Chapter 12 of 13 · Part III · AI in orbit
第 12 章（共 13 章）· 第三部 · 在軌 AI
LLM inference on satellites: the published systems and the assumptions they rest on衛星上的 LLM 推論：已發表的系統與它們所依賴的假設
This chapter calculates decode-token and KV-cache traffic, then examines the separate questions behind SpaceMoE, CollabLLM-LEO, LLMSpace, SkyMemory and Constella. It compares ISL assumptions from 0.5 Gb/s to 10 Tb/s and identifies measurements that remain open.本章計算解碼 token 與 KV 快取在星間鏈路上的流量，並說明 SpaceMoE、CollabLLM-LEO、LLMSpace、SkyMemory 與 Constella 各自處理的問題。接著比較從 0.5 Gb/s 到 10 Tb/s 的 ISL 假設，整理仍待完成的量測。
Before you start
開始之前
This chapter builds on Chapters 2, 3, 8, 9 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。
- Prefill預填（prefill） The first phase of serving a request, in which every prompt token passes through the model at once. 服務一個請求的第一階段，提示的所有 token 一次通過模型。
- Decode解碼（decode） The second phase, generating one output token per step, each step depending on the last. 第二階段，每一步產生一個輸出 token，每一步依賴前一步。
- KV cacheKV 快取 The keys and values that each attention layer stores per token so decode can reuse the prompt’s earlier computation. 每個注意力層為每個 token 儲存的 key 與 value，讓解碼重用先前的提示運算。
- Time per token每 token 時間 The wall-clock time one decode step takes for one request stream, the inference analogue of step time. 單一請求串流中一個解碼步驟所需的時鐘時間，相當於訓練中的步驟時間。
- Mixture of experts (MoE)混合專家（MoE） A transformer whose feed-forward block is split into experts of which a router activates only a few per token. 前饋區塊被切成多個專家、每個 token 只由路由器啟用其中少數幾個的 transformer。
- Expert placement專家配置 The assignment of experts to satellites so that the hidden state travels the shortest expected path. 把專家指派到衛星上，讓隱藏狀態走的期望路徑最短。
- Activation compression激活值壓縮 Sparsifying, quantising and entropy-coding the tensor that crosses a layer boundary between two devices. 對跨越兩個裝置之間層邊界的張量做稀疏化、量化與熵編碼。
12.1Inference and training traffic scales推論與訓練的流量尺度
Recall
回顧
Section 10.1 derives ring all-reduce traffic of \(2(N-1)/N\) times model bytes per worker per step. Section 9.4 lists flown single-satellite inference.
Chapter 10 finds constellation links three to five orders too slow per compute element for training. Inference has different traffic requirements. A decoder-only transformer serves requests in two phases. prefill processes all prompt tokens through every layer in one parallel sweep. Each output tensor has one row per prompt token. decode emits one output token per step, using the previous output, so each layer sends one row.
Thus, the tensor that leaves a layer has a single row. Each attention layer stores the keys and values of earlier tokens for reuse during decode in a KV cache. The cache grows by one entry per layer per generated token and is read in full at every step. The quantity a serving system minimizes is the time per token, the wall-clock time one decode step takes for one request stream. It plays the role step time played in Chapter 10, with the same two-term structure of compute plus exposed communication.
第 10 章拿星座去對照訓練，發現每個運算元件分到的鏈路慢了三到五個數量級。推論問的是另一個問題，差異來自跨越網路的資料內容。對一個 decoder-only transformer 服務一個請求有兩個階段。在預填（prefill）階段，提示中的每個 token 以一次平行掃過的方式通過所有層，因此離開某一層的張量每個提示 token 占一列。而在解碼（decode）階段，模型每一步輸出一個 token、每一步都以前一步為輸入，因此離開某一層的張量只有一列。
為了避免每個解碼步驟重算整段提示，每個注意力層把它為先前每個 token 算出的 key 與 value 留在KV 快取裡，這個快取每產生一個 token 就在每一層多一筆，而且每一步都要整個讀一遍。服務系統要最小化的量是每 token 時間，也就是單一請求串流中一個解碼步驟所需的時鐘時間。它扮演第 10 章中步驟時間的角色，同樣由運算加上暴露的通訊這兩項組成。
Decode sends one hidden vector across each layer boundary, while prefill sends a vector per prompt token. KV state grows across layers and makes migration much larger.
解碼在每個層邊界傳送一個隱藏向量，預填則為每個提示 token 各傳一個。KV 狀態跨層累積，搬移量因而更大。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The bytes follow from the architecture alone. |
| Model and compute | 7B, 4,096, 32, 12.1 | Panel 12.1’s 7B preset is a dense transformer with hidden size 4,096 and 32 layers. Activations use bf16 at two bytes each. |
| Activation transfers | 4,096, 2 B, 8, 33.6 MB | At a two-satellite layer boundary, decode sends one hidden vector: 4,096 × 2 B = 8 KB per token. Prefill sends one vector per prompt token. A 4,096-token prompt therefore produces 33.6 MB. |
| Model and compute | 2, 32, 4,096, 2 B, 0.5 MB, 2.15 GB | The KV cache is larger than the activation by the number of layers and by the factor two for keys and values. Thus, each generated token adds 2 × 32 × 4,096 × 2 B = 0.5 MB to the satellite that holds the attention layers. A 4,096-token context occupies 2.15 GB. |
| Model and compute | Qualitative scope | These numbers are computed from the two architecture parameters and from the bf16 width, derived directly for this primer. They are the yardstick for the whole chapter. |
| Result and scope | Qualitative scope | The contrast with training is the point of the exercise. |
| Model and compute | 10, 7B, 28 GB, 8, 3.4 | Chapter 10’s data-parallel worker moves about twice the model bytes each step. A 7B bf16 model therefore moves 28 GB. Pipelined decode sends 8 KB per boundary, about 3.4 million times less. |
| Capacity and rate | Qualitative scope | At these rates, token-level inference uses little bandwidth. |
| Capacity and rate | Qualitative scope | A bandwidth bottleneck requires prefill-sized or image-sized tensors, or a link far slower than a laser. |
| 結果與適用範圍 | 定性範圍 | 位元組數只由架構決定。 |
| 結果與適用範圍 | 4,096, 32, 7B, 12.1 | 取一個隱藏維度 4,096、32 層的稠密 7B 級 transformer（也就是面板 12.1 的 7B 預設組），激活值以每個兩位元組的 bf16 儲存。 |
| 結果與適用範圍 | 4,096, 2 B, 8 | 當模型在某個層邊界切到兩顆衛星上，解碼時跨越鏈路的張量是一個隱藏向量，即每 token 4,096 × 2 B = 8 KB。 |
| 結果與適用範圍 | 4,096, 33.6 MB | 預填時則是每個提示 token 各一個這樣的向量，4,096 個 token 的提示就是 33.6 MB。 |
