Digest · Computing in orbit · read after Chapter 8
導讀 · 在軌運算 · 第 8 章後閱讀
Deciphering the Enigma of Satellite Computing with COTS Devices: Measurement and Analysis解開 COTS 裝置衛星運算之謎：量測與分析
This digest explains the conductive heat path that limits commercial boards in orbit and how thermal measurements set their duty cycle. It also defines the evidence scope of the campaign’s zero single-event error count.讀完後，你能說明為什麼一塊 5 W 的單板電腦在被冰冷太空包圍的衛星裡會過熱、讀懂決定商用運算在軌工作週期的兩條熱曲線，並說出六個月運算錯誤計數為零的證據範圍。
Main result
主要結果
BUPT-1’s commercial boards show how conductive and radiative heat paths limit sustained orbital compute. Their energy and zero-error observations retain the experiment’s hardware and orbital scope.
BUPT-1 的商用板卡顯示，傳導與輻射散熱路徑會限制持續在軌運算，能源及零錯誤觀察也應依實驗硬體與軌道範圍解讀。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 4B · 200 · 1 · 12 · 17.44 kg · 487.6–494.7 km · 6,000 · 1,000 | This is the only peer-reviewed measurement of commodity computers operating in orbit. BUPT-1 carries two Raspberry Pi 4B and two Huawei Atlas 200 DK boards. The 12U satellite weighs 17.44 kg and follows a 487.6–494.7 km sun-synchronous orbit. Evidence spans over 6,000 orbits and 1,000 experimental hours [85, §2, Tabs. 1–2]. |
| Thermal throttling | 8 · 75% · 80 · 50 · 82 · 15 · 30 · 540 · 40 | Above 75% load, the Pi crosses its 80 °C throttling threshold within about 50 minutes. Its chip saturates near 82 °C, while the board surface rarely exceeds 15 °C. One Atlas reaches the satellite's 30 °C surface limit after 540 minutes. Two Atlases reach it in about 40 minutes. |
| Power and thermal conditions | 2–4 | Heating is 2–4× slower than on the ground replica. Heat conducts into the small satellite body, which radiates it away [85, §3, Tab. 3, Figs. 3–6]. |
| Power and thermal conditions | 40 W · 16 · 50% · 30% | Peak solar power is about 40 W, yielding about 16 Wh per orbit. Battery depth of discharge briefly exceeds 50%, against a 30% target [85, §4, Fig. 13]. |
| Evidence and scope | Qualitative scope | The six-month campaign recorded zero computational errors from single-event effects on these standard commercial boards [85, §6]. |
| Latency comparison | 2–10% · 5% · 10% | Orbital task energy is 2–10% below the ground result. Throttling reduces frequency by about 5% and increases YOLO latency by up to 10%. The recommendations emphasize duty cycling, which larger Part III proposals still need to address. |
| 軌道配置 | 4B · 200 · 1 · 12 · 17.44 kg · 487.6–494.7 km · 6,000 · 1,000 | 這是唯一經同儕審查、對在軌運作之商用電腦的量測，飛的是兩塊 Raspberry Pi 4B 與兩塊華為 Atlas 200 DK，板卡型號為上述兩種，搭載於 BUPT-1，一顆 12U、17.44 kg、位於 487.6–494.7 km 太陽同步軌道的衛星，進行了超過 6,000 圈與 1,000 小時的實驗[85, §2, Tabs. 1–2]。 |
| 證據與範圍 | 8 · 75% · 50 · 80 · 82 · 15 | 組織第 8 章的發現是熱：負載超過 75% 時，Pi 的晶片約 50 分鐘內進入 80 °C 以上的降頻區並在接近 82 °C 飽和，而板子表面很少超過 15 °C。 |
| 證據與範圍 | 540 · 30 · 40 | 單一 Atlas 在 540 分鐘後把衛星表面推到 30 °C 的設計上限，兩塊 Atlas 一起則約 40 分鐘就到。 |
| 證據與範圍 | 2–4 | 升溫比地面複製品慢 2–4 倍，因為熱唯一的出路是傳導進一個會輻射的小機身[85, §3, Tab. 3, Figs. 3–6]。 |
| 電力與熱條件 | 40 W · 16 · 50% · 30% | 能源是第二個限制：太陽能峰值約 40 W、每圈收穫約 16 Wh，電池放電深度曾短暫超過 50%，而目標是 30%[85, §4, Fig. 13]。 |
| 證據與範圍 | 定性範圍 | 這些標準商用板卡在六個月內記錄到的單事件運算錯誤為零[85, §6]。 |
| 證據與範圍 | 2–10% · 5% · 10% | 板子在軌每項任務的能耗比地面少 2–10%，降頻的代價約 5% 的頻率與最多 10% 的 YOLO 延遲，而作者的建議歸結為一種工作週期紀律，第三部分那些更大的提案至今仍待採納。 |
1. Where this paper sits1. 這篇論文的位置
This paper adds flight measurements to earlier compute and power budgets. It shows why thermal behavior must be included even for low-power boards.
