Orbital Edge Computing: Nanosatellite Constellations as a New Class of Computer System軌道邊緣運算：奈米衛星星座作為一類新的電腦系統
Orbital edge computing filters imagery before downlink to reduce ground-contact demand. A formation schedule divides one frame's inference across satellites. The reported infrastructure and latency gains come from the cote simulator, with flight performance still requiring measurement.軌道邊緣運算在下傳前過濾影像，以減少地面接觸需求，並透過編隊排程把一幀影像的推論分配到多顆衛星。基礎設施與延遲增益來自 cote 模擬器，飛行表現則仍待量測。
Orbital edge computing (OEC) runs inference on each nanosatellite and downlinks valuable camera results. A bent pipe instead sends every raw frame to the ground. Every one of these numbers is produced by the authors' cote simulator, with flight performance still pending. The design uses ground links exclusively.Denby 與 Lucia 把軌道邊緣運算（OEC）定義為：在每顆奈米衛星上直接執行推論，只把相機拍到的有價值部分下傳，以減少彎管（bent pipe）模式中每張原始影像皆送回地面的需求。這些數字全部來自作者自己的 cote 模擬器，飛行表現則仍待量測，而且整個設計僅採地面鏈路。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Ground infrastructure | At 200 Mbit/s, one station receives at most 15 GB during a 10-minute pass. An ideally placed station serves 9 satellites per revolution. A 1,000-satellite constellation therefore needs 112 stations. [60, §3.3] |
| Ground infrastructure | On-board filtering retains 0.75 GB per pass. One station then serves 185 satellites per revolution, reducing the constellation requirement to 6 stations. The abstract describes a more than 24-fold infrastructure reduction. Close-spaced computational pipelines reduce edge-processing latency by more than 617×. [60, §7.1–7.2] |
| Model configuration | The hardware model is a Jetson TX2 industrial module in its 7.5 W mode against about 7.1 W of chassis-mounted solar power [60, §3.1, Tab. 1]. |
| 地面設施 | 他們量化的問題是下行鏈路：以 200 Mbit/s 接收的地面站在一次 10 分鐘過頂中最多收到 15 GB，位置理想的地面站每圈只能服務 9 顆衛星，1,000 顆衛星的星座需要 112 座這樣的地面站[60, §3.3]。 |
| 地面設施 | 若機上過濾只留下每次過頂 0.75 GB 的有價值資料，一座地面站每圈可服務 185 顆衛星，同一個星座只需 6 座地面站，摘要把這寫成地面基礎設施減少超過 24 倍，而緊密間隔的運算奈米衛星管線則把邊緣處理延遲降低超過 617 倍（摘要、§7.1–7.2）。 |
| 模型配置 | 硬體模型是以 7.5 W 模式運作的 Jetson TX2 工業模組，對上約 7.1 W 的機身太陽能板[60, §3.1, Tab. 1]。 |
1. Where this paper sits1. 這篇論文的位置
This is the paper that gave the field its name and its first wall. Chapter 8 presents computing in orbit as three successive bottlenecks. The ASPLOS 2020 paper is the one that measures the first of them, the downlink, in a way that a computer architect can reason about. Before it, Earth-observation nanosatellites were treated as cameras with radios.
The architecture question was simply how many ground stations to rent. Denby and Lucia reframed the constellation as a distributed computer whose input is a stream of frames along a ground track. The constellation sends extracted features through a narrow radio downlink. Scheduling assigns each part of each frame to a satellite. Kodan [83] shows that on-board inference shifts the bottleneck to the processor. Xing et al. [85] fly commodity boards and find heat limiting operation before processor or radio capacity.
這是為這個領域命名、也立下第一道牆的論文。第 8 章把在軌運算呈現為三道接連出現的瓶頸，而 ASPLOS 2020 這篇量化的是第一道，也就是下行鏈路，而且用的是電腦架構研究者能推理的方式。在它之前，地球觀測奈米衛星被當成「裝了無線電的相機」，架構問題只是要租幾座地面站。Denby 與 Lucia 把星座重新定義成一台分散式電腦：輸入是沿著地面軌跡連續產生的影像幀，輸出是從狹窄無線電管道流出的少量特徵，排程問題則是哪顆衛星該處理哪一幀的哪一部分。本組另外兩篇後續論文各把這道牆往前推一步：Kodan[83]顯示推論一旦上了機，機上處理器就成為瓶頸。Xing 等人[85]則顯示真正的商用板子一旦飛上去，熱會先於另外兩者成為瓶頸。
The paper also belongs to the bent-pipe lineage that Chapter 3 treats for Starlink user traffic. The shared geometry limits communication to visible stations. Passes last minutes, so satellites store data between contacts. The numbers differ.