| 結果與適用範圍 | 2, 32, 4,096, 2 B, 0.5 MB, 2.15 GB | KV 快取比激活值大上「層數」倍再乘上 key 與 value 的因子二，因此每產生一個 token 就在持有注意力層的衛星上多出 2 × 32 × 4,096 × 2 B = 0.5 MB，4,096 個 token 的上下文佔 2.15 GB。 |
| 結果與適用範圍 | 定性範圍 | 這些數字是由兩個架構參數與 bf16 寬度算出來的，由本教材直接推導，而它們是整章的量尺。 |
| 結果與適用範圍 | 定性範圍 | 與訓練的對比才是這個練習的重點。 |
| 結果與適用範圍 | 10, 7B, 28 GB, 8, 340 | 第 10 章推導出資料平行的工作節點每步大約搬運兩倍的模型位元組，對同一個 bf16 的 7B 模型來說是 28 GB，而一個管線化的解碼步驟每個邊界只搬 8 KB，比值約 340 萬倍。 |
| 結果與適用範圍 | 定性範圍 | 在這些速率下，逐 token 的跨衛星推論只佔用少量頻寬。 |
| 結果與適用範圍 | 定性範圍 | 頻寬成為瓶頸的條件是預填或影像大小的張量，或遠慢於雷射的鏈路。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Activation bytes per decode token per layer boundary (7B-class, hidden 4,096, bf16)每個解碼 token 跨一個層邊界的激活值位元組（7B 級，隱藏維度 4,096，bf16） | \(4{,}096 \times 2\,\text{B}\) | 8,192 B ≈ 8 KBcomputed directly from the architecture直接由架構計算 |
| Same tensor during prefill of a 4,096-token prompt預填 4,096 個 token 的提示時的同一張量 | \(4{,}096 \times 8{,}192\,\text{B}\) | 33.6 MBcomputed. Sequence length matches SpaceMoE §VII-A3計算值。序列長度與 SpaceMoE §VII-A3 相同 |
| KV cache per token per layer每個 token 每層的 KV 快取 | \(2 \times 4{,}096 \times 2\,\text{B}\) | 16 KBcomputed. Factor 2 for keys and values計算值。因子 2 來自 key 與 value |
| KV cache per token across 32 layers每個 token 跨 32 層的 KV 快取 | \(32 \times 16\,\text{KB}\) | 512 KB ≈ 0.5 MBcomputed計算值 |
| KV cache for a 4,096-token context4,096 個 token 上下文的 KV 快取 | \(4{,}096 \times 524{,}288\,\text{B}\) | 2.15 GBcomputed計算值 |
| FLOPs per decode token, 7B dense7B 稠密模型每個解碼 token 的 FLOPs | \(2 \times 7\times10^{9}\) | 14 GFLOPcomputed. The 2 FLOP per parameter per token rule, attention over the context excluded計算值。每參數每 token 2 FLOP 的規則，上下文注意力列於計算範圍之外 |
| Training contrast: ring all-reduce bytes per worker per step, 7B in bf16, large \(N\)訓練對照：bf16 7B 模型環狀 all-reduce 每步每個工作節點的位元組，\(N\) 很大時 | \(\tfrac{2(N-1)}{N}\times 14\,\text{GB}\) | ≈ 28 GBcomputed with the Section 10.1 derivation. Ratio to 8 KB ≈ 3.4 × 10⁶以第 10.1 節的推導計算。與 8 KB 的比值 ≈ 3.4 × 10⁶ |
The KV cache’s location determines the system’s data paths. The satellite that prefills its attention layers stores the cache, adds 0.5 MB per generated token, and releases it when the request ends. Moving KV costs more than moving the activation across a boundary. In this architecture, one token’s KV is sixty-four times its activation. A whole context requires gigabytes.
Figure 12.1 draws the two phases side by side and marks the one place where the traffic is heavy, which is the prefill sweep. The one place where state accumulates, which is the KV store on the attention satellite. The evaluated systems keep KV at its creation site. SkyMemory proposes remote KV caching, whose transfer cost remains open. Section 12.6 returns to what that implies for serving ground users.
KV 快取的位置決定了系統的資料路徑。執行相應注意力層預填的衛星保存快取，每產生一個 token 增加 0.5 MB，請求結束時釋放。相對於跨越邊界的激活值，搬動它是成本較高的，因為在這個架構下一個 token 的 KV 是一個 token 激活值的六十四倍，而搬一整段上下文意味著搬動數 GB。圖 12.1 把兩個階段並排畫出，標出流量唯一沉重的地方（預填掃過）與狀態唯一累積的地方（注意力衛星上的 KV 儲存）。本章每個系統要麼讓 KV 留在它誕生的地方，要麼像 SkyMemory 那樣提議把它快取到其他位置，傳輸成本則待量化。第 12.6 節會回到這對服務地面使用者意味著什麼。
Single-satellite inference keeps activations and KV local. Flown demonstrations use this structure, while distributed systems add inter-satellite traffic assumptions that require separate evaluation.
單衛星推論將激活值與 KV 留在本地，已飛行的示範採用這種結構。分散式系統額外加入星間流量假設，需要另外評估。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | This arithmetic is also why single-satellite inference is feasible and has flown while multi-satellite inference awaits flight demonstration. |
| Model and compute | Qualitative scope | When the model fits one satellite’s memory, activations and KV state stay local. Van Berkel finds the same computation roofline as for an equivalent terrestrial server [101, §IV-D] preprint. |
| Power and thermal conditions | 9.4, 1, 2.1 MB, 325 ms, 1.8 W, 2, 2025 | Section 9.4’s flown systems use this local structure. The Φ-Sat-1 runs a 2.1 MB CNN on Myriad 2 in 325 ms at 1.8 W per inference [87, Tab. 5]. Starcloud-1 runs Gemma-family inference on one H100. Its throughput, power and thermal figures await publication (NVIDIA and DCD, Dec 2025) company. |
| Model and compute | 8B, 2025 | and the Three-Body constellation claims an 8B-parameter model per satellite (Live Science, May 2025) company via press. |
| Latency comparison | 22–183 ms, 13–90 ms | Before a model flew, Celestial compared ground and LEO-edge inference services in emulation. Moving service into orbit reduced latency from 22–183 ms to 13–90 ms [79]. |
| Orbital configuration | Qualitative scope | The chapter's five systems are the papers that take the next step and split a model across satellites. The question for each is what traffic their split creates and what link they assume carries it. |
| 結果與適用範圍 | 定性範圍 | 這套算術也解釋了為什麼單衛星推論可行且已經飛過，多衛星推論則待在軌示範。 |
| 結果與適用範圍 | 定性範圍 | 只要模型裝得進一顆衛星的記憶體，所有激活值都留在本地，KV 快取留在本地，運算的 roofline 與地面相同。 |
| 結果與適用範圍 | 定性範圍 | 這正是 van Berkel 的發現：侷限在單一衛星內的推論與地面伺服器有相同的 roofline[101, §IV-D]preprint。 |
| 結果與適用範圍 | 9.4, 1, 2, 325 ms, 1.8 W, 2.1 MB, 2025, 12, 8B, 5 | 第 9.4 節的飛行紀錄全是這一類：Φ-Sat-1 在 Myriad 2 上以每次推論 325 ms、1.8 W 跑一個 2.1 MB 的 CNN[87, Tab. 5]，Starcloud-1 在單一 H100 上跑 Gemma 系列模型，吞吐量、功率與熱學數字待公開（NVIDIA 與 DCD，2025 年 12 月）company，三體計算星座則宣稱每顆衛星一個 8B 參數模型（Live Science，2025 年 5 月）company via press。 |
| 結果與適用範圍 | 22–183 ms, 13–90 ms | 在任何衛星載著模型之前，Celestial 模擬平台就已經顯示把邊緣推論服務從地面搬到 LEO 邊緣，可以把模擬延遲從 22–183 ms 降到 13–90 ms[79]。 |
| 結果與適用範圍 | 定性範圍 | 本章的五個系統是走出下一步、把模型切到多顆衛星上的論文，而對每一個要問的是：它們的切法產生什麼流量，又假設什麼鏈路來承載。 |
12.2SpaceMoE: experts across 1,056 satellitesSpaceMoE：把專家散布在 1,056 顆衛星上
Recall
回顧
Section 2.1 computed intra-plane hop lengths from \(2(R_E+h)\sin(\pi/N)\) and Section 8.4 introduced the rad-hard processors that fly today.