本篇在早期運算及電力預算之外補上飛行量測，說明低功率板卡也需要納入熱行為。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | 8 | Chapter 8 orders the literature on computing in orbit as three walls. This paper is the third and the only one built from flight data. |
| Evidence and scope | Qualitative scope | Denby and Lucia [60] simulated the downlink wall, Kodan [83] emulated the compute wall on ground hardware, and Bhattacherjee et al. |
| Power and thermal conditions | 350 W · 1.5 kW | [2] budgeted a 350 W server against an estimated 1.5 kW of solar power with thermal budgeting left open. |
| Power and thermal conditions | 5–10 W | The real-board orbital measurements identify thermal transfer as the binding constraint. This occurs at only 5–10 W. |
| Evidence and scope | 8 · 8.6 · 2023 · 1 · 100 | Section 8.6 describes the Tiansuan program. Wang and Li (IEEE IoT Journal 2023) specify the BUPT-1 hardware used here. Its ground-only radio provides 100 Mbps X-band downlink and 1 Mbps uplink. |
| 證據與範圍 | 8 | 第 8 章把在軌運算的文獻排成三道牆，這篇是第三道，也是唯一由飛行資料築成的。 |
| 電力與熱條件 | 1.5 kW · 350 W | Denby 與 Lucia[60]模擬了下行牆，Kodan[83]在地面硬體上模擬了運算牆，Bhattacherjee 等人[2]則在估計的 1.5 kW 太陽能下為一台 350 W 伺服器編列預算，熱預算則待補充。 |
| 電力與熱條件 | 5–10 W | Xing 等人把真的板子放上真的衛星，發現熱才是束縛的那一項，而且規模是 5–10 W，數百瓦酬載則是另一種規模。 |
| 證據與範圍 | 8 · 8.6 · 2023 · 1 · 100 | 這篇論文也是第 8 章第 8.6 節所描述天算星座計畫的實測端點，Wang 與 Li（IEEE IoT Journal 2023）在那裡列出 BUPT-1 的硬體等級：100 Mbps X 頻段下行、1 Mbps 上行、僅有地面鏈路，正是本文使用的平台。 |
The measurements calibrate Part III assumptions about continuous operation. Even single-board workloads require explicit thermal design and duty cycles. Press reports describe 15–20 minute Suncatcher prototype operation, broadly comparable with the 50-minute Pi curve. Google documentation still needs to confirm that reported duty cycle.
對網路讀者而言，它的價值在於作為校準點。第三部分的每個設計都假設只要太陽能陣列供電，酬載就能持續運作。這篇是唯一說「即使在單板規模，持續運作仍需明確的熱設計」而且工作週期必須被設計進去的量測。事實庫指出媒體對 Google Suncatcher 原型的報導提到 15–20 分鐘的工作週期，與本文 50 分鐘的 Pi 曲線同一量級，該數字的 Google 文件確認仍待補充。
2. The problem, measured2. 問題，用量測說話
The experiment compares commercial boards in BUPT-1 with a matching ground thermal structure. Controlled compute and idle periods provide the basis for temperature and energy comparisons.