This Earth-observation downlink uses 200 Mbit/s X-band. Chapter 3 derives hundreds of Mbit/s per Starlink user channel. Both systems still depend on visibility windows.
這篇論文同時屬於第 3 章為 Starlink 使用者流量所討論的彎管傳統，因為幾何完全相同：衛星只能和看得到的地面站通訊，一次過頂只持續幾分鐘，兩次過頂之間的一切都得存在機上。數字各異，因為本文的地球觀測下行是 X 頻段 200 Mbit/s，而第 3 章逆向工程出的 Starlink 使用者鏈路每個通道有數百 Mbit/s，但關於可見窗口的結構性論證是共通的。
2. The problem, measured2. 問題，用量測說話
The measurement here is a measurement of geometry and link capacity rather than of a running system.這裡的「量測」量的是幾何與鏈路容量，運作中系統的表現則須另行量測。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Orbital period | The representative polar sun-synchronous orbit has 410 km altitude, a 92.9-minute period and a two-day revisit. The study estimates ground-station retrieval during 5–10 minute passes. [60, §3.3] |
| Evaluation conditions | At 200 Mbit/s a 10-minute session yields 15 GB. A camera continuously captures 4,096 × 3,072-pixel frames along the ground track. One pass returns less than one revolution of imagery. [60, §3.1, §3.3] |
| Ground infrastructure | One ideally positioned station serves 9 satellites per revolution. A 1,000-satellite fleet requires 112 stations. Growing the fleet from 250 to 1,000 increases station demand by 3.5×. [60, §3.3, §7.1] |
| 星座配置 | 作者取一顆代表性奈米衛星的軌道，也就是 410 km、週期 92.9 分鐘、兩天重訪一次的極地太陽同步軌道，然後問地面站在每次 5 到 10 分鐘的過頂中能下載多少資料[60, §3.3]。 |
| 評估條件 | 以 200 Mbit/s 計，一次 10 分鐘的通聯可得 15 GB，而這還低於一台沿整條地面軌跡以 4,096 × 3,072 像素拍攝的相機一圈所產生的影像量[60, §3.1, §3.3]。 |
| 地面設施 | 後果以地面站數量計算：位置理想的地面站每圈服務 9 顆衛星，支撐 1,000 顆衛星的星座需要 112 座這樣的地面站，而且星座從 250 顆增加到 1,000 顆時地面站數量成長 3.5 倍[60, §3.3, §7.1]。 |
The paper adds two operational facts that make the arithmetic concrete. Existing bent-pipe systems were seen to take 5.5 hours before data reached customers, and one nanosatellite mission reported 88% packet loss on its downlink [60, §1, §3.3] . Reconfiguring a bent-pipe mission, because it requires new ground-side processing and scheduling, is described as taking hours to months [60, §7.4, Fig. 9] . A CubeSat costs about $65,000. A monolithic Earth-observation satellite costs about $650 million, a 10,000-fold difference per device. Thus, the constellation is cheap to build and expensive to listen to [60, §1] .
論文另外補上兩個讓算術具體化的操作事實。既有的彎管系統被觀察到要 5.5 小時資料才送到客戶手上，而某個奈米衛星任務的下行封包遺失率高達 88%[60, §1, §3.3]。彎管任務的重新配置因為需要新的地面端處理與排程，被描述為耗時數小時到數月[60, §7.4, Fig. 9]。經濟面的框架是：一顆 CubeSat 約 $65,000，對比一顆單體地球觀測衛星約 $6.5 億，每台裝置差一萬倍，所以星座蓋起來便宜，聽它說話卻很貴[60, §1]。
3. The design3. 設計
The design has two layers. Each nanosatellite runs inference locally and downlinks features. The DetectNet building-footprint detector trains on SpaceNet at 0.3 m per pixel. A computational nanosatellite pipeline (CNP) divides work among satellites sharing a ground track. [60, §4.1, §6] Two choices define a CNP [60, §4.2, Fig. 5] .