SpaceMoE distributes expert layers across a constellation and routes each hidden state to selected experts. Its processor and simplified link model determine the reported latency regime.
SpaceMoE 將專家層分散到星座，並將隱藏狀態送往選中的專家。處理器與簡化鏈路模型決定了結果所處的延遲區間。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Model and compute | Qualitative scope | SpaceMoE [78] preprint spreads one model over a whole constellation. |
| Model and compute | Qualitative scope | A mixture of experts (MoE) splits each transformer layer’s feed-forward block into experts. A router activates only a few experts per token. Resident parameter count can therefore greatly exceed exercised parameter count. |
| Model and compute | 3.5B, 3.5 B, 6.7 B, 32, 8, 2, 36.3, 4,096 | LLaMA-MoE-3.5B has about 3.5 B active parameters from 6.7 B total. Its 32 MoE layers each contain 8 experts with top-2 routing. The stated forward pass costs 36.3 TFLOPs at sequence length 4,096 (§VII-A2–A3). |
| Capacity and rate | 33, 32, 1,056, 550 km, 87°, 100 | The polar Walker configuration has 33 planes × 32 satellites, totaling 1,056, at 550 km and 87° (§VII-A1). Each satellite has up to four laser ISLs: two intra-orbit and two inter-orbit. The paper specifies “≥ 100 Gbps” (§II-B). |
| Latency comparison | 5, 6 | Latency combines chord-geometry propagation (Eq. 5) with transmission term \(MQ_B/R\) (Eq. 6). The paper describes propagation as dominant. Queueing and per-hop processing are outside the model. |
| Formation geometry | 0.12 rad, 2, 0.95, 3 | A link requires satellite angular-rate difference below 0.12 rad/s (Eq. 2). Every link independently survives each snapshot with Bernoulli probability 0.95 (Eq. 3). |
| Reliability and accuracy | 2, 10.4 GFLOPS, 7.28 GFLOPS, 70% | Each satellite uses a Frontgrade SBC-2A72 radiation-hardened board. Peak compute is 10.4 GFLOPS, reduced to 7.28 GFLOPS effective at 70% utilization (§VII-A1, Table I). |
| 結果與適用範圍 | 定性範圍 | SpaceMoE[78]preprint把一個模型分散到整個星座上。 |
| 結果與適用範圍 | 定性範圍 | 它的模型是混合專家（MoE）：每一層的前饋區塊被切成數個專家，路由器對每個 token 只啟用其中少數幾個，因此常駐的參數量可以遠大於實際動用的參數量。 |
| 結果與適用範圍 | 3.5B, 6.7 B, 3.5 B, 32, 8, 2, 4,096, 36.3 | 實驗使用 LLaMA-MoE-3.5B，總參數 6.7 B 中約 3.5 B 為活躍參數，32 個 MoE 層、每層 8 個專家、top-2 路由，在序列長度 4,096 下一次前向傳遞為 36.3 TFLOPs（§VII-A2–A3）。 |
| 容量與速率 | 33, 32, 1,056, 550 km, 87°, 100 | 星座是極軌 Walker 構型，33 個軌道面 × 32 顆 = 1,056 顆衛星，高度 550 km、傾角 87°（§VII-A1），每顆衛星最多四條雷射 ISL（兩條同軌、兩條跨軌），速率寫作「≥ 100 Gbps」（§II-B）。 |
| 結果與適用範圍 | 5, 6 | 延遲模型是弦長幾何的傳播延遲（Eq. 5）加上傳輸項 \(MQ_B/R\)（Eq. 6），論文自己說由傳播主導，排隊與每跳處理延遲列於模型範圍之外。 |
| 可靠度與準確率 | 0.12 rad, 2, 0.95, 3 | 只有當兩顆衛星的角速率差小於 0.12 rad/s 時鏈路才存在（Eq. 2），而鏈路在每個快照中以獨立的 Bernoulli 事件、機率 0.95 存活，所有鏈路相同（Eq. 3）。 |
| 結果與適用範圍 | 2, 10.4 GFLOPS, 70%, 7.28 GFLOPS | 每顆衛星的運算是 Frontgrade SBC-2A72 抗輻射板，峰值 10.4 GFLOPS，以 70% 使用率計為有效 7.28 GFLOPS（§VII-A1，Table I）。 |
Two-level expert placement minimizes the expected path traveled by a token’s hidden state. First, each MoE layer receives a contiguous ring-based satellite subnet with a central gateway (Eqs. 17–18). Tokens cross layers by hopping between neighboring subnets. Second, activation frequency assigns experts within each subnet (Eqs. 19–20, 29–30), minimizing expected per-layer latency (Eqs. 33–34).
Theorem 1 assigns the \(i\)-th most frequently activated expert to the \(i\)-th lowest-latency satellite. This optimum takes \(O(I\log I + V\log V)\) per layer. Figure 12.2 illustrates the scheme.
Across eight lm-eval-harness datasets, SpaceMoE reports 1.02–1.07 s per token. Random placement within a contiguous region achieves 3.34–3.37 s. Random intra-orbit placement takes 4.13–4.16 s, and fully random placement takes 5.28–5.30 s (Table II, Figs. 6–7). Evaluation samples simulated topology snapshots, with accuracy measurements and a hardware testbed still open (§VII).
貢獻是一個兩層的專家配置，也就是把專家指派到衛星上，使 token 的隱藏狀態穿過各層時走的期望路徑最短。第一層把星座切成以環為基礎的層子網，每個 MoE 層對應一組相鄰的衛星，中央是一顆閘道衛星（Eqs. 17–18），token 逐層前進時就在相鄰子網之間跳躍。第二層在子網內依啟用頻率把專家指派到衛星（Eqs. 19–20、29–30），最小化每層的期望延遲（Eqs. 33–34）。Theorem 1 證明最佳解是把第 \(i\) 常被啟用的專家交給延遲第 \(i\) 低的衛星，每層可在 \(O(I\log I + V\log V)\) 內算出。圖 12.2 畫出這個方案。
在八個 lm-eval-harness 資料集上，SpaceMoE 的結果是每 token 1.02–1.07 s，對照連續區域內最佳隨機配置的 3.34–3.37 s、同軌隨機配置的 4.13–4.16 s 與完全隨機配置的 5.28–5.30 s（Table II，Figs. 6–7）。評估是在抽樣的拓樸快照上做模擬，準確度結果與實體測試平台仍待補充（§VII）。
Original figure note原始圖說出處
Drawn for this primer from Wang et al., arXiv 2605.00515, §IV–V and §VII-A. Hop length computed from the 33 × 32 geometry at 550 km.SpaceMoE’s stated compute inputs explain most of its time per token. Faster processors would expose propagation delay while transmission of the small hidden state remains comparatively cheap.