實驗將 BUPT-1 內的商用板卡與相同熱結構的地面對照比較，受控運算及閒置時段構成溫度與能源比較的基礎。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | Qualitative scope | The problem is that in-orbit measurements of standard commercial boards were still needed. Thus, every design argument about commodity compute in space rested on data sheets and ground tests. |
| Evidence and scope | 1 · 12 · 17.44 kg · 115 · −10 · 30 · 100 · 15 · 2023 | BUPT-1 is a 12U, 17.44 kg satellite launched on 15 January 2023. It carries two 115 Wh batteries. Its internal design range is −10 to 30 °C. Radio rates are 100 Mbps downlink and 1 Mbps uplink [85, Tab. 1, §2.1, §2.3]. |
| Power and thermal conditions | 4 · 14 · 14 mm · 200 · 310 | Two Raspberry Pi 4Bs use quad-core Cortex-A72 processors on 14 × 14 mm dies. Two Atlas 200 DKs use Ascend 310 NPUs with four power levels [85, §2.2, §3.1]. |
| Evidence and scope | 22 · 30 · 1–4 s | After about two months of stability tests, twelve experiments ran from 22 March to 30 July. Each lasted five hours: three computing, then two idle. The compute phase spans roughly two orbits. Telemetry arrives every 1–4 s. A matched ground replica provides the thermal control [85, §2.3, §3.1, Tab. 2]. |
| Evidence and scope | 10 · 22 GB · 30 GB | The campaign produced more than 10 million telemetry lines, 22 GB of telemetry and 30 GB of payload data. The authors state it took two years and more than a million dollars to prepare. |
| 證據與範圍 | 定性範圍 | 問題在於標準商用板卡的在軌行為仍待量測，所以關於太空商用運算的每個設計論證都建立在規格表與地面測試上。 |
| 證據與範圍 | 1 · 12 · 17.44 kg · 115 · −10 · 30 · 100 · 2023 · 15 | 作者的平台是 BUPT-1：12U、17.44 kg、兩顆 115 Wh 電池、內部設計範圍 −10 到 30 °C、100 Mbps 下行與 1 Mbps 上行，於 2023 年 1 月 15 日發射[85, Tab. 1, §2.1, §2.3]。 |
| 證據與範圍 | 4B · 14 · 14 mm · 200 · 310 | 酬載是兩塊 Raspberry Pi 4B（四核 Cortex-A72，14 × 14 mm 晶粒）與兩塊 Atlas 200 DK（Ascend 310 NPU，四個功率等級）[85, §2.2, §3.1]。 |
| 證據與範圍 | 3 · 22 · 7 · 30 · 1–4 | 經過約兩個月的穩定性測試後，十二個實驗在 3 月 22 日到 7 月 30 日之間進行，每個五小時，含三小時運算（約跨兩圈）與兩小時閒置，遙測每 1–4 秒一筆，並以一比一的衛星熱結構地面複製品作為對照[85, §2.3, §3.1, Tab. 2]。 |
| 證據與範圍 | 1,000 · 22 GB · 30 GB | 整個活動產生超過 1,000 萬行遙測、22 GB 遙測與 30 GB 酬載資料，作者表示準備耗時兩年、花費超過一百萬美元。 |
The workloads are deliberately ordinary. The Pi ran stress tests on one to four cores and the SSD-MobileNetV1, YOLO-Fastest, YOLOv3 and YOLOv5-Lite detectors. The Atlas ran JPEG encoding, ResNet-50 and YOLOv3 at one or four threads across its four power levels [85, Tab. 2]. The quantity that turned out to matter is the relevant quantities are the chip and surface temperatures they produced as a function of time. The energy they drew from a battery that the Sun refills only about 16 Wh per orbit.
工作負載刻意平凡。Pi 跑一到四核的壓力測試與 SSD-MobileNetV1、YOLO-Fastest、YOLOv3、YOLOv5-Lite 偵測器。Atlas 跑 JPEG 編碼、ResNet-50 與 YOLOv3，以一或四執行緒跨四個功率等級[85, Tab. 2]。最後證明關注的量是它們隨時間產生的晶片與表面溫度，以及它們從一顆太陽每圈只補回約 16 Wh 的電池中取走的能量。
3. The design3. 設計
The Pi’s chip-to-board path limits cooling, while the Atlas transfers more heat into the satellite body. Their operating limits therefore occur at separate parts of the heat path.
