SpaceMoE: Realizing Distributed Mixture-of-Experts Inference over Space NetworksSpaceMoE：在太空網路上實現分散式混合專家推論
SpaceMoE places mixture-of-experts layers and experts across a satellite constellation. A slow radiation-hardened processor sets most of the reported token time. Interpreting the placement gain requires the paper's link, failure and compute assumptions.SpaceMoE 把混合專家模型的層與專家放置於衛星星座中，較慢的抗輻射處理器決定大部分 token 生成時間。解讀放置策略的增益時，須保留論文的鏈路、故障與運算假設。
它的貢獻是兩層放置：先切出環狀的層子網，再依啟動頻率把專家指派給衛星，而 Theorem 1 證明這個指派是最佳的。此外論文的系統規模估算、加速倍數與處理器速率彼此各有差異。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | SpaceMoE distributes LLaMA-MoE-3.5B, with about 3.5 B active of 6.7 B total parameters. Its 32 layers have 8 experts each and top-2 routing. |
| Constellation configuration | over a simulated polar constellation of 33 planes × 32 satellites = 1,056 satellites at 550 km and 87° |
| ISL configuration | in which every satellite has up to four laser ISLs at "≥ 100 Gbps", delay is propagation only |
| Angular-rate constraint | a link exists only while the angular-rate difference stays below 0.12 rad/s |
| Link-survival model | and each link survives a snapshot as an independent Bernoulli event with probability 0.95 [78, §II, §VII-A] preprint. |
| Placement and evaluation | Its contribution is a two-level placement, ring-based layer subnets first and then an expert-to-satellite assignment by activation frequency that Theorem 1 proves optimal. SpaceMoE reports 1.02–1.07 s per token, compared with 3.34–5.30 s for three random placements. Hardware is a Frontgrade SBC-2A72 at 7.28 GFLOPS effective. [78, Tab. II, Figs. 6–7] |
| Compute limit | Compute accounts for almost the entire second. About 7 GFLOP per token divided by 7.28 GFLOPS gives 0.96 s (computed below). The result mainly measures placement on a slow processor. System sizing, the speedup claim and processor rate contain internal inconsistencies. |
| 模型配置 | SpaceMoE 把一個混合專家語言模型，也就是 LLaMA-MoE-3.5B（6.7 B 總參數中約 3.5 B 活躍、32 層、每層 8 個專家、top-2 路由） |
| 星座配置 | 分布到一個模擬的極軌星座上：33 個軌道面 × 32 顆 = 1,056 顆衛星，位於 550 km、87° |
| ISL 配置 | 每顆衛星最多四條「≥ 100 Gbps」的雷射 ISL，延遲只算傳播 |
| 角速率條件 | 只有當角速率差維持在 0.12 rad/s 以下時鏈路才存在 |
| 鏈路存活模型 | 而每條鏈路在每個快照中以機率 0.95 的獨立 Bernoulli 事件存活[78, §II, §VII-A]預印本。 |
| 評估條件 | 在有效 7.28 GFLOPS 的 Frontgrade SBC-2A72 抗輻射板上，它報告每 token 1.02–1.07 s，三種隨機放置則為 3.34–5.30 s[78, Tab. II, Figs. 6–7]。 |
| 評估條件 | 主要的限制是運算幾乎解釋了那一秒的全部，因為每 token 約 7 GFLOP 除以 7.28 GFLOPS 就是 0.96 s（見下方計算），所以這個結果是關於慢速處理器上的放置的證據，其網路性能的證據範圍則較有限。 |
1. Where this paper sits1. 這篇論文的位置
Chapter 12 compares three orbital-inference preprints whose ISL assumptions differ by over two orders of magnitude. SpaceMoE uses the optimistic rate assumption. It alone spreads a model across the whole constellation. CollabLLM-LEO chains five satellites through 0.5 Gbit/s links, while LLMSpace budgets carbon for one DGX-class payload.
SpaceMoE places mixture-of-experts modules across a thousand satellites. Its objective minimizes the expected path followed by token hidden states. [96] [97] Its output is a placement on a graph whose edges change with orbital motion. This makes network placement central to the problem. It is also the paper whose headline number turns out to say least about the network, for reasons that Section 3 makes concrete.
