，作者指出它「會提高對某些地面位置的延遲」（p. 2）。 |
| 容量與速率 | 9.6 | 一個內部鏈路各承載 9.6 Tbps、但通往地球的唯一路徑只有 TBIRD 等級鏈路的叢集，其總進出頻寬大約比單一內部口徑小五十倍，第 13.5 節會把這個比值轉成一個研究方向。 |
11.5Cluster fabrics: lattice, torus, Clos and the packing question叢集網路架構：格狀、環面、Clos 與堆積問題
Recall
回顧
Section 10.4 compares laser-torus and terrestrial-Clos bisection. Section 2.4 gives modern-terminal acquisition of 2–30 s. Section 11.2 shows direct Suncatcher links stretching from 100 to 200 m and back twice per orbit.
Formation geometry determines clear paths and their lengths, while terminal count limits simultaneous links. Pénot and Balakrishnan compare packing designs under separation, envelope, sunlight and visibility constraints.
編隊幾何決定暢通路徑與距離，終端數則限制同時連線數。Pénot 與 Balakrishnan 在間距、包絡、日照與可見性條件下比較堆積設計。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Because the Google paper stops at the eight-neighborhood, the network inside the cluster has to be designed by someone else. It helps to name the object being designed. |
| Fabric design scope | Qualitative scope | A cluster fabric defines the formation’s internal network. Its choices include terminals per satellite and any satellites that switch traffic for others. |
| Result and scope | Qualitative scope | and the topology that results |
| Geometry and fabric | Qualitative scope | Formation geometry determines visible satellite pairs and their separation. Terrestrial fabrics usually address these choices separately from terminal count. |
| Result and scope | Qualitative scope | while the terminal count decides how many of those pairs can be linked at once. |
| Result and scope | Qualitative scope | Pénot and Balakrishnan treat the first decision as a packing problem. |
| Orbital configuration | 650 km, 98° | The model places satellites in the cluster center’s Hill frame at 650 km and 98° inclination. It maximizes satellite count subject to four whole-orbit conditions. These include minimum collision-avoidance spacing \(R_{\min}\), envelope radius \(R_{\max}\) and clear solar visibility. |
| Orbital configuration | Qualitative scope | and a stable set of nearest neighbors whose links remain clear throughout the orbit [102, Fig. 2] preprint. |
| Orbital configuration | 2:1, 81, 100, 1, 4 | Under those rules the Suncatcher square grid of 2:1 ellipses holds 81 satellites at \(R_{\min} = 100\) m and \(R_{\max} = 1\) km (Fig. 4) |
| Dataset and workload | Qualitative scope | and the authors trace its inefficiency to three causes |
| Result and scope | 2 | The usable area is an ellipse inside the circular envelope. At rectangular phases, the along-track stretch requires spacing \(2R_{\min}\). |
| Result and scope | Qualitative scope | and that the lattice is square where a hexagonal one packs more densely. |
| 結果與適用範圍 | 8 | 由於 Google 論文止步於 8 鄰域，叢集內部的網路仍待後續研究設計，因此先為要設計的對象命名會有幫助。 |
| 結果與適用範圍 | 定性範圍 | 叢集網路架構是編隊叢集內部的網路，包括每顆衛星攜帶的雷射終端數、衛星替其他衛星交換流量的配置，以及由此形成的拓樸，而它受兩個地面網路可分開處理的決定所左右：編隊幾何決定哪些衛星對能互相看見及其距離，終端數則決定其中有多少對能同時連線。 |
| 結果與適用範圍 | 定性範圍 | Pénot 與 Balakrishnan 把第一個決定當成堆積問題處理。 |
| 軌道配置 | 650 km, 98° | 他們把每顆衛星放在 650 km、傾角 98° 的叢集中心 Hill 座標系中，求在整圈軌道上滿足四個條件的最大衛星數：防撞所需的最小間距 \(R_{\min}\)、半徑 \(R_{\max}\) 的包絡、對太陽的完整視線，以及一組鏈路始終保持暢通的穩定最近鄰居[102, Fig. 2]preprint。 |
| 軌道配置 | 2:1, 100, 1, 81, 4, 2 | 在這些規則下，Suncatcher 由 2:1 橢圓構成的方形格在 \(R_{\min} = 100\) m、\(R_{\max} = 1\) km 時容納 81 顆衛星（Fig. 4），作者把它的低效歸結為三個原因：可用面積是內接於圓形包絡的橢圓、沿軌拉伸使矩形相位時該方向需要 \(2R_{\min}\) 的間距，以及格子是方形，六角形的堆積則更密。 |
Circular relative orbits remove this stretch. Each cross-track oscillation is in phase with radial motion and \(\sqrt3\) times larger. Their relative-element condition is \(i_d = \sqrt3\, e_d\). Adding this cross-track term to Section 11.2’s HCW solution gives the following result.
他們的替代方案以圓形相對軌道直接消除拉伸。若給每顆衛星一個與徑向振盪同相、幅度為其 \(\sqrt3\) 倍的橫軌振盪，也就是他們相對軌道根數記法中的條件 \(i_d = \sqrt3\, e_d\)，則打開橫軌項後第 11.2 節的 HCW 解變成
\[ x = A_0\cos(nt+\alpha),\quad y = -2A_0\sin(nt+\alpha),\quad z = \sqrt3\,A_0\cos(nt+\alpha)\ \ \Rightarrow\ \ x^2+y^2+z^2 = 4A_0^2 , \]
Circular relative orbits let the planar cluster rotate as a rigid body. Neighbor distances stay fixed, supporting a denser hexagonal layout with steady link conditions.
圓形相對軌道讓平面叢集以剛體方式旋轉，鄰居距離維持固定，能支援更密的六角配置與穩定鏈路條件。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 2, 0, 3 | Distance from cluster center is always \(2A_0\). The circular relative orbit lies in a plane tilted by \(\pi/3\) from radial toward cross-track. Pénot and Balakrishnan describe this in “Optimal Planar Solution” preprint. |
| Orbital configuration | Qualitative scope | This is the general circular orbit of formation-flying theory, and its consequence for the network is larger than its consequence for packing. |
| Formation geometry | 367, 6 | Every satellite circles in one plane at angular rate \(n\), making the cluster rotate rigidly. Pairwise distances stay fixed, allowing an \(R_{\min}\)-spaced hexagonal lattice. The same envelope then holds 367 satellites, “more than four times” Suncatcher’s count (Fig. 6). |
| Capacity and rate | 100, 200 m | This geometry removes Section 11.3’s square-lattice rate cycle. Direct links retain their spacing instead of stretching from 100 to 200 m and losing beams twice per orbit. |
| Capacity and rate | Qualitative scope | Each satellite retains six neighbors at distance \(R_{\min}\) throughout the formation’s life. The MIT authors use this to guarantee clear links. It also gives network designers a static graph with constant rates. |
| 結果與適用範圍 | 2, 0, 3 | 因此衛星與叢集中心的距離時時刻刻都是 \(2A_0\)，相對軌道是一個圓，位於從徑向朝橫軌方向傾斜 \(\pi/3\) 的平面上（Pénot & Balakrishnan，「Optimal Planar Solution」節）preprint。 |
| 結果與適用範圍 | 定性範圍 | 這就是編隊飛行理論中的一般圓形軌道，而它對網路的影響比對堆積的影響更大。 |
| 結果與適用範圍 | 367, 6 | 由於所有衛星在同一平面上以相同角速率 \(n\) 繞圈，整個叢集剛體式旋轉，每一對衛星的距離都維持固定，間距 \(R_{\min}\) 的六角格得以放入，同樣包絡內的衛星數升到 367 顆，「超過四倍」於 Suncatcher（Fig. 6）。 |
| 結果與適用範圍 | 100 m, 200 m | 第 11.3 節為方形格推導出的速率振盪——直接鏈路每圈兩次從 100 m 拉到 200 m 並失去多工光束——在圓形設計中消失，每顆衛星在編隊的整個壽命中都保有六個恰好相距 \(R_{\min}\) 的鄰居。 |
| 結果與適用範圍 | 定性範圍 | MIT 作者只用這個性質來保證鏈路暢通，但網路設計者會把它視為一張靜態、定速率的圖而格外重視。 |
Stacked formations scale with volume and eventually overtake planar packing. Their greater density also increases self-shading, making solar-array geometry part of the network design.