Pi 的晶片至板卡熱路徑限制了冷卻，Atlas 則將較多熱傳入衛星本體，因此兩者的操作上限出現在熱路徑的各個部位。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | This is a measurement paper, so its design is the experiment and the interpretation. |
| Power and thermal conditions | Qualitative scope | The central interpretation is a heat-path argument. |
| Capacity and rate | Qualitative scope | In vacuum, heat leaves through conduction and radiation. Thus, the heat a chip generates must conduct through its package and mount into the satellite's structure and then radiate from the satellite's outer surface. The rate of that last step is fixed by the surface area and temperature. |
| Heat transfer | 14 · 14 mm · 82 · 15 | The Pi has a 14 × 14 mm die and a mount designed for airflow. Its chip reaches about 82 °C, while its mounting surface rarely exceeds 15 °C. Chip-to-board heat transfer is the bottleneck, with little body warming [85, §3.1, Fig. 3]. |
| Power and thermal conditions | 52.6 · 38.5 mm · 30 | The Atlas module measures 52.6 × 38.5 mm and has better conduction. It keeps its chip cooler while heating the structure. Surface temperature rises over hours toward the 30 °C limit [85, §3.2, Fig. 6]. |
| Reliability and accuracy | Qualitative scope | The two boards therefore fail in opposite ways: the Pi throttles itself and the Atlas threatens the satellite. |
| 證據與範圍 | 定性範圍 | 這是量測論文，所以它的設計就是實驗與詮釋。 |
| 證據與範圍 | 定性範圍 | 核心詮釋是熱路徑論證。 |
| 證據與範圍 | 定性範圍 | 真空中的散熱依靠傳導與輻射，晶片產生的熱必須經封裝與固定座傳導進衛星結構，再從衛星外表面輻射出去，而最後這一步的速率由表面積與溫度決定。 |
| 證據與範圍 | 14 · 14 mm · 82 · 15 | Raspberry Pi 的晶粒只有 14 × 14 mm，固定座是為氣流設計的，所以晶片升到約 82 °C 時，它所鎖定的表面很少超過 15 °C：瓶頸在晶片與板子之間，衛星本體的溫度變化較小[85, §3.1, Fig. 3]。 |
| 證據與範圍 | 52.6 · 38.5 mm · 30 | Atlas 的模組較大（52.6 × 38.5 mm）、傳導路徑較好，自己的晶片較涼，卻把熱推進結構，所以表面溫度數小時內緩慢爬向 30 °C 的設計上限[85, §3.2, Fig. 6]。 |
| 證據與範圍 | 定性範圍 | 兩塊板子因此以相反的方式失敗：Pi 讓自己降頻，Atlas 威脅衛星。 |
A colder initial structure slows in-orbit heating and lowers device energy use. The recommendations combine task duration, mount conductivity, thermal sharing and battery-state scheduling.
較冷的初始結構使在軌升溫較慢，也降低裝置能耗。建議措施結合任務時長、固定座導熱、熱表面分工與電池狀態排程。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Evidence and scope | Qualitative scope | The second design element is the comparison to the ground replica, which shows the counter-intuitive direction of the effect. |
| Evidence and scope | 39 | On the ground, with the same structure in air, the Pi saturates in 39 minutes. |
| Power and thermal conditions | 86 · 147 · 88 | The Pi reaches its orbital temperature limit in 86 minutes. Full-power four-thread Atlas heating takes 147 minutes in orbit, compared with 88 on the ground [85, Tab. 3]. |
| Power and thermal conditions | 0–4 · 90% · 5 · 8 | Orbital heating starts colder: idle surfaces remain at 0–4 °C for 90% of a two-day window. Daylight and eclipse vary the surface by about ±5 °C. An Atlas run starting in daylight ends about 8 °C hotter at the chip than one starting in eclipse [85, §3.4, Fig. 2]. |
| Power and thermal conditions | 4.82 W · 5.92 W · 10.19 W · 11.48 W · 2–10% | Cold starts also lower orbital power. The four-core Pi uses 4.82 W versus 5.92 W on the ground. The full-power Atlas uses 10.19 W versus 11.48 W. Task energy is 2–10% lower [85, §4, Tabs. 3–5]. |
| Evidence and scope | Qualitative scope | The third element is the recommendations: run short, intermittent tasks. |
| Power and thermal conditions | Qualitative scope | Give low-power boards higher-conductivity mounts. |
| Power and thermal conditions | Qualitative scope | Cap the load and duration of high-power boards. |
| Evidence and scope | Qualitative scope | Run one device per shared thermal surface at a time. |
| Power and thermal conditions | 30% | Schedule compute to absorb surplus solar energy and hold depth of discharge at or below 30%. |
| Power and thermal conditions | Qualitative scope | And put thermal and energy models into hardware-in-the-loop simulators. |
| 證據與範圍 | 定性範圍 | 第二個設計要素是與地面複製品的比較，它顯示效應的方向出乎直覺。 |
| 證據與範圍 | 39 | 在地面、同樣的結構放在空氣中，Pi 39 分鐘飽和。 |
| 證據與範圍 | 86 · 147 · 88 | 在軌要 86 分鐘，而四執行緒滿功率的 Atlas 在軌要 147 分鐘，地面是 88 分鐘[85, Tab. 3]。 |
| 證據與範圍 | 90% · 0–4 · 5 · 8 | 軌道上升溫較慢，因為結構起點較冷，閒置表面在兩天窗口中 90% 的時間為 0–4 °C，而日蝕與日照只讓表面擺動約 ±5 °C，日照中開始的 Atlas 運行結束時晶片比日蝕中開始的熱約 8 °C[85, §3.4, Fig. 2]。 |
| 電力與熱條件 | 4.82 W · 5.92 W · 10.19 W · 11.48 W · 2–10% | 同樣的冷啟動也解釋了板子在軌耗能較少：四核 Pi 4.82 W 對 5.92 W，滿功率 Atlas 10.19 W 對 11.48 W，每項任務能耗低 2–10%[85, §4, Tabs. 3–5]。 |
| 證據與範圍 | 定性範圍 | 第三個要素是建議：跑短而間歇的任務。 |
| 證據與範圍 | 定性範圍 | 給低功率板子較高導熱的固定座。 |
| 證據與範圍 | 定性範圍 | 限制高功率板子的負載與時間。 |
| 證據與範圍 | 定性範圍 | 同一熱表面每次只執行一個裝置。 |
| 證據與範圍 | 30% | 排程運算以吸收剩餘太陽能並把放電深度維持在 30% 以下。 |
| 證據與範圍 | 定性範圍 | 以及把熱與能源模型放進硬體在迴路模擬器。 |
Original figure note原始圖說出處
Drawn for this primer from Xing et al., MobiCom 2024, §3.1–3.2, Table 3 and Figs. 3–6.4. Evaluation setup4. 實驗設定
The spacecraft uses ground links for experiment control and data return. The ground comparison measures device power, while radiation evidence is limited to the campaign’s observed computational errors.