SpaceMoE 的運算輸入可解釋大部分每 token 時間。處理器加快後，傳播延遲會顯現，小型隱藏狀態的傳輸成本仍相對較低。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The paper’s inputs reveal the network’s share of the reported result. |
| Model and compute | 3.5 B, 2 FLOP, 7 GFLOP, 7.28 GFLOPS, 0.96 s, 90–94%, 1.02–1.07 s | Each decode token activates 3.5 B parameters. At 2 FLOP per parameter, this gives 7 GFLOP. Dividing by 7.28 GFLOPS gives 0.96 s compute per token, already 90–94% of the reported 1.02–1.07 s. |
| Result and scope | 36.3, 4,096, 8.9 GFLOP, 1.22 s | Alternatively, the stated 36.3 TFLOPs forward pass divided by 4,096 tokens gives 8.9 GFLOP per token. Compute alone would then take 1.22 s, exceeding the reported total. |
| Capacity and rate | 100 | The two estimates bracket the result and both say the same thing: the time per token is set by the rad-hard processor. The "≥ 100 Gbps" link is barely exercised. |
| Latency comparison | 32, 550 km, 1,357 km, 4.5 ms, 0.1 s | A 32-satellite ring at 550 km has a 1,357 km intra-plane hop, or 4.5 ms propagation. At most about 0.1 s of reported time remains for traversing the 32-layer chain. This permits at most about two dozen such hops. |
| Latency comparison | Qualitative scope | The paper's statement that communication is dominated by propagation is therefore true but only because the compute is so slow. |
| Result and scope | 12.5 | Section 12.5 shows that Orin- or H100-class compute makes this placement propagation-bound. ISL rate would still have little effect. |
| 結果與適用範圍 | 定性範圍 | 論文的輸入值可以用來估計網路對結果的影響。 |
| 模型與運算 | 3.5 B, 2 FLOP, 7 GFLOP, 7.28 GFLOPS, 0.96 s, 1.02–1.07 s, 90–94% | 每個解碼 token 動用 3.5 B 活躍參數，依每參數 2 FLOP 的規則是 7 GFLOP，除以 7.28 GFLOPS 得到每 token 0.96 s 的運算時間，這已經是所報告 1.02–1.07 s 的 90–94%。 |
| 結果與適用範圍 | 36.3, 4,096, 8.9 GFLOP, 1.22 s | 若改用論文自己的前向傳遞 36.3 TFLOPs 除以 4,096 個 token 的序列，每個 token 要 8.9 GFLOP，光是運算就要 1.22 s，超過所報告的總時間。 |
| 容量與速率 | 100 | 兩個估計夾住了結果，而且說的是同一件事：每 token 時間由抗輻射處理器決定，「≥ 100 Gbps」的鏈路使用量極低。 |
| 軌道配置 | 550 km, 32, 1,357 km, 4.5 ms, 0.1 s | 550 km 高度、32 顆一環的同軌跳距是 1,357 km、傳播 4.5 ms，因此所報告總時間中最多約 0.1 s 可以分給穿過 32 層鏈的全部傳輸，相當於至多二十幾個這樣的跳數。 |
| 結果與適用範圍 | 定性範圍 | 論文說通訊由傳播主導，這句話是對的，但只因為運算太慢。 |
| 結果與適用範圍 | 12.5 | 第 12.5 節會顯示，換成 Orin 級或 H100 級的載荷，同樣的配置會變成傳播受限，而 ISL 速率仍然影響較小。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Compute per token from active parameters由活躍參數算出的每 token 運算時間 | \(2\times 3.5\times10^{9} / 7.28\times10^{9}\) | 0.96 scomputed. 3.5 B active and 7.28 GFLOPS from SpaceMoE §VII-A計算值。3.5 B 活躍參數與 7.28 GFLOPS 取自 SpaceMoE §VII-A |
| Compute per token from the paper's forward-pass FLOPs由論文前向傳遞 FLOPs 算出的每 token 運算時間 | \(36.3\times10^{12} / 4{,}096 / 7.28\times10^{9}\) | 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 |
| Share of the reported 1.02–1.07 s explained by compute所報告 1.02–1.07 s 中由運算解釋的比例 | \(0.96 / 1.07 \;\ldots\; 0.96/1.02\) | 90–94%computed. Table II values計算值。Table II 數值 |
| Intra-plane hop in a 32-satellite ring at 550 km550 km 高度 32 顆一環的同軌跳距 | \(2\times 6{,}921\,\text{km}\times\sin(\pi/32)\) | 1,357 km → 4.5 mscomputed with the Section 2.1 chord formula. \(R_E\) = 6,371 km以第 2.1 節的弦長公式計算。\(R_E\) = 6,371 km |
| 8 KB activation on a 100 Gb/s link8 KB 激活值在 100 Gb/s 鏈路上 | \(65{,}536\,\text{bit} / 10^{11}\,\text{bit/s}\) | 0.66 µscomputed. 1/6,900 of one 4.5 ms hop計算值。為一跳 4.5 ms 的 1/6,900 |
SpaceMoE’s model sizing, throughput statements and expert order differ across sections. The result table defines the safest quotation scope, while correlated failures and acquisition require added models.
SpaceMoE 各節的模型估算、加速說法與專家排序存在差異。引用應以結果表的設定為準，相關性故障與擷取則待加入模型。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 38 | The paper also carries internal discrepancies that a reader quoting it must navigate (fact base Part C, item 38). |
| Model and compute | 70, 140 GB, 6.7 B | Its system sizing in §IV-A and Table I is done for a Switch Transformer of "over 70 billion parameters" at 140 GB in FP16. In comparison, every experiment uses the 6.7 B model. |
| Result and scope | Qualitative scope | The abstract and Table II imply at least a threefold reduction against the best baseline while §VII-B says "at least a twofold" reduction. |
| Result and scope | 10 GFLOPS, 7.28 GFLOPS, 10.4 | Table I lists 10 GFLOPS where the text derives 7.28 GFLOPS from 10.4. |
| Result and scope | 1 | And the introduction orders experts by ascending activation probability where Theorem 1 uses descending. |
| Model and compute | 6.7 B | The safe quotation is the Table II time-per-token range with the 6.7 B model. |
| Reliability comparison | 5%, 0.057% | The failure model assumes independent 5% link loss per snapshot. This is two orders above a production training fabric’s 0.057% monthly failure rate [65, §2.3]. |
| Result and scope | 2–30 s, 2 | Inference can reroute hidden states under that loss model. Correlated outages, 2–30 s reacquisition and pointing losses from Chapter 2 remain outside its scope. |
| 結果與適用範圍 | 38 | 這篇論文也帶著引用者必須小心的內部差異（事實庫 Part C 第 38 項）。 |
| 結果與適用範圍 | 700, 140 GB, 6.7 B | §IV-A 與 Table I 的系統規模估算是針對一個「超過 700 億參數」、FP16 下 140 GB 的 Switch Transformer，而所有實驗用的是 6.7 B 模型。 |
| 結果與適用範圍 | 定性範圍 | 摘要與 Table II 暗示相對最佳基準至少三倍的降幅，§VII-B 卻寫「至少兩倍」。 |
| 結果與適用範圍 | 10 GFLOPS, 10.4, 7.28 GFLOPS | Table I 列 10 GFLOPS，內文卻從 10.4 推得 7.28 GFLOPS。 |
| 結果與適用範圍 | 1 | 引言把專家依啟用機率遞增排序，Theorem 1 用的卻是遞減。 |
| 結果與適用範圍 | 6.7 B | 安全的引用方式是 Table II 中 6.7 B 模型的每 token 時間區間。 |
| 結果與適用範圍 | 5%, 0.057% | 更有影響的建模選擇是失效模型：每個快照 5% 的獨立鏈路失效，比生產級訓練網路每月 0.057% 的鏈路失效率[65, §2.3]差了兩個數量級。 |
| 結果與適用範圍 | 2, 2–30 s | 對一個會重新路由隱藏狀態的推論系統而言這可以容忍，但其範圍之外還有第 2 章為真實雷射終端記錄的相關性中斷、2–30 s 的重新擷取時間與指向損失。 |
12.3CollabLLM-LEO: a pipeline over five Orin satellites at 0.5 Gbit/sCollabLLM-LEO：五顆 Orin 衛星上 0.5 Gbit/s 的管線
CollabLLM-LEO partitions vision transformers across a satellite pipeline. A learned sparse mask, quantization and entropy coding reduce the activation passed at each boundary.