堆疊編隊隨體積成長，規模擴大後可超越平面堆積。較高密度也增加自我遮蔽，因此太陽能板幾何需納入網路設計。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 264, 39°, 7 | A third design stacks inclined planes along track and staggers adjacent planes at spacing \(R_{\min}\). Satellites leaving the envelope during an orbit are pruned. At Suncatcher dimensions, it fits 264 satellites at local inclination 39°, below the planar design’s count (Fig. 7). |
| Result and scope | Qualitative scope | Its advantage lies in scaling. |
| Result and scope | 0.80, 1.996, 3.63, 2.00, 0.27, 2.99, 3, 1, 13.5, 9 | The sweeps fit \(N = a\,(R_{\max}/R_{\min})^b\). Coefficients are \(a = 0.80,\ b = 1.996\) for Suncatcher’s grid, \(a = 3.63,\ b = 2.00\) for optimal planar packing and \(a = 0.27,\ b = 2.99\) for 3-D packing (Table 1). Planar counts scale with envelope area, while stacked counts scale with volume. The 3-D design overtakes the planar designs at \(R_{\max}/R_{\min} \ge 13.5\) (Fig. 9). |
| Power and thermal conditions | 50, 19 m, 3, 3 m, 10, 5 | Solar visibility is the third constraint and limits packing density. Satellites are modeled as discs of radius \(R_{\rm sat}\). Suncatcher’s grid keeps full illumination up to \(R_{\rm sat} = 50\) m. Optimal planar packing starts self-shading at 19 m, and 3-D packing at 3 m (Fig. 10, Table 5). |
| Power and thermal conditions | 575 kg, 105 m, 5.8 m | The paper’s 575 kg power analogue has about 105 m² of panels [36, p. 10], equivalent to a disc radius near 5.8 m. It fits planar packing while exceeding stacked packing’s shading threshold. The best-scaling layout would therefore shade Suncatcher-class satellites. |
| 結果與適用範圍 | 定性範圍 | 第三種設計沿軌堆疊多個傾斜平面，彼此錯開使相鄰平面的衛星保持 \(R_{\min}\) 的距離，並剔除任何在軌道期間離開包絡的衛星。 |
| 結果與適用範圍 | 39°, 264, 7 | 在 Suncatcher 的尺寸下它比平面設計差，在局部傾角 39° 時只有 264 顆（Fig. 7）。 |
| 結果與適用範圍 | 定性範圍 | 它的優勢在於縮放。 |
| 結果與適用範圍 | 0.80, 1.996, 3.63, 2.00, 3, 0.27, 2.99, 1, 13.5, 9 | 作者對掃描結果擬合 \(N = a\,(R_{\max}/R_{\min})^b\)，得到 Suncatcher 格子 \(a = 0.80,\ b = 1.996\)、最佳平面設計 \(a = 3.63,\ b = 2.00\)、3-D 設計 \(a = 0.27,\ b = 2.99\)（Table 1），所以平面設計隨包絡的面積成長，堆疊設計則隨其體積成長，而一旦 \(R_{\max}/R_{\min} \ge 13.5\)，3-D 設計就超越平面設計（Fig. 9）。 |
| 結果與適用範圍 | 50, 19 m, 3, 3 m, 10, 5 | 密度在日照上有代價，這是四個條件中的第三個：把每顆衛星模型化為半徑 \(R_{\rm sat}\) 的圓盤時，Suncatcher 格子在 \(R_{\rm sat} = 50\) m 以內都保有完整日照，最佳平面設計在 19 m 起開始自我遮蔽，3-D 設計則在 3 m 起（Fig. 10，Table 5）。 |
| 結果與適用範圍 | 575 kg, 105 m, 5.8 m | 論文自己的電力類比——一顆 575 kg、約 105 m² 太陽板的衛星[36, p. 10]——相當於半徑約 5.8 m 的圓盤，輕鬆放得進平面佈局，卻超過堆疊佈局的遮蔽門檻，所以縮放最好的設計，恰恰也是 Suncatcher 等級衛星會互相遮蔽的那一種。 |
| Formation design編隊設計 | Relative orbit相對軌道 | Lattice and neighbors格子與鄰居 | N at 100 m / 1 km100 m / 1 km 時的 N | Fit \(a,\ b\)擬合 \(a,\ b\) | Self-shading from自我遮蔽起點 |
|---|---|---|---|---|---|
| Suncatcher gridSuncatcher 格子 | 2:1 ellipse in the orbital plane軌道面內的 2:1 橢圓 | square. 8-neighborhood at 100–283 m, oscillating方形。8 鄰域距離 100–283 m 且振盪 | 81 | 0.80, 1.996 | \(R_{\rm sat}\) = 50 m\(R_{\rm sat}\) = 50 m |
| Optimal planar最佳平面 | circle in a plane tilted by π/3傾斜 π/3 平面上的圓 | hexagonal. 6 neighbors at fixed 100 m六角形。6 個鄰居固定在 100 m | 367 | 3.63, 2.00 | 19 m |
| 3-D stacked3-D 堆疊 | inclined planes staggered along track沿軌錯開的傾斜平面 | 3-D. Line of sight checked numerically3-D。視線需數值驗證 | 264 (\(i_{\rm local}\) = 39°)264（\(i_{\rm local}\) = 39°） | 0.27, 2.99 | 3 m |
Source: Pénot & Balakrishnan, arXiv 2605.15335 v2, Figs. 4, 6, 7, 9, 10, Tables 1, 5 preprint. Neighbor distances for the Suncatcher grid from the calc table in Section 11.2.來源：Pénot & Balakrishnan，arXiv 2605.15335 v2，Figs. 4、6、7、9、10，Tables 1、5 preprint。Suncatcher 格子的鄰居距離取自第 11.2 節的計算表。
The Clos mapping assigns each satellite one network role and requires clear whole-orbit links. Its feasibility depends on switch degree, layer count and the formation’s visibility graph.