太空船透過地面鏈路控制實驗及回傳資料，地面對照量的是裝置功率，輻射證據則限於活動中觀察到的運算錯誤。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | Qualitative scope | This is an in-orbit measurement, the only one in this primer's fact base for commodity compute. Thus, the setup statement is about what the satellite capabilities and operating limits. |
| Latency comparison | 1 · 100 · 9.6/4.8 | BUPT-1 uses ground-only links: 100 Mbps X-band downlink, 1 Mbps uplink, and 9.6/4.8 kbps telemetry/command. ISL rate, latency and topology lie outside the platform. Ground passes carry experiment control and data, with contact times unspecified [85, Tab. 1, §2.3]. |
| Power and thermal conditions | Qualitative scope | The ground control was a thermal replica of the satellite with the same boards, powered from a programmable monitor rather than a battery. This explains why the energy comparison is of device power rather than of a whole power system. |
| Orbital configuration | 490 km | The six-month campaign records zero computational single-event errors. Particle counters, upset counts and latch-up counts lie outside the instrumentation. At about 490 km, the authors expect weaker single-event exposure than at higher orbits because of magnetospheric shielding [85, §6]. |
| 證據與範圍 | 定性範圍 | 這是在軌量測，也是本教材事實庫中唯一關於商用運算的在軌量測，所以設定的說明是關於衛星的能力與操作範圍。 |
| 證據與範圍 | 1 · 100 · 9.6/4.8 | BUPT-1 有 100 Mbps 的 X 頻段下行與 1 Mbps 上行，加上 9.6/4.8 kbps 的遙測與指令通道，僅提供地面連線，ISL 速率、延遲與拓樸列於平台範圍之外，所有實驗控制與資料回傳都經由地面過頂，而通聯時間則待補充[85, Tab. 1, §2.3]。 |
| 證據與範圍 | 定性範圍 | 地面對照是搭載相同板子的衛星熱複製品，由可程式電源監測器供電，這也是為什麼能源比較是裝置功率，整體電力系統則需另外評估。 |
| 軌道配置 | 490 km | 輻射儀器量測列於實驗範圍之外：單事件的發現是六個月活動中運算錯誤計數為零，粒子計數器待配置，翻轉與閂鎖計數待補充，而在約 490 km 的高度，作者指出磁層讓單事件效應較更高軌道弱[85, §6]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Pi enters the >80 °C throttling zone within ≈50 min at >75% load. Chip saturates ≈82 °C. Throttling ≈5%. YOLOv3/v5-Lite latency +≈10%, other tasks +5–8% [85, §3.1–3.2, Fig. 3].Pi 在負載 >75% 時約 50 分鐘內進入 >80 °C 降頻區。晶片約 82 °C 飽和。降頻約 5%。YOLOv3/v5-Lite 延遲 +約 10%，其他任務 +5–8%[85, §3.1–3.2, Fig. 3]。 | Two Pi units, stock mount. The board surface rarely exceeds 15 °C, so the chip-to-board conduction path sets the limit.兩塊 Pi，原廠固定座。板子表面很少超過 15 °C，所以限制來自晶片到板子的導熱路徑。 |
| One Atlas reaches the 30 °C surface limit at 540 min. Two concurrent Atlases at ≈40 min, one killed by the OS at ≈50 min and the other shut down at ≈110 min [85, §3.2, Fig. 6].單一 Atlas 在 540 分鐘達 30 °C 表面上限。兩塊同時運作約 40 分鐘即達，一塊約 50 分鐘被作業系統終止、另一塊約 110 分鐘關機[85, §3.2, Fig. 6]。 | The reported value is surface temperature. The limit is the satellite's −10 to 30 °C design range, and recovery to ≈17 °C took 15 min after stopping.數值描述表面溫度。上限是衛星 −10 到 30 °C 的設計範圍，停止後 15 分鐘回到約 17 °C。 |