第 12 章把已發表的在軌語言模型推論研究歸成三篇預印本，它們對 ISL 能承載多少的假設相差超過兩個數量級，而 SpaceMoE 位於這個範圍樂觀的那一端，也是唯一把單一模型分散到整個星座的論文。CollabLLM-LEO[96]用 0.5 Gbit/s 鏈路串起五顆衛星，LLMSpace[97]為一個 DGX 等級的酬載估算碳排，SpaceMoE 則問：混合專家模型的專家應該如何擺放在一千顆衛星上，才能讓一個 token 的隱藏狀態以最短的期望路徑穿過各層。因此它是這組論文中最像網路問題的一篇，因為它的輸出是一張圖上的放置，而這張圖的邊會隨軌道運動出現又消失。它同時也是頭條數字對網路說明最少的一篇，原因會在第 3 節具體說明。
A mixture of experts transformer divides each layer's feed-forward block into experts. A small router activates only a few experts per token. Thus, the parameters that must be resident somewhere are many more than the parameters exercised for any one token. That property is what makes the idea attractive for a constellation.
This occurs because each satellite can hold a slice of a model too large for any single board while the per-token work stays bounded. It is also what makes the traffic distinctive. This occurs because the hidden state must visit whichever experts the router picks at each of the 32 layers.
混合專家（mixture of experts）是一種 transformer，它每一層的前饋區塊被拆成數個專家，由一個小型路由器對每個 token 只啟動其中少數幾個，因此必須常駐在某處的參數遠多於任何一個 token 實際用到的參數。這個性質正是它對星座有吸引力的原因：每顆衛星可以只存放一個大到任何單板都記憶體需求超出單板容量的模型的一片，而每個 token 的工作量仍有上限。這個性質也讓它的流量很尤其，因為隱藏狀態在 32 層中的每一層都必須拜訪路由器所選中的專家。
2. The problem, measured2. 問題，用量測說話
The problem the authors pose is latency per generated token when the experts live on different satellites. The radiation-hardened single-board computer has realistic flight heritage. Chapter 12 compares it with CollabLLM-LEO's Orin-class modules and LLMSpace's H100-class payload. Its compute rate is three to four orders of magnitude below modern accelerators, strongly shaping the evaluation.作者提出的問題是：當專家分住在各異衛星上時，每產生一個 token 的延遲是多少。選用抗輻射單板電腦在飛行資歷上是務實的，第 12 章把它拿來與 CollabLLM-LEO 假設的 Orin 等級模組以及 LLMSpace 假設的 H100 等級酬載對照，但這也讓處理器比現代加速器低了三到四個數量級，而這個差距左右了評估所能顯示的一切。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | All experiments use LLaMA-MoE-3.5B with about 3.5 B active of 6.7 B total parameters. Its 32 layers contain 8 experts each, totaling 256 experts with top-2 routing. A 4,096-token forward pass requires 36.3 TFLOPs. [78, §VII] preprint |
| Memory | Each satellite has a Frontgrade SBC-2A72 in SpaceVPX 3U format with 8 GB memory and 10.4 GFLOPS peak. A 70% utilization assumption gives 7.28 GFLOPS effective. RAD5545 and SpaceCloud iX10 are alternative candidate platforms. [78, §VII, Tab. I] |
| 模型配置 | 所有實驗的模型都是 LLaMA-MoE-3.5B，6.7 B 總參數中約 3.5 B 活躍，32 個 MoE 層、每層 8 個專家、共 256 個專家，top-2 路由，序列長度 4,096 時一次前向傳遞為 36.3 TFLOPs[78, §VII]預印本。 |
| 機上運算平台 | 每顆衛星搭載一塊 SpaceVPX 3U 規格的 Frontgrade SBC-2A72，記憶體 8 GB、峰值 10.4 GFLOPS，作者以 70% 使用率折算為有效 7.28 GFLOPS，並列出 RAD5545 與 SpaceCloud iX10 作為其他候選平台[78, §VII, Tab. I]。 |
3. The design3. 設計
The expert placement works at two levels. The first level creates one contiguous ring-based subnet per MoE layer. Each subnet has a central gateway satellite [78, Eqs. 17–18] . Thus, a token moves from layer to layer by stepping between neighboring subnets rather than by crossing the constellation. The second level assigns one layer's experts to subnet satellites to minimize expected latency.