Clos 對應為每顆衛星分配一個網路角色，並要求鏈路整圈保持暢通。可行性取決於交換度數、層數及編隊的可見性圖。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Terminal-count constraint | Qualitative scope | The second decision, how to wire the satellites that fit, is where the terminal count binds. |
| Capacity and rate | 1/2, 3, 2/3, 2 | Nearest-neighbor meshes scale poorly beyond a few dozen satellites. In planar meshes, diameter, mean path length and bisection bandwidth grow as \(N^{1/2}\), while cut-crossing traffic grows with \(N\). Three-dimensional mesh bisection grows as \(N^{2/3}\) (Table 2). |
| Orbital configuration | Qualitative scope | The authors map a VL2-like Clos onto the cluster, assigning one satellite per Clos node. Every Clos edge needs an ISL maintaining line of sight throughout the orbit. |
| Clos node capacity | 4, 12, 1, 3, 2 | Switch satellites have at most \(k\) simultaneous links, swept from 4 to 12 in steps of two. With \(L\) layers, node capacity is \(k+1\) at \(L = 1\) and \(3k/2\) at \(L = 2\). For \(L \ge 3\), capacity is \((k/2)^{L-1} + (2L-3)(k/2)^{L-2}\). The compute-bearing top-of-rack nodes number \((k/2)^{L-1}\) (Table 3). |
| Result and scope | Qualitative scope | The mapping is a pure integer feasibility problem solved with Gurobi with a feasibility-only objective |
| Result and scope | 500, 2,000 m, 4, 12, 0, 15 m, 2, 3 | The sweep covers \(R_{\max}\) = 500–2,000 m, \(k\) = 4–12 and \(R_{\rm sat}\) = 0–15 m. Mapping is usually infeasible at \(L \le 2\) and feasible throughout at \(L \ge 3\). |
| Orbital configuration | 37, 300, 30, 500, 10, 3, 4, 13–14 | Worked examples fit 37 planar satellites at \(R_{\max} = 300\) m and 30 three-dimensional satellites at \(R_{\max} = 500\) m. Both use \(k = 10\) and \(L = 3\) (Table 4, Figs. 13–14). |
| 結果與適用範圍 | 定性範圍 | 第二個決定——如何把放得下的衛星連起來——正是終端數成為限制之處。 |
| 結果與適用範圍 | 1/2, 3, 2/3, 2 | Pénot 與 Balakrishnan 首先指出，最近鄰網格在超過幾十顆衛星後就擴展效率較低，因為平面網格的直徑、平均路徑長與對分頻寬都只隨 \(N^{1/2}\) 成長，而必須穿越中央的流量卻隨 \(N\) 成長，3-D 網格也只把對分頻寬提高到 \(N^{2/3}\)（Table 2）。 |
| 結果與適用範圍 | 定性範圍 | 接著他們問：如何把類 VL2 的 Clos 鋪到叢集上，讓每顆衛星恰好扮演 Clos 的一個節點，而每條 Clos 邊都由一對在整圈軌道上保持視線的衛星之間的 ISL 承載。 |
| 結果與適用範圍 | 4, 12, 2 | 擔任交換的衛星最多同時有 \(k\) 條鏈路，\(k\) 從 4 到 12 以 2 為步長掃描。 |
| 結果與適用範圍 | 1, 2, 3 | \(L\) 層的架構在 \(L = 1\) 時容納 \(k+1\) 個全互連節點，\(L = 2\) 時容納 \(3k/2\) 個，\(L \ge 3\) 時容納 \((k/2)^{L-1} + (2L-3)(k/2)^{L-2}\) 個，其中 \((k/2)^{L-1}\) 個是承載運算的機架頂端（ToR）節點（Table 3）。 |
| 結果與適用範圍 | 定性範圍 | 這個對應是純粹的整數可行性問題，以 Gurobi 求解、僅檢驗可行性。 |
| 結果與適用範圍 | 500, 2,000 m, 4, 12, 0, 15 m, 2, 3, 300, 37, 30, 10, 13–14 | 在 \(R_{\max}\) 從 500 到 2,000 m、\(k\) 從 4 到 12、\(R_{\rm sat}\) 從 0 到 15 m 的掃描中，\(L \le 2\) 通常落在可行範圍之外，\(L \ge 3\) 則處處可行，示範解包括 \(R_{\max} = 300\) m 的 37 顆平面衛星與 \(R_{\max} = 500\) m 的 30 顆 3-D 衛星，兩者都是 \(k = 10\)、\(L = 3\)（Table 4，Figs. 13–14）。 |
Switch satellites consume positions, launch mass and power that could otherwise support accelerators. The compute fraction therefore measures a physical cost of the orbital Clos.
交換衛星佔用原本可供加速器使用的位置、發射質量與電力，因此運算比例量化了軌道 Clos 的實體代價。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The result is encouraging about geometry and sobering about compute, because a Clos needs nodes dedicated to switching. |
| Power and thermal conditions | 4, 6, 8 | A switch satellite dedicates its payload to switching in a Clos mapping. It occupies a formation position, launch mass and power that could otherwise support accelerators. The remaining compute fraction is \(r(k,L) = k/(k+4L-6)\) (Pénot & Balakrishnan, Eq. 8) preprint. |
| Compute fraction | 81, 55%, 367, 36% | At six ports and three layers, half the satellites switch traffic. An 81-satellite cluster needs four layers even with twelve ports, leaving about 55% computing. A 367-satellite cluster needs five layers at eight ports, leaving about 36% computing. The calculation table reproduces these costs. |
| Capacity and rate | 51.2 | A terrestrial switch is a 51.2 Tbps chip inside a rack [65, §3]. An orbital switch occupies a whole spacecraft. This difference prevents direct transfer of the companion topology chapter’s assumptions. |
| 結果與適用範圍 | 定性範圍 | 這個結果確立了幾何可行性，也量化了運算容量的代價，因為 Clos 需要只做交換的節點。 |
| 結果與適用範圍 | 4, 6, 8 | 交換衛星是在 Clos 對應中專責交換的衛星，它佔據編隊中的一個位置，以及一份原本可以承載加速器的發射質量與電力，而架構中仍在運算的比例為 \(r(k,L) = k/(k+4L-6)\)（Pénot & Balakrishnan，Eq. 8）preprint。 |
| 結果與適用範圍 | 定性範圍 | 六個埠、三層時有一半衛星是交換衛星。 |
| 結果與適用範圍 | 81, 55%, 367, 36% | 下方計算表顯示，81 顆衛星的叢集即使每顆交換衛星有十二個埠也需要四層，只剩約 55% 的架構在運算，而 367 顆衛星在八個埠時需要五層，只剩約 36%。 |
| 容量與速率 | 51.2 | 在地面上交換器是機架裡的一顆 51.2 Tbps 晶片[65, §3]，在軌道上它卻是編隊中的一整艘太空船，這個差異就是姊妹教材拓樸章節需要重新設計的全部原因。 |
Van Berkel’s torus comparison separates the effect of link speed from the remaining bisection gap. Faster Suncatcher-class links reduce the gap while topology still limits scaling.