| Saturation 2–4× slower in orbit than on the ground replica: Pi four-core 86 vs 39 min, Atlas four-thread full 147 vs 88 min [85, Tab. 3].在軌飽和比地面複製品慢 2–4 倍：Pi 四核 86 對 39 分鐘，Atlas 四執行緒滿功率 147 對 88 分鐘[85, Tab. 3]。 | The Pi pair gives 2.2× and the Atlas pair 1.7× (calc table). Thus, the "2–4×" range is wider than the two Table 3 rows alone support.Pi 一組為 2.2 倍、Atlas 一組為 1.7 倍（計算表），所以「2–4 倍」的範圍比 Table 3 這兩列單獨能支持的更寬。 |
| Device power lower in orbit: Pi four-core 4.82 vs 5.92 W, Atlas full 10.19 vs 11.48 W. Task energy 2–10% lower [85, §4, Tabs. 3–5].在軌裝置功率較低：Pi 四核 4.82 對 5.92 W，Atlas 滿功率 10.19 對 11.48 W。任務能耗低 2–10%[85, §4, Tabs. 3–5]。 | Attributed to the colder ambient. Ground side powered by a programmable monitor. Satellite-vs-ground latency differed by <1% for the Atlas.歸因於較冷的環境。地面端由可程式監測器供電。Atlas 的衛星對地面延遲差異 <1%。 |
| Solar peak ≈40 W, ≈16 Wh harvested per orbit. Computing 30–40% of short-term peak energy. ≈6% of solar energy wasted with a full battery. DoD average 16%, briefly >50% against a 30% target [85, §4, Fig. 13].太陽能峰值約 40 W、每圈收穫約 16 Wh。運算佔短期峰值能量 30–40%。電池滿時約 6% 太陽能被浪費。DoD 平均 16%、曾短暫 >50%，目標為 30%[85, §4, Fig. 13]。 | Two 115 Wh batteries. The paper's communication power is given as ≈18 W in Fig. 11 and 38 W in §4.2, an internal inconsistency the fact base records.兩顆 115 Wh 電池。論文的通訊功率在 Fig. 11 為約 18 W、在 §4.2 為 38 W，事實庫記錄了這個內部差異。 |
| Zero computational errors from single-event effects in six months on unhardened boards [85, §6].商用板子六個月內的單事件運算錯誤計數為零[85, §6]。 | A negative finding with event counts unspecified, one satellite, one phase of the solar cycle, ≈490 km.運算錯誤計數為零的觀察，粒子事件計數待補充，一顆衛星、太陽週期的一個相位、約 490 km。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| How much slower is saturation in orbit, per Table 3 row?依 Table 3 各列，在軌飽和慢多少？ | \( \dfrac{86}{39} \approx 2.2,\qquad \dfrac{147}{88} \approx 1.7 \) | 2.2× (Pi) · 1.7× (Atlas)computed from Table 3. The paper's text states 2–4×由 Table 3 計算。論文內文寫 2–4 倍 |
| What does a three-hour compute block cost against two orbits of harvest?三小時的運算區塊相對兩圈的收穫耗多少？ | \( E_{\text{Pi}} = 4.82 \times 3 \approx 14.5\ \text{Wh},\quad E_{\text{Atlas}} = 10.19 \times 3 \approx 30.6\ \text{Wh},\quad E_{\text{sun}} = 2 \times 16 = 32\ \text{Wh} \) | 45% / 96% of harvestcomputed. Device powers from Tables 3–5, 16 Wh per orbit from Fig. 13, 3 h ≈ two orbits from Table 2. Ignores bus and communication loads計算值。裝置功率來自 Tables 3–5，每圈 16 Wh 來自 Fig. 13，3 小時約兩圈來自 Table 2。忽略平台與通訊負載 |
| How much energy does a 50% depth of discharge remove?50% 的放電深度抽走多少能量？ | \( 0.5 \times 2 \times 115\ \text{Wh} \) | 115 Whcomputed. Two 115 Wh batteries from Table 1, 50% DoD from §4計算值。兩顆 115 Wh 電池來自 Table 1，50% DoD 來自 §4 |
| How much lower is device power in orbit?在軌裝置功率低多少？ | \( 1 - \dfrac{4.82}{5.92} \approx 0.19,\qquad 1 - \dfrac{10.19}{11.48} \approx 0.11 \) | 19% (Pi) · 11% (Atlas)computed from Tables 3–5. The paper's 2–10% figure is energy per task, which includes throttling-lengthened runtimes由 Tables 3–5 計算。論文的 2–10% 是每項任務的能耗，包含降頻拉長的執行時間 |