Theorem 1 assigns the i -th most active expert to the i -th lowest-latency satellite. Per-layer cost is \(O(I\log I + V\log V)\) for \(I\) experts and \(V\) satellites. Frequent experts sit near the gateway, as popular items do in a cache. [78, Eqs. 19–20, 29–30, 33–34] [78, §IV-C, §V] The rarely used ones can afford to be far away, which Figure A draws on the left.
專家放置（expert placement）分兩層進行。第一層把星座切成環狀的層子網，每個 MoE 層對應一組相鄰的衛星，中心有一顆閘道衛星[78, Eqs. 17–18]，讓 token 在層與層之間是跨到相鄰的子網，使路徑集中於相鄰子網之間。第二層把一層的專家指派給該子網的衛星，這是一個最小化每層期望延遲的二元指派[78, Eqs. 19–20, 29–30, 33–34]，而 Theorem 1 證明最佳解是把第 i 常被啟動的專家放到第 i 低延遲的衛星，對 \(I\) 個專家與 \(V\) 顆衛星每層的成本為 \(O(I\log I + V\log V)\)[78, §IV-C, §V]。直覺就像快取：路由器最常選的專家應該離閘道一跳，很少用到的則可以放遠一些，圖 A 左半畫的就是這個概念。
The network model under that placement is simple on purpose. Link delay adds chord propagation to transmission time \(MQ_B/R\). The hidden state has dimension \(M\), quantization \(Q_B\) bits and link rate \(R\). The paper states that propagation dominates communication.
Hidden-state dimension \(M\), quantization \(Q_B\) and slot duration remain unspecified. [78, Eq. 5] [78, Eq. 6] Queueing and per-hop processing lie outside the model [78, §II-B] . What makes the result interpretable is a check that uses only the paper's inputs.
在這個放置之下的網路模型刻意簡化。鏈路上的延遲是弦長幾何的傳播時間[78, Eq. 5]加上一個傳輸項 \(MQ_B/R\)，其中隱藏狀態維度為 \(M\)、量化為 \(Q_B\) 位元、鏈路速率為 \(R\)[78, Eq. 6]。論文說通訊由傳播主導，但 \(M\)、\(Q_B\) 與時槽長度皆仍待補充，排隊與每跳處理則位於模型範圍之外[78, §II-B]。讓結果得以解讀的是一個只用論文自身輸入的檢查。因此這個放置相對隨機基準的增益，是「每層花多少跳」上的增益，量測的基準線則是處理器所設下的下限。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | A decode token uses about 3.5 B active parameters, requiring 7 GFLOP at two operations per parameter. At 7.28 GFLOPS, this takes 0.96 s. It accounts for 90–94% of the reported 1.02–1.07 s. The paper's 36.3 TFLOPs per 4,096-token pass instead gives 8.9 GFLOP and 1.22 s. Thus, the two estimates bracket the result (calc table). |
| Propagation delay | What remains for the network is at most about 0.1 s per token. This is the propagation time of at most about two dozen intra-plane hops of 4.5 ms. Placement gains reduce per-layer hop counts. The processor sets the lower time bound used for comparison. |
| 評估條件 | 一個解碼 token 動用約 3.5 B 活躍參數，以每參數兩次浮點運算計為 7 GFLOP，在 7.28 GFLOPS 的板子上就是 0.96 s，已經是所報告 1.02–1.07 s 的 90–94%。 |
| 評估條件 | 而用論文自己的「4,096 token 一次前向 36.3 TFLOPs」換算，每 token 為 8.9 GFLOP、1.22 s，兩個估計把結果夾在中間（見計算表）。 |
| 路徑長度與跳數 | 留給網路的最多約為每 token 0.1 s，相當於至多為約兩打、每跳 4.5 ms 的同軌跳躍的傳播時間。 |
Original figure note原始圖說出處
Drawn for this primer from Wang et al., arXiv 2605.00515, §IV–V, Table II and Figs. 6–7. The subnet layout is schematic and the dashed floor is computed in the table in Section 5.4. Evaluation setup4. 實驗設定