CollabLLM-LEO 將視覺 transformer 切分到衛星管線，透過學習式稀疏遮罩、量化與熵編碼，減少每個邊界傳送的激活值。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | CollabLLM-LEO [96] preprint answers the opposite question. |
| Result and scope | Qualitative scope | Where SpaceMoE assumes a fast link and a slow processor, CollabLLM-LEO assumes a fast processor and a slow link. Its contribution is to shrink what crosses it. |
| Orbital configuration | 0, 1 | ISLs chain sensing satellite \(S_0\) with \(K\) compute satellites \(S_1 \ldots S_K\). The model is divided layer-wise into \(K\) sub-models. Shortest-path search over a directed acyclic graph chooses among candidate cuts (§III-A, §V). Multithreading overlaps computation and transmission (§III-B3). |
| Model and compute | 0.086 B, 2 GB, 1.8 B, 12 GB | The vision transformers classify Earth-observation images on EuroSAT and RESISC45. ViT-B has 0.086 B parameters and about 2 GB footprint. ViT-G reaches 1.8 B parameters and exceeds 12 GB (Table III). |
| Power and thermal conditions | 1–5, 8, 6 | Each boundary applies activation compression. A learned Gumbel-Sigmoid mask sparsifies the tensor using a straight-through estimator, sparsity loss and temperature annealing (Eqs. 1–5). Surviving values are quantized to 8 bits (Eq. 6), then entropy-coded. |
| Result and scope | 3.96, 11.56, 25.82, 25.60, 24.14, 23.28, 20, 80%, 8 | On the first boundary the cumulative ratio is 3.96× after the mask, 11.56× after quantisation and 25.82× after entropy coding. The other boundaries reach 25.60×, 24.14× and 23.28×. Thus, every boundary compresses by more than 20× at 80% sparsity (Fig. 8, Tables IV–V). |
| 結果與適用範圍 | 定性範圍 | CollabLLM-LEO[96]preprint回答的是相反的問題。 |
| 結果與適用範圍 | 定性範圍 | SpaceMoE 假設鏈路快、處理器慢，CollabLLM-LEO 則假設處理器快、鏈路慢，它的貢獻是縮小跨越鏈路的東西。 |
| 結果與適用範圍 | 0, 1 | 系統以 ISL 把一顆感測衛星 \(S_0\) 與 \(K\) 顆運算衛星 \(S_1 \ldots S_K\) 串成一條鏈，把模型逐層切成 \(K\) 個子模型，切點由候選切割構成的有向無環圖上的最短路徑搜尋決定（§III-A，§V），運算以多執行緒與傳輸重疊（§III-B3）。 |
| 結果與適用範圍 | 0.086 B, 2 GB, 1.8 B, 12 GB | 模型是用於地球觀測分類的視覺 transformer ，處理對象是影像，從 0.086 B 參數、約 2 GB 的 ViT-B 到 1.8 B 參數、超過 12 GB 的 ViT-G（Table III），資料集為 EuroSAT 與 RESISC45。 |
| 結果與適用範圍 | 1–5, 8, 6 | 每個邊界都有一個激活值壓縮階段：先用以 straight-through 估計器、稀疏損失與溫度退火訓練出的 Gumbel-Sigmoid 遮罩稀疏化張量（Eqs. 1–5），再把留下的元素量化為 8 位元（Eq. 6），最後做熵編碼。 |
| 結果與適用範圍 | 3.96, 11.56, 25.82, 25.60, 24.14, 23.28, 80%, 20, 8 | 第一個邊界的累積壓縮比在遮罩後為 3.96×、量化後 11.56×、熵編碼後 25.82×，其餘邊界達 25.60×、24.14× 與 23.28×，因此在 80% 稀疏度下每個邊界都壓縮超過 20×（Fig. 8，Tables IV–V）。 |
The evaluation combines real Orin timing with simulated orbital links. Its ground baseline processes images captured in orbit, so the result concerns image downlink versus on-board processing.
評估結合實體 Orin 計時與模擬軌道鏈路。地面基準處理的是在軌拍攝影像，因此結果比較了影像下傳與機上處理。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 275 TOPS, 15, 30, 50 W, 8 GB | Compute timing comes from four physical NVIDIA Jetson AGX Orin modules, each rated up to 275 TOPS. Tests use 15, 30 and 50 W modes, with 8 GB memory per compute satellite. Orbits, visibility and links are simulated (§VI-A, Table II). |
| Capacity and rate | 0.5 Gbit/s, 6 Gbit/s, 0.2–0.8 Gbit/s, 4 | The five-satellite chain comes from twelve satellites in one Walker-Delta plane. ISL rate is 0.5 Gbit/s. Table II’s ground link is 6 Gbit/s, while Fig. 4 sweeps 0.2–0.8 Gbit/s. The paper leaves this difference unexplained. |
| Latency comparison | 42%, 4070, 800 Mbit/s, 100%, 4, 71%, 1% | At 800 Mbit/s ground rate, the paper reports 42% lower delay than downlinking raw images to an RTX 4070 Ti server. It claims “over 100%” at lower ground rates (Fig. 4). Communication drops 71%, with accuracy loss below 1%. |
| Result and scope | 194, 200, 1.0%, 10 | with more than 194 of 200 split strategies within 1.0% of the original baseline (Fig. 10). |
| Result and scope | 12.3 | Figure 12.3 draws the chain and the baseline it is compared against. |
| Result and scope | 12.6 | This ground baseline processes images created on the sensing satellite, requiring downlink before ground processing. Section 12.6 asks a separate question: where to serve ground-originated requests. |
| 結果與適用範圍 | 275 TOPS, 15, 30, 50 W, 8 GB | 評估是混合式的：運算時間在四片 NVIDIA Jetson AGX Orin 模組上量測（每片最高 275 TOPS，15、30、50 W 三種功率模式，每顆運算衛星 8 GB 記憶體），而軌道、可見性與鏈路則是模擬（§VI-A，Table II）。 |
| 容量與速率 | 0.5 Gbit/s, 6 Gbit/s, 4, 0.2–0.8 Gbit/s | 五顆衛星的鏈取自單一軌道面、十二顆衛星的 Walker Delta，ISL 速率 0.5 Gbit/s，地面鏈路在 Table II 是 6 Gbit/s、在 Fig. 4 的掃描中卻是 0.2–0.8 Gbit/s，這項差異的原因仍待論文說明。 |
| 結果與適用範圍 | 800 Mbit/s, 4070, 42%, 100%, 4, 71%, 1%, 200, 194, 1.0%, 10 | 主結果是：當地面鏈路為 800 Mbit/s 時，相較於把原始影像下傳到配備 RTX 4070 Ti 的地面伺服器，端到端延遲低 42%，在更低的地面鏈路速率下增長到「超過 100%」（Fig. 4），同時通訊量少 71%、準確度損失低於 1%，200 種切割策略中有超過 194 種與原始基準相差低於 1.0%（Fig. 10）。 |
| 結果與適用範圍 | 12.3 | 圖 12.3 畫出這條鏈與它所比較的基準。 |
| 結果與適用範圍 | 12.6 | 這個基準對第 12.6 節很重要：它確實是一個地面端的替代方案，但針對的是誕生在感測衛星上、必須先下傳才能在任何地方處理的資料，這與「源自地面的請求該在哪裡服務」各有問題範圍。 |
Original figure note原始圖說出處
Drawn for this primer from Zhang et al., arXiv 2604.04654, §III, §VI-A, Tables II–III and Figs. 4 and 8.Compression reduces transmission time while propagation stays fixed. Image-sized tensors explain its benefit, and faster hardware makes the link assumptions more influential.