6. Limits and open questions6. 範圍與待解問題
- The evidence covers Raspberry Pi and Atlas boards. Raspberry Pi 4B and Atlas 200 DK operate at 5–14 W. Peer-reviewed Jetson and GPU-class orbital thermal measurements remain pending. Chapter 8 extrapolates the 50-minute and 540-minute curves to 700 W only through Stefan–Boltzmann calculations, spanning two orders of magnitude.證據涵蓋 Raspberry Pi 與 Atlas 板卡。板子是 5–14 W 的 Raspberry Pi 4B 與 Atlas 200 DK。事實庫記載Jetson 與 GPU 級板卡的同儕審查在軌熱量測仍待完成，所以把 50 分鐘與 540 分鐘的曲線外推到 700 W 的加速器是跨越兩個數量級的跳躍，第 8 章只能靠 Stefan–Boltzmann 算術來做。
- One satellite, twelve experiments. Two device types with two units each on a single 12U bus in a single sun-synchronous orbit over four months. The thermal numbers are properties of BUPT-1's structure and mounts as much as of the boards.一顆衛星、十二個實驗。兩種裝置各兩塊，在單一 12U 平台、單一太陽同步軌道上四個月。熱的數字既是板子的性質，也同樣是 BUPT-1 結構與固定座的性質。
- Radiation evidence is a zero-error observation with event counts unspecified. Six-month zero-error operation agrees with Bhattacherjee et al.'s below-the-belt argument. Particle flux, upset and latch-up counts remain open. Rate comparison needs dose and event counts. Section 8.4 gives Suncatcher's proton-test reference of 1 silent-data-corruption event per 17 rad.輻射證據是零錯誤觀察，事件計數待補充。六個月的單事件錯誤計數為零令人鼓舞，也與 Bhattacherjee 等人「內帶以下」的論證一致，但粒子通量待記錄、翻轉或閂鎖計數，所以換算速率還需事件與劑量計數，才能與第 8.4 節 Suncatcher 質子測試每 17 rad 一次的靜默資料損毀數字比較。
- Internal inconsistencies. Communication power is about 18 W in Fig. 11 and 38 W in §4.2. The claimed "2–4× slower" heating spans more than Table 3's 2.2× and 1.7×. The fact base retains both.內部差異。通訊功率在 Fig. 11 約 18 W、在 §4.2 為 38 W，而「慢 2–4 倍」的說法比 Table 3 兩列給出的 2.2 倍與 1.7 倍更寬。事實庫記錄了這兩點。
- Link measurements remain open. Contact times and realized downlink throughput remain unspecified. Checking the 100 Mbps specification and the downlink time for 30 GB requires those measurements. The platform uses ground connectivity only.實際鏈路量測待補充。通聯時間與實際下行吞吐量仍待補充，核對 100 Mbps 規格及 30 GB 酬載資料下傳時間都需要這些量測。平台僅提供地面連線。
- The ground baseline uses a laboratory power supply. A programmable monitor powers the replica. The 2–10% saving therefore compares devices. Satellite power-system efficiency and depth-of-discharge excursions require separate evidence.地面基準以電源供應器供電。複製品由可程式監測器供電，所以 2–10% 的節能是裝置層級的比較，衛星電力系統效率與放電深度偏移則需另外量測。
- Guidelines await scheduler validation. Recommendations include short intermittent tasks, mount changes and a 30% DoD ceiling. A runtime enforcing them still needs flight tests. Its interaction with Kodan frame deadlines and Chapter 12 inference pipelines remains open.建議作為排程器仍待測試。短而間歇的任務、固定座改動與 30% DoD 上限都只是指引。強制執行這些規則的執行期系統仍待在軌測試，而這種工作週期如何與 Kodan 式的逐幀期限或第 12 章的多衛星推論管線互動仍是待解。
7. How it connects7. 與其他內容的關聯
- Section 8.3 uses Pi and Atlas curves to show conduction into a small radiating body limiting orbital computation. These measurements motivate Panel 8.2 radiator sizing.第 8.3 節：Pi 與 Atlas 的曲線顯示，商用運算在冰冷太空中的熱行為取決於傳導到小輻射面的路徑，並為面板 8.2 的輻射器尺寸計算提供證據。