The evaluation is a simulation over sampled topology snapshots, with testbeds, flight hardware and accuracy results outside its scope, over eight lm-eval-harness datasets [78, §VII] preprint.評估是在取樣的拓樸快照上進行的模擬，測試平台、飛行硬體及準確度結果皆位於範圍之外，使用八個 lm-eval-harness 資料集[78, §VII]預印本。每顆衛星最多四條雷射 ISL，兩條同軌、兩條跨軌，也就是第 2 章經典的 +Grid 度數。ISL 速率寫成「≥ 100 Gbps」，延遲模型是弦長傳播加 \(MQ_B/R\) 傳輸，排隊、交換及每跳處理皆位於模型範圍之外。只有當兩顆衛星的角速率差低於 0.12 rad/s 時鏈路才可行[78, Eq. 2]。故障模型讓每條鏈路在每個快照以機率 0.95 的獨立 Bernoulli 事件存活，所有鏈路相同且在時間上相互獨立[78, Eq. 3]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Plane configuration | The polar Walker constellation has 33 planes × 32 satellites, phasing F = 13, altitude 550 km and inclination 87°. It is sampled over 200 time slots and described as aligned with existing megaconstellations. Each satellite has up to four ISLs, two intra-orbit and two inter-orbit, matching classic +Grid degree. Chapter 2 describes this +Grid degree. [78, §VII] |
| Link or data rate | ISL rate is "≥ 100 Gbps". Delay includes chord propagation and \(MQ_B/R\) transmission, with queueing, switching and per-hop processing outside the model. Links require angular-rate difference below 0.12 rad/s. Each link survives a snapshot with probability 0.95 under an independent Bernoulli model. Survival is identical across links and uncorrelated over time. [78, Eq. 2] [78, Eq. 3] |
| Orbital altitude | Each satellite has a Frontgrade SBC-2A72 at 7.28 GFLOPS effective and 8 GB memory. Figure 7 varies altitude, constellation size, link survival and angular-rate threshold. |
| 軌道面配置 | 拓樸是 33 個軌道面 × 32 顆衛星的極軌 Walker 星座，相位 F = 13，高度 550 km、傾角 87°，取樣 200 個時槽，並被描述為與現有的巨型星座一致[78, §VII]。 |
| 星座配置 | 硬體是每顆衛星一塊有效 7.28 GFLOPS、8 GB 的 Frontgrade SBC-2A72，Fig. 7 則以此為中心掃描高度、星座規模、存活機率與角速率門檻。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| SpaceMoE achieves 1.02–1.07 s per token over eight datasets [78, Tab. II, Figs. 6–7].SpaceMoE 在八個資料集上達到每 token 1.02–1.07 s[78, Tab. II, Figs. 6–7]。 | Simulated, LLaMA-MoE-3.5B on a 7.28 GFLOPS board. Compute alone accounts for 90–94% (calc table).模擬，LLaMA-MoE-3.5B 跑在 7.28 GFLOPS 的板子上。僅運算就佔 90–94%（計算表）。 |
| Random baselines take 5.28–5.30 s (RandPlace), 4.13–4.16 s (RandIntra) and 3.34–3.37 s (RandIntra-CG) [78, Tab. II].隨機基準各自為 5.28–5.30 s（完全隨機）、4.13–4.16 s（同軌隨機）與 3.34–3.37 s（連續區隨機）[78, Tab. II]。 | All baselines are placements of the same model on the same constellation. Single-satellite, ground-side and dense-model baselines lie outside the comparison.所有基準都是同一模型在同一星座上的放置方式。單衛星、地面端與稠密模型基準皆位於比較範圍之外。 |
| At least a threefold reduction against the best baseline [78, Tab. II].相對最佳基準至少減少三倍（摘要、Table II）。 | Table II gives 3.1–3.3× (calc table), but §VII-B says "at least a twofold" reduction (fact base Part C, item 38).Table II 給出 3.1–3.3 倍（計算表），但 §VII-B 寫「至少兩倍」（事實庫 Part C，第 38 項）。 |
| Theorem 1: rank-matching experts by activation frequency to satellites by latency is optimal per layer, at \(O(I\log I + V\log V)\) [78, §V].Theorem 1：依啟動頻率排名專家、依延遲排名衛星並一一對應，在每層是最佳的，成本 \(O(I\log I + V\log V)\)[78, §V]。 | Optimal under the paper's latency model of propagation plus transmission with i.i.d. link survival. The introduction orders experts ascending where the theorem uses descending.在論文「傳播加傳輸、鏈路獨立同分布存活」的延遲模型下最佳。導論中專家依遞增排序，定理則用遞減。 |