壓縮縮短傳輸時間，傳播時間則維持固定。影像大小的張量解釋了其效益，較快硬體也會提高鏈路假設的影響。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 12.1 | Two observations follow from reading the design against Section 12.1. |
| Result and scope | Qualitative scope | The first is that compression attacks the transmission term and only that term. |
| Latency comparison | 0.5 Gbit/s, 8, 131 µs, 25.82, 5.1 µs, 659 km, 2.2 ms | At 0.5 Gbit/s, an 8 KB decode activation takes 131 µs, reduced to 5.1 µs by 25.82× compression. One 659 km hop still adds 2.2 ms propagation. Token-sized compression therefore targets the smaller communication term. |
| Result and scope | 42% | The 42% gain concerns image-sized tensors. Vision transformers emit one activation row per image patch, and the baseline downlinks raw images at a few hundred Mbit/s. |
| Model and compute | 275 TOPS, 3.8, 10, 7.28 GFLOPS | The two papers also expose a compute gap. Orin’s 275 TOPS peak is 3.8 × 10⁴ times SpaceMoE’s 7.28 GFLOPS radiation-hardened board. |
| Latency comparison | Qualitative scope | so the same ViT-G forward pass that would take seconds on one board takes tens of microseconds on the other |
| Model and compute | Qualitative scope | and an author who swaps one assumption for the other moves the bottleneck from compute to the link with the same network model. |
| Latency comparison | 39, 6, 0.2–0.8 Gbit/s, 103%, 12 | Fact-base item 39 lists the paper’s discrepancies. Ground rate is 6 versus 0.2–0.8 Gbit/s. Figure 12 claims a delay reduction of “103%”, exceeding the metric’s definition. Model names alternate between ViT-B/L/H/G and ViT-Tiny/Base/Large. |
| Reliability and accuracy | 4.1%, 0.3% | and a Top-k accuracy loss of 4.1% in the text against 0.3% in Table IV |
| Result and scope | 42%, 800 Mbit/s, 20 | so the safe quotations are the 42% at 800 Mbit/s and the greater-than-20× compression. |
| 結果與適用範圍 | 12.1 | 把這個設計對照第 12.1 節讀，有兩個觀察。 |
| 結果與適用範圍 | 定性範圍 | 第一，壓縮攻擊的是傳輸項，而且只有傳輸項。 |
| 容量與速率 | 0.5 Gbit/s, 8, 131 µs, 25.82, 5.1 µs, 659 km, 2.2 ms | 在 0.5 Gbit/s 下，原始的 8 KB 解碼激活值要 131 µs，壓縮 25.82× 後是 5.1 µs，而一跳 659 km 的同軌鏈路有 2.2 ms 的傳播延遲，壓縮後仍保持固定。 |
| 結果與適用範圍 | 定性範圍 | 所以對 token 大小的張量而言，這個設計處理的是兩個通訊項中較小的那個。 |
| 結果與適用範圍 | 42% | 它能賺到 42%，是因為它的張量是影像大小（視覺 transformer 每個影像區塊，每個 token 則有相異的輸出尺度 輸出一列激活值），也因為它的基準是以每秒幾百 Mbit 下傳原始影像。 |
| 模型與運算 | 275 TOPS, 7.28 GFLOPS, 3.8, 10 | 第二個觀察是它暴露出兩篇論文之間的運算差距：Orin 的 275 TOPS 峰值是 SpaceMoE 抗輻射板 7.28 GFLOPS 的 3.8 × 10⁴ 倍，同樣的 ViT-G 前向傳遞在一塊板上要幾秒、在另一塊上只要幾十微秒，因此作者只要把一個假設換成另一個，就能在網路模型維持相同的情況下把瓶頸從運算搬到鏈路。 |
| 容量與速率 | 39, 6, 0.2–0.8 Gbit/s, 12, 103%, 4.1%, 0.3% | 論文中的差異（事實庫 Part C 第 39 項）包括 6 對 0.2–0.8 Gbit/s 的地面鏈路、Fig. 12 中對延遲而言超出延遲降幅定義的「103% 降幅」、在 ViT-B/L/H/G 與 ViT-Tiny/Base/Large 之間切換的模型名稱，以及內文 4.1% 對 Table IV 0.3% 的 Top-k 準確度損失。 |
| 結果與適用範圍 | 800 Mbit/s, 42%, 20 | 安全的引用是 800 Mbit/s 下的 42% 與超過 20× 的壓縮比。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| 8 KB activation on a 0.5 Gbit/s ISL, original8 KB 激活值在 0.5 Gbit/s ISL 上，原始 | \(65{,}536 / 5\times10^{8}\) | 131 µscomputed. Rate from CollabLLM-LEO Table II計算值。速率取自 CollabLLM-LEO Table II |
| Same tensor after 25.82× compression同一張量經 25.82× 壓縮後 | \(131\,\mu\text{s} / 25.82\) | 5.1 µscomputed. Ratio from Fig. 8計算值。比例取自 Fig. 8 |
| Propagation over one 659 km intra-plane hop, untouched by compression一跳 659 km 同軌鏈路的傳播延遲，壓縮後維持固定 | \(659\,\text{km}\times 3.336\,\mu\text{s/km}\) | 2.2 mscomputed. Hop length from Section 2.1 (24 × 66 geometry)計算值。跳距取自第 2.1 節（24 × 66 幾何） |
| Compute per token for a 1.8 B model at the 275 TOPS peak1.8 B 模型在 275 TOPS 峰值下的每 token 運算時間 | \(2\times 1.8\times10^{9} / 2.75\times10^{14}\) | 13 µscomputed. Peak integer TOPS treated as an upper bound on throughput計算值。峰值整數 TOPS 視為吞吐量上限 |
| Compute ratio between the two papers' hardware兩篇論文硬體的運算比值 | \(2.75\times10^{14} / 7.28\times10^{9}\) | 3.8 × 10⁴computed. Orin peak vs SBC-2A72 effective計算值。Orin 峰值對 SBC-2A72 有效值 |
12.4LLMSpace, SkyMemory, Constella and the cached model: the rest of the setLLMSpace、SkyMemory、Constella 與快取模型：其餘系統
LLMSpace estimates embodied and operational carbon for orbital inference hardware. Its link parameters enter the hardware accounting, while token-path latency remains outside the model.