- Section 8.2 uses 40 W peak solar input, 16 Wh per orbit and the 50% discharge excursion as a small measured eclipse/battery example.第 8.2 節：40 W 峰值、每圈 16 Wh 與 50% 的放電深度偏移是日蝕與電池預算最小的真實實例。
- Section 8.4: the six-month zero-error observation sits in the radiation table beside the Trillium and Jetson TX2 proton tests.第 8.4 節：六個月錯誤計數為零的觀察列在輻射表中，與 Trillium 和 Jetson TX2 的質子測試並列。
- Section 8.6: BUPT-1 and the Tiansuan programme in the table of what has flown.第 8.6 節：BUPT-1 與天算星座計畫在「已飛行」表中。
- Section 3.6: the 100 Mbps downlink and 1 Mbps uplink of a small Earth-observation-class bus.第 3.6 節：小型地球觀測等級平台的 100 Mbps 下行與 1 Mbps 上行。
- Section 11.4: thermal behavior in orbit is one of the items the Suncatcher paper leaves open. This is the only flight measurement available to reason about it.第 11.4 節：在軌熱行為是 Suncatcher 論文待解的項目之一，而這是唯一可用來推理它的飛行量測。
- OEC [60] and Kodan [83] assume a compute budget this paper shows must be duty-cycled. Bhattacherjee et al. [2] budget 350 W with thermal budget pending. Suncatcher [36] is the proposal whose prototype will produce the next such curve.OEC[60]與 Kodan[83]假設的運算預算，本文顯示必須以工作週期切分。Bhattacherjee 等人[2]編列 350 W ，熱預算則待補充。Suncatcher[36]是其原型將產生下一條這種曲線的提案。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 12U · 17.44 kg · 487.6–494.7 km · 97.371° | BUPT-1 bus and sun-synchronous orbit [85, Tab. 1] | BUPT-1 平台與太陽同步軌道[85, Tab. 1] | |
| 2 × Pi 4B + 2 × Atlas 200 DK | Flown devices: Pi and Atlas [85, §2.2] | 飛行裝置：Pi 與 Atlas[85, §2.2] | |
| ≈50 min · ≈82 °C · ≈5% | Pi time to the >80 °C zone at >75% load, chip saturation, frequency throttling [85, §3.1–3.2] | Pi 在負載 >75% 下到 >80 °C 區的時間、晶片飽和溫度、頻率降幅[85, §3.1–3.2] | |
| 540 min · ≈40 min | One Atlas vs two Atlases to the 30 °C surface limit [85, §3.2, Fig. 6] | 單一 Atlas 對兩塊 Atlas 到 30 °C 表面上限的時間[85, §3.2, Fig. 6] | |
| 86 vs 39 min · 147 vs 88 min | Saturation time in orbit vs ground replica, Pi four-core and Atlas full four-thread [85, Tab. 3] | 在軌對地面複製品的飽和時間，Pi 四核與 Atlas 滿功率四執行緒[85, Tab. 3] | |
| 4.82 vs 5.92 W · 10.19 vs 11.48 W | Device power in orbit vs ground, Pi four-core and Atlas full [85, Tabs. 3–5] | 在軌對地面的裝置功率，Pi 四核與 Atlas 滿功率[85, Tabs. 3–5] | |
| ≈40 W · ≈16 Wh · 2 × 115 Wh | Solar peak, energy harvested per orbit, battery capacity [85, §4, Fig. 13, Tab. 1] | 太陽能峰值、每圈收穫能量、電池容量[85, §4, Fig. 13, Tab. 1] | |
| 30% → >50% | Depth-of-discharge target vs brief excursion. Two-week average 16% [85, §4] | 放電深度目標對短暫偏移。兩週平均 16%[85, §4] | |
| 0 in 6 months | Single-event computational errors: zero observed. Particle-event counts unspecified [85, §6] | 單事件運算錯誤計數為零，粒子事件計數待補充[85, §6] | |
| 100 Mbps / 1 Mbps | Downlink / uplink. Ground connectivity only. Contact times unspecified [85, Tab. 1] | 下行／上行。僅地面鏈路。通聯時間待補充[85, Tab. 1] |