| Communication is "dominated by propagation" [78, §II-C].通訊「由傳播主導」[78, §II-C]。 | True in the model because transmission of one hidden state at ≥ 100 Gbps is negligible and queueing excluded exists. The hidden-state size \(M\), \(Q_B\) and slot length remain unspecified.在模型中成立，因為一個隱藏狀態以 ≥ 100 Gbps 傳輸可忽略，而且模型排除排隊。隱藏狀態大小 \(M\)、\(Q_B\) 與時槽長度皆仍待補充。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| How long does one token take on the board from active parameters alone?僅以活躍參數計，一個 token 在板子上要多久？ | \( \dfrac{2 \times 3.5\times10^{9}}{7.28\times10^{9}\ \text{FLOP/s}} \) | 0.96 scomputed. 3.5 B active parameters and 7.28 GFLOPS from §VII-A. 2 FLOP per parameter per token計算值。3.5 B 活躍參數與 7.28 GFLOPS 取自 §VII-A。每 token 每參數 2 FLOP |
| How long from the paper's own forward-pass FLOPs?以論文自己的前向傳遞 FLOPs 計要多久？ | \( \dfrac{36.3\times10^{12}/4{,}096}{7.28\times10^{9}} \) | 8.9 GFLOP → 1.22 scomputed. 36.3 TFLOPs at sequence 4,096 from §VII-A3計算值。序列 4,096 下 36.3 TFLOPs 取自 §VII-A3 |
| What share of the reported time is compute?所報告的時間中運算佔多少？ | \( 0.96/1.07 \;\ldots\; 0.96/1.02 \) | 90–94%computed. Table II range計算值。Table II 範圍 |
| How long is one intra-plane hop in a 32-satellite ring at 550 km?550 km、32 顆一環的同軌一跳有多長？ | \( 2\,(6{,}371+550)\,\sin(\pi/32)\ \text{km},\quad d/c \) | 1,357 km · 4.5 mscomputed with the chord formula of Chapter 2. \(R_E\) = 6,371 km以第 2 章弦長公式計算。\(R_E\) = 6,371 km |
| How many such hops fit in what compute leaves over?運算剩下的時間能容納幾跳？ | \( \dfrac{1.02-0.96}{0.0045} \;\ldots\; \dfrac{1.07-0.96}{0.0045} \) | ≈ 13–24 hopscomputed. An upper bound, since any transmission time would reduce it計算值。為上限，任何傳輸時間都會讓它變少 |
| What reduction does Table II itself imply against the best baseline?Table II 本身相對最佳基準意味多少倍的減少？ | \( 3.34/1.07 \;\ldots\; 3.37/1.02 \) | 3.1–3.3×computed. §VII-B states "at least twofold"計算值。§VII-B 寫「至少兩倍」 |
6. Limits and open questions6. 限制與待解的問題
- The headline is compute-bound. Because 0.96 of the 1.02–1.07 s is the processor, the "≥ 100 Gbps" ISL is barely exercised. The paper's statement that communication is dominated by propagation holds only because the compute is so slow. On an Orin-class or H100-class payload the same placement would become propagation-bound, and Chapter 12's Section 12.5 shows where that flip happens.頭條數字受運算限制。因為 1.02–1.07 s 中有 0.96 s 是處理器，「≥ 100 Gbps」的 ISL 在評估中僅占少量時間。論文說通訊由傳播主導，只是因為運算太慢才成立。在 Orin 等級或 H100 等級的酬載上，同樣的放置會變成受傳播限制，第 12 章第 12.5 節會指出這個翻轉發生在哪裡。
- Baselines are random placements only. Every comparison is between placements of the same model on the same constellation. Ground-side serving by a terrestrial cluster lies outside the comparison. A dense model on one satellite is also outside its scope. The time per token still requires comparison with what a ground data center would deliver to the same user.基準只有隨機放置。每項比較都是同一模型在同一星座上的各異放置，所以論文將地面端服務替代方案列於比較範圍之外，即（由地面叢集回答請求）比較，單顆衛星上的稠密模型亦位於比較範圍之外，每 token 的時間仍待與地面資料中心對同一使用者的交付時間比較。