Frame-spaced satellites remain one frame apart, each capturing and processing a whole frame. Close-spaced satellites remain close enough to observe the same frame. The partition choice is frame-parallel, where each satellite owns whole frames, or tile-parallel, where each satellite processes one tile of a shared frame. Close-spaced tile-parallel pipelines let a frame be finished within one pass of the group rather than queued behind a single satellite's compute. This is where the latency reduction comes from.
設計分兩層。第一層是逐衛星推論：每顆奈米衛星對自己拍到的影像幀執行偵測網路，下傳特徵以取代原始像素，論文用在 SpaceNet 資料集（每像素 0.3 m）上訓練的 DetectNet 建築輪廓偵測器來建模[60, §4.1, §6]。第二層是運算奈米衛星管線（CNP），讓沿同一條地面軌跡飛行的多顆衛星分工。兩個選擇定義了一條 CNP[60, §4.2, Fig. 5]。
間隔的選擇是「幀間隔」（frame-spaced），相鄰衛星相距一幀，每顆衛星拍攝並處理一整幀。或「緊密間隔」（close-spaced），衛星靠得很近，全都看到同一幀。切分的選擇是「幀平行」（frame-parallel），每顆衛星負責完整的幀。或「區塊平行」（tile-parallel），每顆衛星處理共享幀中的一個區塊。
緊密間隔加區塊平行的管線讓一幀在整組衛星的一次過頂內就處理完，縮短在單顆衛星處等待運算的時間，延遲的降低就來自這裡。
Coordination is deliberately minimal. The cote-lib runtime derives local decisions from orbit and schedule. It uses only local formation knowledge. [60, §4–5] The pipeline coordinates through time and position. Each satellite identifies its tile from its formation position.
Inter-satellite message exchange lies outside the design. Because detection draws more than the panels supply, the satellite computes in bursts and sleeps between them. Thus, coverage in this paper is energy-limited rather than compute-limited.
協調刻意極簡。執行期 cote-lib 從軌道與排程在本地做決定，論文明確說它僅依本地的編隊資訊運作[60, §4–5]。這對網路讀者很重要，因為這表示管線是一種「時間與位置」協定，通訊需求由既有編隊排程取代：每顆衛星知道哪個區塊是自己的，是因為它知道自己在編隊中的位置，星間資料交換則位於設計範圍之外。因為偵測耗電超過太陽能板供電，衛星只能間歇運算、其間休眠，這也是為什麼本文的覆蓋主要受能源供應限制。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Energy | A discrete-time energy model runs the Jetson TX2 at 7.5 W. Surface-mounted panels supply about 7.1 W, while deployable panels supply 14.2 W. |
| Jetson detection power | with Table 1 listing 11.3 W average for Jetson detection |
| Sleep power | 0.5 W in sleep |
| Imaging power | 3.5 W for camera imaging |
| Readout power | 2.5 W for readout |
| Attitude-control power | and 1.13 W for attitude control |
| Capacitor configuration | buffered by 1.0 F capacitors in a 5.0 F baseline bank [60, §3.1, §6, Tab. 1]. |
| 模型配置 | 能源面用離散時間模型處理：Jetson TX2 以 7.5 W 模式運作，對上機身板約 7.1 W 或可展開板 14.2 W 的供電 |
| 偵測功率 | Table 1 列出 Jetson 偵測平均 11.3 W、 |
| 休眠功率 | 休眠 0.5 W、 |
| 成像功率 | 相機成像 3.5 W、 |
| 讀出功率 | 讀出 2.5 W、 |
| 姿態控制功率 | 姿態控制 1.13 W |
| 電容配置 | 由 1.0 F 電容組成的 5.0 F 基準電容組做緩衝[60, §3.1, §6, Tab. 1]。 |
Original figure note原始圖說出處
Drawn for this primer from Denby & Lucia, ASPLOS 2020, §3.3, §4.2 and Fig. 5.4. Evaluation setup4. 實驗設定
Everything is simulation. The model uses ground links only. ISL rate, latency and topology lie outside its scope. Latency is counted as edge-processing latency, the time from capture to a finished inference for a frame. This is the quantity reduced 617×. The absolute minutes-to-customer value remains unreported for OEC.一切都是模擬。模型僅採地面鏈路，ISL 速率、延遲及拓樸皆位於範圍之外。延遲的定義是邊緣處理延遲，即從拍攝到該幀推論完成的時間，這就是被降低 617 倍的量，而論文將其保留為待補： OEC 從拍攝到送達客戶的絕對分鐘數。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Energy | The cote simulator combines SGP4 orbits, link budgets and discrete-time energy. Planet TLEs define 250- and 1,000-satellite constellations at 97.3° inclination. Ground-station rings lie at 87° N and S. [60, §6] |