LLMSpace 估算軌道推論硬體的體現碳排與營運碳排。鏈路參數用於硬體計算，逐 token 路徑延遲則列於模型範圍之外。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Related evaluation scope | Qualitative scope | Other papers study carbon, cached-state placement, satellite roles and satellite relays. Their per-token paths remain unquantified. |
| Model and compute | Qualitative scope | LLMSpace [97] preprint is a bottom-up carbon model for LLM inference on a LEO satellite. |
| Power and thermal conditions | 130 kg, 4 nm, 10 kW, 14 nm, 20 kW, 191 kg, 4.1, 5 | LLMSpace assumes a roughly 130 kg, eight-GPU DGX-H100-class payload. The 4 nm COTS version uses 10 kW and lasts two years, limited by total dose and single-event effects. The 14 nm radiation-hardened version uses 20 kW and lasts ten years. Both mount on a Starlink-V1 bus modeled at 191 kg, a mass combination requiring explanation (§4.1, §5). |
| Result and scope | 200, 1 W, 28 kg, 150, 96/32 Gbit/s, 90–100, 30–35, 80%, 4.1 | Laser terminals provide “up to 200 Gbps at ~1 W optical power”. Each weighs about 28 kg and accounts for 150 kgCO2e. Ground rates exceed 96/32 Gbit/s, with latency modeling open. The orbit lasts 90–100 min, including 30–35 min eclipse. Batteries use 80% depth of discharge (§4.1). |
| Orbital configuration | 16.57, 30.45, 14.5, 9, 6, 2, 20, 380 | Embodied carbon is 16.57 tCO2e for COTS and 30.45 tCO2e for radiation-hardened hardware. Falcon 9 launch intensity is 14.5 kgCO2e per kilogram (Table 6, Fig. 2). Orbital operational carbon is described as “effectively negligible” against grids at 20 and 380 gCO2e/kWh. |
| Dataset and workload | 40, 15, 2, 10–20 | Fact base Part C item 40 records its inconsistent rad-hard area overheads of 15×, 2× and 10–20×. Thus, the embodied totals are the safe quotation. |
| 結果與適用範圍 | 定性範圍 | 其餘論文問的是碳排、快取狀態該放哪裡、衛星角色，以及衛星作為中繼，而每 token 路徑的數值仍待補充。 |
| 結果與適用範圍 | 定性範圍 | LLMSpace[97]preprint是 LEO 衛星上 LLM 推論的由下而上碳排模型。 |
| 電力與熱條件 | 130 kg, 4 nm, 10 kW, 14 nm, 20 kW, 191 kg, 4.1, 5 | 它的載荷是約 130 kg 的 DGX-H100 級八 GPU 系統，或者是 4 nm COTS、10 kW、受總劑量與單事件效應限制壽命兩年，或者是 14 nm 抗輻射、20 kW、壽命十年，裝在它取為 191 kg 的 Starlink-V1 衛星平台上，這個質量組合的說明待補充（§4.1，§5）。 |
| 結果與適用範圍 | 200, 1 W, 28 kg, 150, 96/32 Gbit/s | 鏈路是「最高 200 Gbps、約 1 W 光功率」的雷射 ISL，每個終端約 28 kg、150 kgCO2e，地面鏈路高於 96/32 Gbit/s，延遲模型待建立。 |
| 結果與適用範圍 | 90–100, 30–35, 80%, 4.1 | 軌道週期 90–100 分鐘、地影 30–35 分鐘、電池以 80% 放電深度設計（§4.1）。 |
| 結果與適用範圍 | 16.57, 30.45, 9, 14.5, 6, 2, 20, 380 | 結果是體現碳排總量：COTS 版本 16.57 tCO2e，抗輻射版本 30.45 tCO2e，Falcon 9 的發射碳強度為每公斤 14.5 kgCO2e（Table 6，Fig. 2），並指出相對於 20 與 380 gCO2e/kWh 的地面電網，在軌營運碳排「實際上可忽略」。 |
| 結果與適用範圍 | 40, 15, 2, 10–20 | 事實庫 Part C 第 40 項記錄了它 15×、2× 與 10–20× 各有差異的抗輻射面積開銷，因此體現碳排總量是安全的引用。 |
SkyMemory proposes remote KV storage, Constella separates processing and communication roles, and the cached-model framework uses satellites as relays. Their evaluation scopes address separate system questions.
SkyMemory 提議遠端 KV 儲存，Constella 區分處理與通訊角色，快取模型框架則以衛星擔任中繼，各自評估的系統問題也相異。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 2505.14427 | SkyMemory directly addresses KV caching (Sandholm, Mukherjee, Cheng & Huberman, arXiv 2505.14427) preprint. It extends KV memory across LEO satellites connected by free-space optical ISLs. A neighboring satellite can then serve cache hits. |
| Result and scope | 19, 5 | Its KVC protocol is evaluated in simulation and on a proof-of-concept testbed. Five Intel NUCs host a 19 × 5 constellation, while a Jetson Nano hosts the LLM. |
| Cache performance | 12.5 | Its abstract leaves ISL rate, cache size and hit rate unspecified, so placement requires further details for the ladder in Section 12.5. It is the only system in the set whose traffic would be KV-sized, with much larger transfers than activations. |
| Orbital configuration | 2609.05427, 4, 1, 3.1 | Constella partitions DNNs across processor and communicator satellites (Stanisic, Gravara, Herrera & Nastic, arXiv 2609.05427) preprint. Processors carry sensors, compute, ISLs and downlink, at relative cost α = 4. Communicators carry ISLs and downlink only, at β = 1. Split-layer selection trades FLOPs against intermediate tensor size (§3.1). |
| Result and scope | 5,100 s, 300 s, 900 s, 6,300 s, 10 s, 625,000 bit/s, 100, 1, 3.1 | Each 6,300 s orbit has fixed phases: 5,100 s compute, 300 s communication and 900 s idle. Inference runs every 10 s, with 625,000 bit/s downlink. The paper calls ISL latency and energy “negligible” against ground transfer. It cites modern ISLs at “up to 100 Gbps” (§1, §3.1, Table 1). |
| Dataset and workload | 1, 1,541 s, 4,129 s, 3,518 s, 2.7, 74, 81.9%, 4.1 | Evaluation uses a BUPT-1 dataset and Jetson Orin Nano energy coefficients. Medium-scenario latency is 1,541 s, versus 4,129 s for naive processing and 3,518 s for bent pipe. Reported improvements reach 2.7× in latency and 74× in energy, with at least 81.9% inference success (§4.1, Table 1). |