van Berkel 的環面比較可分開檢視鏈路速率與剩餘對切差距。較快的 Suncatcher 級鏈路縮小差距，拓樸仍限制規模擴展。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Capacity and rate | 8,000, 3 | Van Berkel quantifies the laser-torus bisection loss against terrestrial Clos at 8,000 nodes. His Table 3 states the gap [101, §IV-A, Tab. 3] preprint. |
| Capacity and rate | 8,000, 72, 2:1, 100 GB/s, 89, 90, 20, 100 Gb/s, 12.5 GB/s | The terrestrial Clos has 8,000 racks, 72 ToR ports, 2:1 oversubscription and 100 GB/s links. The compared tori contain 89 × 90 satellites in 2-D and 20 × 20 × 20 in 3-D. Their 100 Gb/s laser ISLs equal 12.5 GB/s, with four to six terminals per satellite. |
| Latency comparison | 28,800 TB/s, 2.25 TB/s, 2, 10 TB/s, 3, 4 µs, 1 µs, 1,800 µs, 600 µs, 20 µs | Bisection is 28,800 TB/s for Clos, 2.25 TB/s for 2-D torus and 10 TB/s for 3-D torus. Clos diameter latency is 4 µs, assuming \(\log N\) switch traversals at 1 µs each. Torus diameter latencies are 1,800 µs and 600 µs at 20 µs per hop. |
| Capacity and rate | 2, 1/2 | His text writes the 2-D torus bisection as \(N^{1/2}B_l\) while the table uses \(2N^{1/2}B_l\), and this chapter quotes the table. |
| Orbital configuration | 8,000, 32,000–48,000 | An 8,000-satellite torus needs adaptive steering of 32,000–48,000 beams. Its planar wraparound links are “particularly challenging”. Van Berkel proposes a bi-planar “Clos satellite constellation” with separate compute and communication satellites, leaving quantification open. Pénot and Balakrishnan quantify this switch-satellite concept. |
| Capacity and rate | 3, 9.6 Tb/s, 960 TB/s | His rate uses the Starlink reference, below Suncatcher’s. Scaling the 3-D torus to 9.6 Tb/s per link raises bisection to 960 TB/s, still roughly thirty times below Clos. Suncatcher narrows the gap from over three orders to between one and two, leaving a substantial residual. |
| 結果與適用範圍 | 8,000, 3 | van Berkel 從另一端切入同一個問題：在 8,000 節點的規模下，以雷射佈線的環面相對地面 Clos 會損失多少對分頻寬，而他的 Table 3 是這個差距最清楚的單一陳述[101, §IV-A, Tab. 3]preprint。 |
| 容量與速率 | 8,000, 72, 2:1, 100 GB/s, 89, 90, 20, 12.5 GB/s, 100 Gb/s | 地面參考是 8,000 個機架、ToR 寬度 72、2:1 超額訂閱、100 GB/s 鏈路的 Clos，與之對照的是 89 × 90 的二維環面與 20 × 20 × 20 的三維環面，其鏈路是 12.5 GB/s 的 100 Gb/s 雷射 ISL，每顆衛星四到六個終端。 |
| 容量與速率 | 28,800 TB/s, 2, 2.25 TB/s, 3, 10 TB/s | 對分頻寬各自為 Clos 28,800 TB/s、2-D 環面 2.25 TB/s、3-D 環面 10 TB/s。 |
| 結果與適用範圍 | 4 µs, 1 µs, 20 µs, 1,800 µs, 600 µs | 延遲則是 Clos 4 µs（\(\log N\) 次交換、每次 1 µs），兩種環面則以每跳 20 µs 計為 1,800 µs 與 600 µs。 |
| 結果與適用範圍 | 2, 1/2 | 他的正文把 2-D 環面對分頻寬寫成 \(N^{1/2}B_l\)，表格卻用 \(2N^{1/2}B_l\)，本章引用表格的值。 |
| 結果與適用範圍 | 8,000, 32,000–48,000 | 他還指出 8,000 顆衛星的環面需要自適應操控 32,000–48,000 道光束、環面的環繞鏈路在平面編隊中相當困難，並提出（量化分析則待完成）一種由運算衛星與通訊衛星分開組成的雙平面「Clos 衛星星座」，這正是 Pénot 與 Balakrishnan 所量化的交換衛星構想。 |
| 容量與速率 | 3, 9.6 Tb/s, 960 TB/s | 他的鏈路速率是 Starlink 的數字，Suncatcher 的速率則更高，若把 3-D 環面線性放大到每鏈路 9.6 Tb/s，其對分頻寬升到 960 TB/s，仍比 Clos 低約三十倍，所以 Suncatcher 鏈路把差距從三個數量級以上縮小到一到兩個數量級之間，仍留有差距。 |
| Fabric, N = 8,000架構，N = 8,000 | Bisection formula對分頻寬公式 | Bisection對分頻寬 | Latency formula延遲公式 | Latency延遲 |
|---|---|---|---|---|
| Terrestrial Clos (W = 72, 2:1, 100 GB/s)地面 Clos（W = 72、2:1、100 GB/s） | \(\tfrac12 N W B_l\) | 28,800 TB/s | \(\log N \cdot 1\ \mu\text{s}\) | 4 µs |
| 2-D torus 89 × 90, 12.5 GB/s ISLs2-D 環面 89 × 90，12.5 GB/s ISL | \(2 N^{1/2} B_l\) | 2.25 TB/s | \(N^{1/2} \cdot 20\ \mu\text{s}\) | 1,800 µs |
| 3-D torus 20 × 20 × 20, 12.5 GB/s ISLs3-D 環面 20 × 20 × 20，12.5 GB/s ISL | \(2 N^{2/3} B_l\) | 10 TB/s | \(\tfrac32 N^{1/3} \cdot 20\ \mu\text{s}\) | 600 µs |
Source: [101, §IV-A]–B, Table 3 preprint. The table's 2-D torus formula is used, the text's differs by a factor of two (factbase Part C, item 41).來源：[101, §IV-A]–B，Table 3 preprint。採用表格中的 2-D 環面公式，正文的公式相差兩倍。
Joint network design must include range-dependent rate, partner acquisition and partial blocking. These effects change the value of a geometrically feasible Clos or lattice.