| Model configuration | The downlink uses 8.15 GHz and 20 MHz bandwidth. Satellite patch-antenna gain is 6.0 dB, and ground-dish gain is 44.1 dB. This yields the 200 Mbit/s figure used throughout [60, §6] . |
| Evaluation conditions | The camera uses 1.1 µm pixels and 20 cm focal length. Its ground sampling is 2.26 m per pixel, 7.5× coarser than a monolithic satellite. [60, §3.1] |
| 鏈路建立 | 作者建立了 cote 模擬器，以 SGP4 傳播軌道、評估鏈路預算、以離散時間步進能源模型，並餵入由 Planet 的 TLE 推導、傾角 97.3° 的 250 顆與 1,000 顆衛星星座，地面站環設在南北緯 87°[60, §6]。 |
| 頻率 | 下行鏈路建模為 8.15 GHz、20 MHz 頻寬、衛星端 6.0 dB 貼片天線與地面 44.1 dB 碟形天線，由此得到全文使用的 200 Mbit/s[60, §6]。 |
| 星座配置 | 相機像素 1.1 µm，在 20 cm 焦距下地面取樣距離為每像素 2.26 m，論文自己也指出這比單體衛星粗 7.5 倍[60, §3.1]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Ground stations for 1,000 satellites fall from 112 to 6, reported as a >24× reduction in ground infrastructure [60, §7.1].1,000 顆衛星所需地面站從 112 座降到 6 座，摘要寫成地面基礎設施減少 >24 倍（摘要、§7.1）。 | cote simulation with 200 Mbit/s stations and 10-minute passes. The ratio of the two station counts is about 18.7 and of the per-station satellite counts about 20.6 (calc table below). Thus, the 24× headline is quoted from the abstract rather than reproduced.cote 模擬，200 Mbit/s 地面站與 10 分鐘過頂。兩個地面站數量之比約 18.7、每站衛星數之比約 20.6（見下方計算表），所以 24 倍是直接引用摘要，其權重仍須進一步資料才能重現。 |
| Edge-processing latency reduced by more than 617× with close-spaced pipelines [60, §7.2].緊密間隔管線使邊緣處理延遲降低超過 617 倍（摘要、§7.2）。 | Relative to sequential single-satellite processing in simulation. Absolute latency remains unreported.相對於模擬中的單衛星依序處理。報告資料待補絕對延遲。 |
| Full ground-track coverage needs about 250 satellites with surface panels or about 100 with deployable panels [60, §7.3].完整地面軌跡覆蓋需約 250 顆（機身板）或約 100 顆（可展開板）衛星[60, §7.3]。 | Driven by the energy model (7.1 vs 14.2 W of solar input). Thus, the count is an energy result, with compute capacity held fixed.由能源模型（7.1 對 14.2 W 太陽能輸入）決定，所以這是能源條件的結果，運算容量則維持固定。 |
| After on-board discard, 5–25% of captured data is downlinked [60, Fig. 6].機上丟棄後只下傳 5–25% 的拍攝資料[60, Fig. 6]。 | The 0.75 GB of 15 GB example corresponds to the 5% end of this range.15 GB 中留 0.75 GB 的例子對應此範圍的 5% 一端。 |
| Bent-pipe station count grows 3.5× from 250 to 1,000 satellites [60, §7.1].彎管模式地面站數在 250 到 1,000 顆衛星間成長 3.5 倍[60, §7.1]。 | Sub-linear because stations at 87° N/S see every polar orbit. The OEC count stays small in both cases.因為南北緯 87° 的地面站能看到每一條極軌道，所以成長低於線性。OEC 的站數在兩種情況下都很少。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| How long does a 15 GB pass take at 200 Mbit/s?200 Mbit/s 下傳 15 GB 要多久？ | \( t = \dfrac{15 \times 8 \times 10^{9}}{2 \times 10^{8}} \) | 600 s = 10 mincomputed. 15 GB and 200 Mbit/s from §3.3計算值。15 GB 與 200 Mbit/s 來自 §3.3 |