- Energy and orbital dynamics outside the evaluation. The evaluation reports time only. Power, eclipse, battery discharge depth and thermal duty cycling lie outside the evaluation. Xing et al. measure duty cycling on a 17 kg bus. Sustained orbital operation therefore requires energy analysis. An overlapping-author perspective preprint lists battery degradation and thermal limits, with quantification pending. [106]能源與軌道動態皆位於評估範圍之外。評估只報告時間。功耗、日蝕、電池放電深度，以及 Xing 等人在 17 kg 平台上量到的熱工作週期皆位於範圍之外，所以若要判斷星座整圈軌道的持續運作能力，仍須加入能源分析。與本文作者部分重疊的一篇觀點預印本[106]把電池老化與熱限制列為挑戰，量化結果則仍待補充。
- An optimistic failure model and queueing excluded. The 5% independent snapshot loss exceeds a production fabric's 0.057% monthly failure rate by two orders of magnitude. Correlated outages and reacquisition require separate modeling. [65, §2.3] This occurs because it excludes temporal correlation, re-acquisition delay after loss and handover dynamics. Together with the absence of queueing, this leaves out exactly the effects that bursty all-to-all expert routing would stress.樂觀的故障模型，排隊則位於模型範圍之外。每個快照 5% 的獨立鏈路損失比生產訓練網路每月 0.057% 的鏈路故障率差兩個數量級[65, §2.3]，但從另一方向看又太溫和，因為它將時間相關性、斷線後重新擷取延遲及換手動態排除於模型之外。加上排隊也位於模型範圍之外，它恰好遺漏了突發性全對全專家路由會加壓的那些效應。
- Internal inconsistencies. Section IV-A and Table I size a Switch Transformer with "over 70 billion parameters" and 140 GB at FP16. Experiments instead use the 6.7 B model. The abstract and Table II imply at least a threefold reduction while §VII-B says "at least a twofold". Table I lists 10 GFLOPS where the text derives 7.28 from 10.4. And the introduction orders experts ascending where Theorem 1 uses descending (fact base Part C, item 38). The safe quotation is the Table II range with the 6.7 B model.內部差異。§IV-A 與 Table I 的系統規模估算用的是「超過 700 億參數」、FP16 下 140 GB 的 Switch Transformer，而所有實驗都用 6.7 B 的模型。摘要與 Table II 暗示至少三倍的減少，§VII-B 卻說「至少兩倍」。Table I 列出 10 GFLOPS，內文則由 10.4 推得 7.28。導論把專家遞增排序，Theorem 1 則用遞減（事實庫 Part C，第 38 項）。安全的引用方式是 6.7 B 模型的 Table II 範圍。
- Missing quantities and assumptions to update. The hidden-state dimension, its quantization and the slot duration are unspecified. Thus, the transmission term requires those inputs for recalculation. The fact base leaves ground access outside its record. Thus, the 15-second user-link reconfiguration of Chapter 3 stays outside the model. And the four-ISL, "≥ 100 Gbps", zero-setup-delay link is an assumption that still requires an operator's published fleet specification.缺少的量與需要更新的假設。隱藏狀態維度、其量化與時槽長度都說明資料待補，所以傳輸項須取得這些輸入才能重算。地面接取段位於事實庫描述範圍之外，所以第 3 章使用者鏈路每 15 秒的重新配置位於模型範圍之外。而四條 ISL、「≥ 100 Gbps」、零建立延遲的鏈路，是仍待業者公布機隊規格支持的假設。
7. How it connects7. 與其他內容的關聯