聯合網路設計需納入隨距離變化的速率、連線對象擷取與部分遮擋，這些效應會改變幾何可行的 Clos 或晶格之實際效益。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Model scope | 11.2, 11.3 | Sections 11.2 and 11.3 identify three formation properties outside both analyses’ models, each with a separate effect on their conclusions. |
| Capacity and rate | Qualitative scope | The first is range-dependent rate. |
| Capacity and rate | Qualitative scope | Pénot and Balakrishnan state that maximum ISL range, link budget and data rate are outside their model |
| Orbital configuration | 1.5 km, 100 m | so a Clos edge between satellites 1.5 km apart counts the same as one between neighbors 100 m apart |
| Capacity and rate | 0.32 km, 4×4, 2×2, 1.25 km, 16 | The paper supports 4×4 multiplexing within 0.32 km. Beyond the 2×2 range of 1.25 km, a link uses one beam. This changes rate by a factor of 16 ([36, p. 3, Fig. 1]). |
| Capacity and rate | Qualitative scope | and van Berkel assigns one rate to every link of a torus. |
| Result and scope | Qualitative scope | The second is re-pointing. |
| Result and scope | 2–30 s, 6.7 µs | The MIT mapping is static, and van Berkel excludes acquisitions when counting beams. A fabric changing partners for collectives or failure recovery pays Section 2.4’s 2–30 s acquisition per change. This exceeds the cluster’s 6.7 µs light-time by five to six orders of magnitude. |
| Capacity and rate | Qualitative scope | The third is blocking treated as a rate effect. |
| Orbital configuration | 3 | The line-of-sight matrix is binary across a full orbit. A pair blocked for part of that orbit is discarded, with scheduled use outside the model. The authors require numerical visibility checks for 3-D clusters. Their remark implies partial links occur most often in the densest designs. |
| Result and scope | Qualitative scope | The remaining question is how to choose terminals, geometry and fabric together. |
| Capacity and rate | Qualitative scope | Rate depends on range, partner changes cost seconds, and visibility varies over the orbit. |
| Result and scope | Qualitative scope | This is the problem that Section 13.2 takes up. |
| 結果與適用範圍 | 11.2, 11.3 | 第 11.2 與 11.3 節指出三項編隊特有性質，均列於兩份分析的模型範圍之外，各自影響了結論。 |
| 結果與適用範圍 | 定性範圍 | 第一項是隨距離變化的速率。 |
| 軌道配置 | 1.5 km, 100 m | Pénot 與 Balakrishnan 明言 ISL 最大距離、鏈路預算與資料速率列於其模型範圍之外，因此相距 1.5 km 的衛星之間的 Clos 邊與相距 100 m 的鄰居之間的邊被視為等價。 |
| 軌道配置 | 0.32 km, 4×4, 2×2, 1.25 km, 16 | 然而依論文的多工數字，0.32 km 以內的鏈路可承載 4×4 陣列，超過 2×2 距離 1.25 km 的鏈路只能承載單一光束，速率相差 16 倍（[36, p. 3, Fig. 1]），而 van Berkel 則給環面的每條鏈路同一個速率。 |
| 結果與適用範圍 | 定性範圍 | 第二項是重新指向。 |
| 結果與適用範圍 | 2–30 s, 6.7 µs | MIT 的對應在建構上就是靜態的，van Berkel 計算了光束數擷取次數則列於範圍之外，然而任何會更換連線對象的架構——包括配合集體通訊模式或繞過故障衛星——每次更換都要付出第 2.4 節的 2–30 s 擷取時間，這比穿越叢集的 6.7 µs 光行時間長五到六個數量級。 |
| 結果與適用範圍 | 定性範圍 | 第三項是把遮擋當成速率效應。 |
| 結果與適用範圍 | 3 | 視線矩陣在整圈軌道上是二元的，所以被另一顆衛星的圓盤遮擋部分軌道的鏈路會被直接捨棄，依遮擋時程安排鏈路容量則留待後續研究，而作者自己指出 3-D 叢集需要數值驗證視線，這意味著此類部分可用鏈路恰恰最常出現在堆積最好的設計中。 |
| 結果與適用範圍 | 定性範圍 | Pénot 與 Balakrishnan 已分析 Clos 的編隊對應[102]，van Berkel 則量化了環面的差距。 |
| 結果與適用範圍 | 定性範圍 | 後續問題是在速率隨距離變化、更換對象要花數秒、可見性隨軌道變化的情況下，如何同時選擇終端、幾何與架構，這就是第 13.2 節要處理的問題。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Compute fraction, k = 6 ports, L = 3 layers運算比例，k = 6 埠、L = 3 層 | \(r = \frac{6}{6+12-6}\) | 6/12 = 0.5computed from Pénot & Balakrishnan Eq. 8由 Pénot & Balakrishnan Eq. 8 計算 |
| Layers for 81 satellites at k = 12k = 12 時 81 顆衛星所需層數 | \(6^2 + 3\cdot 6 = 54 < 81;\ 6^3 + 5\cdot 6^2 = 396\) | L = 4, r = 12/22 ≈ 0.55L = 4，r = 12/22 ≈ 0.55computed from Table 3 capacity and Eq. 8由 Table 3 的容量與 Eq. 8 計算 |
| Layers for 367 satellites at k = 8k = 8 時 367 顆衛星所需層數 | \(4^3 + 5\cdot 4^2 = 144;\ 4^4 + 7\cdot 4^3 = 704\) | L = 5, r = 8/22 ≈ 0.36L = 5，r = 8/22 ≈ 0.36computed. Fraction of the unpruned fabric計算值。剔除節點前架構的比例 |
| ISL hops between two compute satellites, L = 4L = 4 時兩顆運算衛星間的 ISL 跳數 | \(2(L-1)\) | 6computed: up to the top layer and down計算值：上到頂層再下來 |
| Neighbours strictly within 200 m on a 100 m square lattice100 m 方形格上嚴格小於 200 m 的鄰居數 | \(4 \text{ at } 100\ \text{m} + 4 \text{ at } 141\ \text{m}\) | 8computed. The paper's "8-neighborhood" (Fig. 2)計算值。論文的「8 鄰域」（Fig. 2） |
| Fits at Suncatcher's ratio \(R_{\max}/R_{\min} = 10\)Suncatcher 比值 \(R_{\max}/R_{\min} = 10\) 時的擬合值 | \(0.80\cdot10^{1.996};\ 3.63\cdot10^{2};\ 0.27\cdot10^{2.99}\) | 79 · 363 · 264computed. Reported counts 81, 367, 264計算值。報告的數量為 81、367、264 |
| Planar–3-D crossover from the fits由擬合求平面與 3-D 的交叉點 | \(3.63\,\rho^{2} = 0.27\,\rho^{2.99}\) | ρ ≈ 13.8computed. The paper reports 13.5 from its data (Fig. 9)計算值。論文由資料報告為 13.5（Fig. 9） |
| Disc radius of the paper's 105 m² array論文 105 m² 太陽板的等效圓盤半徑 | \(\sqrt{105/\pi}\) | ≈ 5.8 mcomputed. Below the 19 m planar threshold, above the 3 m 3-D one計算值。低於平面的 19 m 門檻，高於 3-D 的 3 m |
| 3-D torus bisection at 9.6 Tb/s (1.2 TB/s) per link每鏈路 9.6 Tb/s（1.2 TB/s）時的 3-D 環面對分頻寬 | \(2 \cdot 8000^{2/3} \cdot 1.2\ \text{TB/s}\) | 960 TB/s, 30× below 28,800960 TB/s，比 28,800 低 30 倍computed from van Berkel Table 3 with the Joule link rate以 van Berkel Table 3 與 Joule 鏈路速率計算 |
| Acquisition time over light time across the cluster擷取時間與穿越叢集光行時間之比 | \((2\ \text{to}\ 30\ \text{s}) / 6.7\ \mu\text{s}\) | 3×10⁵ – 4.5×10⁶computed. Acquisition from Section 2.4計算值。擷取時間取自第 2.4 節 |
| Rate of a 4×4 array over a single beam4×4 陣列相對單一光束的速率 | \(4^2 \times 9.6\ \text{Tb/s}\ \text{vs}\ 9.6\ \text{Tb/s}\) | 154 vs 9.6 Tb/s, 16×154 對 9.6 Tb/s，16 倍computed. Array sizes from [36, p. 3, Fig. 1]計算值。陣列大小取自 [36, p. 3, Fig. 1] |
Original figure note原始圖說出處
Drawn for this primer from Agüera y Arcas et al., Joule 2026, p. 4, Fig. 2, and Pénot & Balakrishnan, arXiv 2605.15335 v2, Figs. 4, 6, 12, Table 3, Eq. 8.Stop and think停下來想一想
Figure 11.4(b)’s hexagonal links stay at 100 m, inside the paper’s 0.32 km 4×4 range. A Clos edge across that cluster can exceed one kilometer. Assume six apertures per satellite, each pointed at a hexagonal neighbor. Compare the fully multiplexed mesh with a Clos whose long edges use one beam. Explain how all-to-all-dominated training changes the choice relative to ring all-reduce.