| How long does the 0.75 GB OEC downlink take?OEC 的 0.75 GB 下傳要多久？ | \( t = \dfrac{0.75 \times 8 \times 10^{9}}{2 \times 10^{8}} \) | 30 scomputed. Matches the paper's "about 30 s" [60, §3.3]計算值。與論文的「約 30 秒」一致[60, §3.3] |
| What reduction do the stated station counts imply?所述地面站數隱含多少倍的減少？ | \( \dfrac{112}{6} \approx 18.7,\quad \dfrac{185}{9} \approx 20.6 \) | ≈ 19–21×computed from §3.3 counts. The abstract's >24× requires weighting beyond these two ratios由 §3.3 的數字計算。摘要的 >24 倍須加入這兩個比值以外的權重才能重現 |
| How far short of detection power is the solar input?太陽能供電比偵測功耗差多少？ | \( \dfrac{7.1}{11.3} \approx 0.63 \) | ≈ 63% duty limitcomputed. 7.1 W solar [60, §3.1] and 11.3 W Jetson detection [60, Tab. 1]. Ignores camera, ADACS and conversion losses計算值。7.1 W 太陽能[60, §3.1]與 11.3 W Jetson 偵測[60, Tab. 1]。忽略相機、姿態控制與轉換損耗 |
6. Limits and open questions6. 限制與待解的問題
- Simulation evidence with flight validation still pending. The 24× and 617× figures come from cote with a modeled Jetson TX2. OEC pipeline flight validation was pending at publication. Chapter 8 records Xing et al. using Raspberry Pi 4B and Huawei Atlas 200 DK in orbit. That COTS study supplies separate hardware evidence. [85]證據來自模擬，飛行驗證仍待完成。24 倍與 617 倍來自以模型化 Jetson TX2 跑的 cote。論文發表時OEC 管線的飛行驗證仍待完成，而第 8 章記載唯一的在軌 COTS 量測 Xing 等人[85]飛的是 Raspberry Pi 4B 與華為 Atlas 200 DK，其硬體為 Pi 與 Atlas 平台。
- The headline ratio requires weighting beyond the printed counts. Reducing 112 stations to 6 gives 18.7×. Increasing service from 9 to 185 satellites gives 20.6×. Thus, the abstract's "more than 24×" must rest on a cost or infrastructure weighting that remain unspecified in the text.標題倍數須加入印出數字以外的權重。112 到 6 座是 18.7 倍、每站 9 到 185 顆是 20.6 倍（皆為上方計算值），所以摘要的「超過 24 倍」必然依賴內文仍待揭露的成本或基礎設施權重。
- Ground links with local formation coordination. cote-lib coordinates using local formation knowledge. Thus, the paper leaves ISL rate, setup delay and topology outside its scope. The "pipeline" is a formation-flying schedule rather than a network. The later Tiansuan design in Chapter 8 adds ISLs only from its fourth satellite onward.設計僅採地面鏈路。cote-lib 以本地編隊資訊協調，所以論文對 ISL 速率、建鏈延遲與拓樸皆位於模型範圍之外，「管線」是編隊飛行排程，星間通訊網路則位於範圍之外。第 8 章提到的後續天算星座設計要到第四顆衛星才加入 ISL。
- Radiation is deferred. The Jetson TX2 is used as a power model only. A NASA GSFC proton test later reports single-event functional interrupts on every run of the same module. Each requires a power cycle (Wyrwas 2019, agency report). An OEC runtime must tolerate these interruptions.輻射被延後處理。Jetson TX2 只被當成功耗模型。NASA GSFC 後來對同一模組做的質子測試回報每一輪都出現單事件功能中斷，每次都需重新上電（Wyrwas 2019，機構報告），OEC 執行期必須能撐過這種情況。
- Fixed downlink and pass length. 200 Mbit/s and 10-minute passes are the only operating point. Chapter 3 shows 200 Gbps-class optical downlinks have since flown on TBIRD. This would move the wall by three orders of magnitude and change the conclusion about where to compute.固定的下行速率與過頂長度。200 Mbit/s 與 10 分鐘過頂是唯一的工作點。第 3 章顯示 TBIRD 已飛過 200 Gbps 等級的光學下行，那會把這道牆推開三個數量級，並改變「該在哪裡運算」的結論。