- Section 12.2: the chapter's full treatment of SpaceMoE, with Figure 12.1 on layer-subnet placement and the compute-bound arithmetic reproduced here.第 12.2 節：該章對 SpaceMoE 的完整討論，包含關於層子網放置的圖 12.1，以及此處重現的運算限制算術。
- Section 12.3: CollabLLM-LEO makes the opposite assumption, a fast processor and a slow link, and works on the transmission term SpaceMoE ignores.第 12.3 節：CollabLLM-LEO 做了相反的假設，快處理器配慢鏈路，處理的正是 SpaceMoE 忽略的傳輸項。
- Section 12.4: LLMSpace and the rest of the set, which leave per-token path time unquantified.第 12.4 節：LLMSpace 與這組論文的其餘部分，每 token 路徑時間皆仍待量化。
- Section 12.5: Panel 12.1 moves the bottleneck between compute and link, and SpaceMoE is its compute-bound corner.第 12.5 節：面板 12.1 讓瓶頸在運算與鏈路之間移動，SpaceMoE 位於受運算限制的那一角。
- Section 12.6: ground and orbit as two serving sites, the comparison still required beyond this paper.第 12.6 節：地面與軌道作為兩個服務地點，正是本文範圍之外仍待完成的比較。
- Section 10.6 places SpaceMoE in the constellation regime of Gbps links and millisecond hops. Chapter 10 finds that regime suitable for inference. Synchronous training requires another regime.第 10.6 節：SpaceMoE 位於星座尺度的區間，Gbps 鏈路與毫秒級跳躍，第 10 章判斷這個區間適合推論、同步訓練則須另行評估。
- Section 13.6: research direction E, serving across ground and orbit with time and energy modeled together, starts from the gaps listed above.第 13.6 節：研究方向 E，在地面與軌道之間服務並同時模擬時間與能源，正是從上述缺口出發。
- Sibling digests are CollabLLM-LEO [96] and LLMSpace [97]. van Berkel [101] gives terrestrial rooflines for inference confined to one satellite. Suncatcher [36] provides a Tbps-formation reference for constellation-wide placement. Xing et al. [85] supplies the thermal duty cycle the paper omits.CollabLLM-LEO[96]與 LLMSpace[97]是兩篇姊妹導讀。van Berkel[101]主張限於單顆衛星的推論具有與地面相同的 roofline，Suncatcher[36]則是可拿來與星座尺度放置相比的 Tbps 編隊。Xing 等人[85]提供了本文遺漏的熱工作週期。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 6.7 B · ≈3.5 B active · 32 × 8 · top-2 | LLaMA-MoE-3.5B: total and active parameters, layers × experts, routing [78, §VII] | LLaMA-MoE-3.5B：總參數與活躍參數、層數 × 專家數、路由[78, §VII] | |
| 33 × 32 = 1,056 · 550 km · 87° | Simulated polar Walker constellation, F = 13, 200 slots [78, §VII] | 模擬的極軌 Walker 星座，F = 13，200 個時槽[78, §VII] | |
| ≤ 4 ISLs · ≥ 100 Gbps | Laser ISLs per satellite and their rate. Propagation-only delay [78, §II-B] | 每顆衛星的雷射 ISL 數與速率。只算傳播延遲[78, §II-B] | |
| < 0.12 rad/s · 0.95 | Angular-rate feasibility threshold. I.i.d. Bernoulli link survival per snapshot [78, Eqs. 2–3] | 角速率可行門檻。每個快照的獨立 Bernoulli 鏈路存活機率[78, Eqs. 2–3] | |
| 7.28 GFLOPS · 8 GB | Frontgrade SBC-2A72, 10.4 GFLOPS peak × 70% [78, §VII][78, Tab. I] | Frontgrade SBC-2A72，峰值 10.4 GFLOPS × 70%[78, §VII][78, Tab. I] | |
| 1.02–1.07 s/token | SpaceMoE, eight datasets, simulated [78, Tab. II] | SpaceMoE，八個資料集，模擬[78, Tab. II] | |
| 3.34–5.30 s/token | Three random placements, best to worst [78, Tab. II] | 三種隨機放置，由最佳到最差[78, Tab. II] | |
| ≈ 0.96 s | Compute-only time per token, 7 GFLOP ÷ 7.28 GFLOPS (computed) | 僅運算的每 token 時間，7 GFLOP ÷ 7.28 GFLOPS（計算值） |