圖 11.4(b) 的六角設計讓每條鏈路固定在 100 m，遠在論文 4×4 多工距離 0.32 km 之內，而鋪在同一叢集上的 Clos 邊可能跨越一公里以上。若每顆衛星帶六個口徑、各自指向六個六角鄰居，你寧可用全多工速率運行網格，還是用長邊降到單一光束的 Clos？如果訓練工作以全對全流量，環狀全歸約則作為對照為主，答案又會如何改變？
11.6Redesign list: what from the terrestrial stack survives重新設計清單：地面網路堆疊中哪些能存活
Recall
回顧
Section 10.3 walked the companion primer's design chapters onto a constellation and sorted them into keep, adapt and replace. A formation cluster is a more favorable case, with short, fast and nearly stationary links, so this section repeats the audit against the Suncatcher numbers.
第 10.3 節把姊妹教材的設計章節逐一對應到星座上，分為保留、調整與替換三類。編隊叢集是更有利的情況，鏈路短、快且幾乎靜止，因此本節以 Suncatcher 的數字重做一次這項檢查。
The terrestrial parallelism hierarchy still applies: keep the heaviest traffic on ICI and carry lighter collectives across apertures. The step budget remains compute plus exposed communication.
地面的平行化層級仍適用，最重流量留在 ICI，較輕的集體通訊再跨越孔徑。步時間預算仍為運算加上暴露通訊。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | Qualitative scope | The companion primer's design chapters were written for a hall of racks joined by switch ASICs. The useful question for each is which satellite fact, if any, breaks its assumptions once the racks become satellites a hundred meters apart. |
| Model and compute | 10.3 | Parallelism follows Section 10.3’s hierarchy. A satellite carrying Google’s “dozens of TPU chips” keeps tensor parallelism on its inter-chip interconnect [100] company. |
| Capacity and rate | 6.4 | which runs at 6.4 Tbps per Trillium chip [99] company |
| Capacity and rate | 9.6, 3 | Pipeline and data parallelism cross 9.6 Tbps apertures. This tier resembles the terrestrial host-to-NIC boundary. Joule describes ICI as “hundreds of Gbps per chip” (p. 3). |
| Result and scope | Qualitative scope | The step budget survives as well, because its identity of compute time plus exposed communication applies across transmission media |
| Latency comparison | 6.7 µs, 2 km, 3 | The step identity keeps its form while orbital inputs change. Light crosses a 2 km cluster in 6.7 µs, comparable with Meta’s reported multi-hop “tens of microseconds” [77, §3.3.3] preprint. |
| Constellation-scale latency | 1,800 µs, 600 µs, 8,000 | while the 1,800 µs and 600 µs of van Berkel's tori belong to 8,000-satellite constellations rather than to a kilometer-scale formation. |
| 結果與適用範圍 | 定性範圍 | 姊妹教材的設計章節是為以交換晶片相連的一整廳機架而寫，對每一章有用的問題是：當機架變成相距一百公尺的衛星時，哪一項衛星事實（若有的話）會打破它的假設。 |
| 容量與速率 | 10.3, 6.4, 9.6 | 平行化以第 10.3 節所給的形式存活：一顆搭載 Google 現在所說「數十顆 TPU 晶片」的衛星[100]company，把張量平行留在晶片間互連上，每顆 Trillium 晶片的 ICI 為 6.4 Tbps[99]company，並把流水線與資料平行送過每個 9.6 Tbps 的口徑。 |
| 容量與速率 | 3 | 這是一條與地面主機對網路卡邊界形狀相同的層級邊界，儘管 Joule 論文本身只把 ICI 描述為「每晶片數百 Gbps」（p. 3）。 |
| 結果與適用範圍 | 2 km, 6.7 µs, 3, 1,800 µs, 600 µs, 8,000 | 步時間預算也存活，因為「運算時間加上暴露通訊」這個恆等式適用於各種傳輸介質，叢集只改變它的輸入，而其中一項對軌道有利：穿越 2 km 叢集的 6.7 µs 光行時間，與 Meta Llama 3 報告中多跳網路的「數十微秒」相當[77, §3.3.3]preprint，van Berkel 環面的 1,800 µs 與 600 µs 則屬於 8,000 顆衛星的星座，公里尺度編隊則採另一組條件。 |
Collectives and routing can use a capacity graph known from formation dynamics. Checkpoint and recovery logic still applies, with radiation and thermal events supplying new inputs.