- Resolution is traded away. The 2.26 m/px camera is 7.5× coarser than a monolithic satellite by the paper's own account. Thus, the station savings are partly purchased with less data per frame. The paper leaves energy per image and Planet data rates unspecified to close that comparison.解析度被犧牲。依論文自己所述，2.26 m/px 的相機比單體衛星粗 7.5 倍，所以地面站的節省有一部分是用每幀較少資料換來的。論文每張影像能耗與 Planet 資料速率皆仍待補充可供封閉這項比較。
- Energy supply sets coverage under the evaluated compute capacity. Solar input is 7.1 W, while detection draws 11.3 W. Detection duty is therefore at most about 63%, before camera and attitude-control demand. Thus, the 250- and 100-satellite coverage counts are statements about panels, with processor capacity held fixed.在既定運算容量下，能源供應決定覆蓋。7.1 W 太陽能對上 11.3 W 偵測，即使將相機及姿態控制列於範圍之外，衛星最多只能約 63% 的時間執行偵測，所以 250 顆與 100 顆的覆蓋數字說的是太陽能板，處理器容量則維持固定。
7. How it connects7. 與其他內容的關聯
- Section 8.1 opens with the downlink bottleneck and 112→6 station comparison. Figure 8.1 redraws the pipeline.第 8.1 節：下行瓶頸與 112→6 座地面站的結果是該章的開場論證，管線重繪為圖 8.1。
- Section 8.2 compares 7.1 W solar input with 11.3 W compute demand. Panel 8.1 scales this power problem to accelerator-class payloads.第 8.2 節：7.1 W 對 11.3 W 的預算是電力問題的最小實例，面板 8.1 讓你把它放大到加速器等級的酬載。
- Section 8.4: the Jetson TX2 proton results belong to the radiation table there.第 8.4 節：Jetson TX2 的質子測試結果列在該節的輻射表中。
- Section 8.6: what has actually flown, against which this simulated design should be read.第 8.6 節：真正飛過的系統，用來對照這個模擬設計。
- Section 3.6: the 200 Mbit/s station and 15 GB per pass sit beside L2D2's distributed-downlink numbers.第 3.6 節：200 Mbit/s 地面站與每次過頂 15 GB 與 L2D2 的分散式下行數字並列。
- Kodan [83] moves the bottleneck from downlink to on-board compute. Xing et al. [85] fly real boards and find heat. Bhattacherjee et al. [2] ask the same per-satellite compute question for a 15.6 kg server. L2D2 [61] attacks the same downlink wall from the ground side.Kodan[83]把瓶頸從下行移到機上運算。Xing 等人[85]飛了真實板子並發現熱是問題。Bhattacherjee 等人[2]對一台 15.6 kg 伺服器問同樣的逐衛星運算問題。L2D2[61]從地面端攻擊同一道下行牆。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 200 Mbit/s · 15 GB | Ground-station downlink rate and data per 10-minute pass [60, §3.3] | 地面站下行速率與每次 10 分鐘過頂的資料量[60, §3.3] | |
| 9 → 185 | Satellites one ideal station serves per revolution, bent pipe vs OEC [60, §3.3] | 一座理想地面站每圈服務的衛星數，彎管對 OEC[60, §3.3] | |
| 112 → 6 | Stations for 1,000 satellites, bent pipe vs OEC. Abstract reports >24× [60, §7.1] | 1,000 顆衛星所需地面站，彎管對 OEC。摘要寫 >24 倍[60, §7.1] | |
| >617× | Edge-processing latency reduction for close-spaced pipelines, relative, simulated [60, §7.2] | 緊密間隔管線的邊緣處理延遲降幅，相對值、模擬[60, §7.2] | |
| 7.5 W vs 7.1 W | Jetson TX2 power mode vs chassis-mounted solar. 14.2 W deployable [60, §3.1] | Jetson TX2 功耗模式對機身太陽能板。可展開板 14.2 W[60, §3.1] | |
| 11.3 W · 0.5 W | Jetson detection average and sleep power [60, Tab. 1] | Jetson 偵測平均功耗與休眠功耗[60, Tab. 1] | |
| ≈250 / ≈100 | Satellites for full ground-track coverage, surface vs deployable panels [60, §7.3] | 完整地面軌跡覆蓋所需衛星數，機身板對可展開板[60, §7.3] | |
| 410 km · 92.9 min | Example polar sun-synchronous orbit and period. Two-day revisit [60, §3.3] | 範例極地太陽同步軌道與週期。兩天重訪[60, §3.3] | |
| $65,000 vs $650M | CubeSat vs monolithic satellite cost framing [60, §1] | CubeSat 對單體衛星的成本框架[60, §1] |