集體通訊與路由可利用編隊動力學推得的容量圖。檢查點與恢復邏輯仍適用，輻射及散熱事件則提供新的輸入條件。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | Qualitative scope | Collectives and routing must be adapted, and the formation makes the adaptation concrete with concrete orbital constraints. |
| Capacity and rate | 1, 3, 2 | In the square lattice each direct link's length, and through spatial multiplexing its rate, follows \(s\sqrt{1+3\sin^2 nt}\). Thus, the capacity graph varies with time but is known years ahead from the HCW solution. This means a collective can be scheduled against that graph instead of discovering it by probing. |
| Orbital configuration | Qualitative scope | In the circular-orbit design of Section 11.5 the graph remains static, which turns the adaptation into the far easier problem of a static, low-degree fabric. |
| Result and scope | Qualitative scope | Routing uses a precomputed orbital-phase function for the first case and a fixed table for the second. Both replace terrestrial ECMP hashing for unpredictable flows with predetermined routes. |
| Power and thermal conditions | 17 rad, 5 krad, 150 rad, 15–20 | Failure handling keeps the companion’s checkpoint-and-restart logic with revised inputs. Trillium tests give about one silent corruption per 17 rad and one SEFI per 5 krad per chip [36, pp. 5, 9]. Annual dose is roughly 150 rad. The flown MVP reportedly shuts down chips every 15–20 minutes for cooling press. |
| Result and scope | Qualitative scope | and van Berkel notes that a satellite lost for good leaves a non-isotropic network that must be rebalanced and rerouted [101, §IV-A] preprint. |
| 結果與適用範圍 | 定性範圍 | 集體通訊與路由必須調整，而編隊讓這種調整變得具體且有軌道條件可依循。 |
| 結果與適用範圍 | 1, 3, 2 | 在方形格中，每條直接鏈路的長度——以及透過空間多工決定的速率——遵循 \(s\sqrt{1+3\sin^2 nt}\)，所以容量圖隨時間變化，卻能由 HCW 解提前數年得知，這表示集體通訊可以依照這張圖排程，可以直接利用軌道模型。 |
| 結果與適用範圍 | 定性範圍 | 在第 11.5 節的圓形軌道設計中，這張圖維持靜態，調整就變成容易得多的靜態、低度數架構問題。 |
| 結果與適用範圍 | 定性範圍 | 路由同樣地在第一種情況下成為軌道相位的預先計算函數，在第二種情況下成為一張固定的表，兩者都以預定路由取代地面網路分散隨機流量所用的等價多路徑雜湊。 |
| 可靠度與準確率 | 17 rad, 5 krad, 150 rad, 15–20 | 故障處理採用調整，因為姊妹教材的檢查點與重啟邏輯仍然適用，只是輸入改變：Trillium 測試給出每約 17 rad 一次靜默資料損毀、每晶片每 5 krad 一次 SEFI，年劑量約 150 rad[36, pp. 5, 9]，已飛行的 MVP 據報每 15–20 分鐘必須關閉晶片散熱press，而 van Berkel 指出永久失效的衛星會留下一個方向性相異的網路，必須重新平衡與改道[101, §IV-A]preprint。 |
Three chapters need replacement, while sharing falls outside the present scope. Satellite degree follows independently pointed terminal count, which the paper leaves open. Orbital Clos costs whole switch satellites at compute fraction \(r(k,L)\). Tori instead lose orders of bisection bandwidth (Section 11.5). The companion’s lossless RDMA uses priority flow control through hop-by-hop switch pause frames. Orbital links also fade through pointing jitter and thermal flexing, with congestion as a separate cause.
Formation motion removes Doppler and point-ahead (Section 11.2), while Tbps free-space transport still needs measurement. The front end requires ground-transfer design. TBIRD demonstrates 200 Gbps and 4.8 TB in roughly five minutes [37, §I–IV]. Suncatcher’s pilot uses radio, and its dawn–dusk orbit constrains station visibility [36, pp. 2, 7]. The companion’s sharing chapter covers multiple tenants. This operator runs one job, so the present cluster has a single tenant.
The table connects each verdict to its determining fact.
As in Figure 11.3, compute and physics preserve some concepts, while the network requires new designs.
有三章必須替換，第四章則省略。拓樸是最明顯的例子，因為衛星的度數就是它獨立指向的終端數，而論文對此仍待指定。軌道上的 Clos 要以運算比例 \(r(k,L)\) 付出整顆交換衛星，環面則會損失數個數量級的對分頻寬（第 11.5 節）。傳輸層被替換，因為姊妹教材的無損 RDMA 依賴優先權流量控制，也就是由交換器逐跳發出的暫停訊框，而叢集鏈路的衰落來自指向抖動與熱撓曲，壅塞則是另一種原因。編隊消除了多普勒與前置指向（第 11.2 節），但Tbps 自由空間鏈路上的傳輸協定仍待量測。
前端被替換，因為通往地面的唯一路徑是下行鏈路，其光學現況技術是 TBIRD 的 200 Gbps、約五分鐘的一次過頂傳 4.8 TB[37, §I–IV]，Suncatcher 試驗任務使用射頻，晨昏軌道又限制哪些地面站能看見叢集[36, pp. 2, 7]。共用——姊妹教材關於多租戶的章節——暫時省略，因為一個由單一營運者為單一工作擁有並運行的叢集，僅需滿足單一租戶需求。下表彙整各項結論及迫使其改變的事實，其模式重複了圖 11.3 對論文本身所呈現的模式：能存活的是取決於運算與物理的部分，必須重建的是網路。
| Companion chapter姊妹教材章節 | Verdict for a Suncatcher cluster對 Suncatcher 叢集的結論 | Satellite fact that forces it迫使改變的衛星事實 | Where出處 |
|---|---|---|---|
| 3 Parallelism3 平行化 | keep保留 | ICI 6.4 Tbps per chip on board company. 9.6 Tbps per aperture between satellites. "dozens of TPU chips" per satellite company機上每晶片 ICI 6.4 Tbps company。衛星間每口徑 9.6 Tbps。每星「數十顆 TPU 晶片」company | 11.3, 10.3 |
| 4 Collectives4 集體通訊 | adapt調整 | Square lattice: link length and beam count follow \(s\sqrt{1+3\sin^2 nt}\), known in advance. Circular design: static方形格：鏈路長度與光束數遵循 \(s\sqrt{1+3\sin^2 nt}\)，可預先得知。圓形設計：靜態 | 11.2, 13.2 |
| 5 Step budget5 步時間預算 | keep保留 | Identity unchanged. Light time 6.7 µs across 2 km vs "tens of µs" on the ground恆等式保持成立。2 km 光行時間 6.7 µs，對照地面「數十 µs」 | 11.3 |
| 6 Topology6 拓樸 | replace替換 | Degree = apertures, unspecified. Clos needs L ≥ 3 and switch satellites, r(k, L). Tori lose 1–3 orders of bisection preprint度數＝口徑數，待指定。Clos 需 L ≥ 3 與交換衛星，r(k, L)。環面損失 1–3 個數量級對分頻寬 preprint | 11.5 |
| 7 Transport7 傳輸層 | replace替換 | PFC switch support absent. Loss from pointing and flexing. Doppler ≈ 0.14 MHz is negligible. Protocol measurement pendingPFC 交換器支援待建立，損失來自指向與撓曲，約 0.14 MHz 多普勒可忽略，協定待實測 | 11.2, 6.6 |
| 8 Routing8 路由 | adapt調整 | Geometry is a function of orbital phase (lattice) or constant (circular). Precomputed paths replace ECMP hashing幾何是軌道相位的函數（方形格）或常數（圓形）。預先計算的路徑取代 ECMP 雜湊 | 11.5 |
| 9 Front end9 前端 | replace替換 | Radio for the pilot. Optical "eventually". TBIRD 200 Gbps, 4.8 TB per pass. Dawn–dusk visibility試驗用射頻。光學「最終」。TBIRD 200 Gbps、每次過頂 4.8 TB。晨昏軌道可見性 | 11.4, 13.5 |
| 10 Failures10 故障 | adapt調整 | SDC ≈ 1/17 rad, SEFI ≈ 1/5 krad per chip, ~150 rad/yr. 15–20 min power-down on the MVP press. Lost satellites make the graph non-isotropicSDC 約每 17 rad 一次、SEFI 約每晶片每 5 krad 一次、年約 150 rad。MVP 每 15–20 分鐘關機 press。失效衛星使網路圖呈現方向差異 | 8.4, 10.5 |
| 11 Sharing11 共用 | drop省略 | One operator, one job, one tenant單一營運者、單一工作、單一租戶 | — |
Chapter numbers refer to the companion AI Training Network Primer, as in Section 10.3. Numbers in the third column are sourced in the sections listed in the fourth.章號指姊妹教材 AI Training Network Primer，與第 10.3 節相同。第三欄的數字來源見第四欄所列章節。
Chapter summary
本章摘要
Suncatcher specifies the physics of a close optical formation while leaving the cluster network open. Packing, terminal count, switching, range-dependent rate and recovery must be designed together.
Suncatcher 指定了緊密光學編隊的物理條件，叢集網路則仍待設計。堆積、終端數、交換、距離相關速率與恢復機制需一併考量。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 81, 1 km, 650 km, 2027 | Suncatcher’s peer-reviewed design specifies 81 TPU satellites in a square lattice of 1 km radius at 650 km [36]. It uses a dawn–dusk orbit. The flown prototype has ground connectivity only, with a two-satellite optical test planned for 2027. |
| Orbital configuration | 0, -2, 2:1 | The formation rests on the HCW equations, whose bounded solution requires \(\dot y_0 = -2nx_0\) and traces a 2:1 ellipse |
| Orbital configuration | 100, 200 m | so the lattice rotates and stretches, direct neighbors swing between 100 and 200 m and the shape repeats twice per orbit |
| Orbital configuration | 3 m, 2:1.0037, 30 m | Axis-ratio tuning to 2:1.0037 holds J₂ drift below 3 m/s yearly per kilometer. Joule budgets formation maintenance below 30 m/s yearly. |
| Result and scope | 100–200 m | At 100–200 m every link sits deep in the near field |
| Capacity and rate | 24, 9.6–12.8 | so power is abundant and the paper spends it on 24-channel DWDM, 9.6–12.8 Tbps per aperture |
| Result and scope | Qualitative scope | and on spatial multiplexing |
| Result and scope | Qualitative scope | which trades range for beams |
| Capacity and rate | 800, 5 km, 2×2, 1.25 km, 4×4, 0.32 km | The bench reaches 800 Gbps in one direction. Confocal near-field length is about 5 km. Multiplexing ranges are 1.25 km for 2×2 and 0.32 km for 4×4. |
| Latency comparison | Qualitative scope | Everything that would turn a link into a network, the terminal count, switching, topology, hop count, latency budget and bisection, is left open. |
| Result and scope | 367, 100 m, 4, 6 | Circular relative orbits fit 367 satellites in the same envelope with fixed 100 m links. Pénot and Balakrishnan’s Clos mapping requires at least three layers. Switching costs reduce compute fraction to k/(k + 4L − 6). Van Berkel’s laser tori lose orders of bisection bandwidth against terrestrial Clos. |
| Capacity and rate | 2–30 s | Both leave range-dependent rate outside their models, 2–30 s re-pointing or partial blocking. |
| Result and scope | Qualitative scope | Parallelism and step budgets transfer from the companion. Collectives, routing and failures need adaptation. Topology, transport and front ends need replacement designs. Sharing remains outside the single-tenant scope. |
| 軌道配置 | 81, 1 km, 650 km, 2027 | Suncatcher 是一份經同儕審查的設計[36]：81 顆 TPU 衛星排成半徑 1 km 的方形格，位於 650 km 的晨昏軌道，已有一顆僅有地面鏈路的原型機升空，並計畫在 2027 年以兩顆衛星測試鏈路。 |
| 結果與適用範圍 | 0, -2, 2:1, 100, 200 m, 2:1.0037, 3 m, 30 m | 編隊建立在 HCW 方程之上，其有界解要求 \(\dot y_0 = -2nx_0\) 並描出 2:1 橢圓，因此格子會旋轉與拉伸，直接鄰居在 100 到 200 m 之間擺盪，形狀每圈重複兩次，J₂ 則以 2:1.0037 的微調壓在每公里每年 3 m/s 以下，Joule 內文則把編隊維持預算定為每年低於 30 m/s。 |
| 容量與速率 | 100–200 m, 24, 9.6–12.8, 800, 5 km, 2×2, 1.25 km, 4×4, 0.32 km | 在 100–200 m 處每條鏈路都深在近場內，功率充裕，論文把它花在 24 通道 DWDM（每口徑 9.6–12.8 Tbps）與以距離換光束的空間多工上，實驗台結果為單向 800 Gbps，而約 5 km 的共焦近場長度讓 2×2 陣列可用到 1.25 km、4×4 陣列可用到 0.32 km。 |
| 結果與適用範圍 | 定性範圍 | 所有能把鏈路變成網路的項目——終端數、交換、拓樸、跳數、延遲預算與對分頻寬——都待指定。 |
| 結果與適用範圍 | 100 m, 367, 4, 6 | Pénot 與 Balakrishnan 證明圓形相對軌道能在同一包絡內以固定 100 m 鏈路放入 367 顆衛星，Clos 只有在三層以上才放得進去，且要以運算比例 k/(k + 4L − 6) 付出交換衛星。 |
| 結果與適用範圍 | 定性範圍 | van Berkel 則證明以雷射佈線的環面相對地面 Clos 損失數個數量級的對分頻寬。 |
| 結果與適用範圍 | 2–30 s | 兩者的模型範圍都將隨距離變化的速率列在之外、2–30 s 的重新指向或部分遮擋。 |
| 結果與適用範圍 | 定性範圍 | 姊妹教材中，平行化與步時間預算存活，集體通訊、路由與故障處理需要調整，拓樸、傳輸層與前端必須重建，共用則列於單一租戶的範圍之外。 |
Where these ideas go next
這些概念接下來用在哪裡
Chapter 12 examines inference, including van Berkel’s finding that single-satellite inference has the same roofline as terrestrial execution. Published inference systems assume 0.5 Gbit/s to 200 Gbps links, compared with Suncatcher’s Tbps. Chapter 13 develops the open problems in Sections 11.4–11.5. These concern collectives on predictable changing graphs, joint terminal-geometry-fabric design and the ground segment.
第 12 章從訓練轉向推論：van Berkel 指出限制在單顆衛星內的推論與地面有相同的 roofline，這完全改變了網路問題，而已發表的系統所假設的鏈路從 0.5 Gbit/s 到 200 Gbps，Suncatcher 則以 Tbps 為目標。第 13 章把第 11.4 與 11.5 節的待確定項目——可預測時變圖上的集體通訊、終端、幾何與架構的聯合選擇，以及地面段——轉化為研究問題。
Papers behind this chapter本章所依據的論文
Joule 2026Toward a Future Space-Based, Highly Scalable AI Infrastructure System Design
arXiv 2605.15335Designing Dense Satellite Clusters for Distributed Space-based Datacenters
arXiv 2607.14172The Cost and Network Limits of Space-Based AI Compute
ICSOS 2023Recent On-Orbit Results and ARQ Performance Analysis for the TBIRD 200-Gbps Mission