| Result and scope | 52 | Its submission date and arXiv identifier disagree (fact base Part C, item 52). |
| Cached-model allocation | 2403.05826, 23% | “Cached Model-as-a-Resource” provisions LLM agents from terrestrial edge and cloud servers through satellite relays in a space-air-ground network (Xu et al., arXiv 2403.05826) preprint. It defines “age of thought” and least-age cache replacement. A DQN-based auction improves allocation efficiency by 23%. |
| Latency comparison | Qualitative scope | The satellites serve only as relays, and orbit, ISL and latency assumptions remain unspecified in the abstract. |
| Result and scope | Qualitative scope | A ToN version is indicated by a DOI record that awaits verification. |
| 結果與適用範圍 | 2505.14427 | SkyMemory（Sandholm、Mukherjee、Cheng 與 Huberman，arXiv 2505.14427）preprint是唯一直接碰到 KV 快取的提案：它把 KV 快取記憶體延伸到以自由空間光學 ISL 相連的 LEO 星座，讓快取命中可以由鄰近衛星提供。 |
| 結果與適用範圍 | 19, 5 | 其 KVC 協定以模擬以及一個概念驗證測試平台評估，平台由五台 Intel NUC 承載 19 × 5 的星座、以 Jetson Nano 作為 LLM 主機。 |
| 結果與適用範圍 | 12.5 | 摘要的 ISL 速率、快取大小與命中率待補充，取得這些資訊後才能將提案放入第 12.5 節的比較，而它是整組中唯一流量會是 KV 大小，傳輸量高於激活值的系統。 |
| 結果與適用範圍 | 2609.05427, 4, 1, 3.1 | Constella（Stanisic、Gravara、Herrera 與 Nastic，arXiv 2609.05427）preprint把 DNN 切到兩種角色的衛星上：處理者帶感測器、運算、ISL 與下行鏈路，通訊者只帶 ISL 與下行鏈路，相對成本 α = 4 與 β = 1，並以 FLOPs 對中間張量大小的權衡選擇切割層（§3.1）。 |
| 結果與適用範圍 | 6,300 s, 5,100 s, 300 s, 900 s, 10 s | 它的軌道是固定的三個階段：6,300 s 的軌道中 5,100 s 運算、300 s 通訊、900 s 閒置，每 10 s 一次推論。 |
| 結果與適用範圍 | 625,000 bit/s | 下行鏈路 625,000 bit/s。 |
| 容量與速率 | 100, 1, 3.1 | ISL 的延遲與能耗相對地面段被宣告為「可忽略」，並引用現代 ISL「最高 100 Gbps」（§1，§3.1，Table 1）。 |
| 結果與適用範圍 | 1, 1,541 s, 4,129 s, 3,518 s, 2.7, 74, 81.9%, 4.1 | 在 BUPT-1 資料集上以 Jetson Orin Nano 的能耗係數評估，它報告中等情境的延遲為 1,541 s，對照天真方案的 4,129 s 與彎管的 3,518 s，延遲最多低 2.7×、能耗低 74×，推論成功率至少 81.9%（§4.1，Table 1）。 |
| 結果與適用範圍 | 52 | 它的投稿日期與 arXiv 編號互相矛盾（事實庫 Part C 第 52 項）。 |
| 模型與運算 | 2403.05826, 23% | 最後，「Cached Model-as-a-Resource」框架（Xu 等，arXiv 2403.05826）preprint在空天地整合網路中，從地面邊緣與雲端伺服器經衛星中繼供應 LLM 代理，定義「思考年齡」與最少年齡的快取替換策略，並報告以 DQN 為基礎的拍賣帶來 23% 的配置效率增益。 |
| 結果與適用範圍 | 定性範圍 | 其中衛星僅擔任中繼，摘要的軌道、ISL 與延遲假設待補充。 |
| 結果與適用範圍 | 定性範圍 | 一個 DOI 記錄指向 ToN 版本，驗證待完成。 |
| System系統 | Model模型 | Hardware硬體 | Constellation星座 | ISL assumptionISL 假設 | Latency model延遲模型 | Failure model失效模型 | Evaluation評估方式 | Result結果 | Internal inconsistencies (Part C)內部差異（Part C） |
|---|---|---|---|---|---|---|---|---|---|
| SpaceMoE ppSpaceMoE pp | LLaMA-MoE-3.5B (3.5 B active / 6.7 B), 32 layers × 8 experts, top-2LLaMA-MoE-3.5B（3.5 B 活躍 / 6.7 B），32 層 × 8 專家，top-2 | Frontgrade SBC-2A72, 7.28 GFLOPS effectiveFrontgrade SBC-2A72，有效 7.28 GFLOPS | 33 × 32 = 1,056 polar, 550 km, 87°33 × 32 = 1,056 顆極軌，550 km，87° | ≥ 100 Gbps, ≤ 4 lasers≥ 100 Gbps，≤ 4 條雷射 | propagation + \(MQ_B/R\). Queueing and node delay excluded傳播 + \(MQ_B/R\)。排隊與節點延遲列於範圍之外 | i.i.d. Bernoulli 0.95 per link per snapshot每鏈路每快照獨立 Bernoulli 0.95 | topology-snapshot simulation, accuracy results pending拓樸快照模擬，準確度結果待補充 | 1.02–1.07 s/token vs 3.34–5.30 random每 token 1.02–1.07 s，對照隨機 3.34–5.30 s | 70 B sizing vs 6.7 B experiments. ≥3× vs 2×. 10 vs 10.4 GFLOPS. Expert ordering (item 38)70 B 估算對 6.7 B 實驗。≥3× 對 2×。10 對 10.4 GFLOPS。專家排序（第 38 項） |
| CollabLLM-LEO ppCollabLLM-LEO pp | ViT-B 0.086 B to ViT-G 1.8 B, EuroSAT / RESISC45ViT-B 0.086 B 至 ViT-G 1.8 B，EuroSAT / RESISC45 | 4 × Jetson AGX Orin, ≤ 275 TOPS, 15/30/50 W, 8 GB4 × Jetson AGX Orin，≤ 275 TOPS，15/30/50 W，8 GB | 5 of 12 in one Walker Delta plane單一 Walker Delta 軌道面 12 顆中的 5 顆 | 0.5 Gbit/s. GSL 6 or 0.2–0.8 Gbit/s0.5 Gbit/s。地面鏈路 6 或 0.2–0.8 Gbit/s | transmission with overlap. Orbits simulated含重疊的傳輸。軌道模擬 | unspecified零 | hybrid: real Orin timing + simulated links混合：實體 Orin 計時 + 模擬鏈路 | −42% delay at 800 Mbit/s. >20× compression. −71% traffic800 Mbit/s 下延遲 −42%。壓縮 >20×。流量 −71% | GSL rate. "103% reduction". Model names. 4.1% vs 0.3% loss (item 39)地面鏈路速率。「103% 降幅」。模型名稱。4.1% 對 0.3% 損失（第 39 項） |
| LLMSpace ppLLMSpace pp | generic LLM inference payload泛用 LLM 推論載荷 | DGX-H100 class, 130 kg, 10 kW COTS / 20 kW rad-hardDGX-H100 級，130 kg，10 kW COTS / 20 kW 抗輻射 | single Starlink-V1-class bus, 191 kg單顆 Starlink-V1 級平台，191 kg | terminals up to 200 Gbps, 28 kg each終端最高 200 Gbps，每個 28 kg | unspecified零 | 2- vs 10-year TID/SEE life2 年對 10 年的 TID/SEE 壽命 | analytical bottom-up carbon model解析式由下而上碳排模型 | 16.57 / 30.45 tCO2e embodied. 14.5 kgCO2e/kg launch體現碳排 16.57 / 30.45 tCO2e。發射 14.5 kgCO2e/kg | 130 kg payload on 191 kg bus. 15× / 2× / 10–20× overheads (item 40)191 kg 平台載 130 kg 載荷。15× / 2× / 10–20× 開銷（第 40 項） |
| SkyMemory ppSkyMemory pp | KV-cache edge cache for transformer inferencetransformer 推論的 KV 快取邊緣快取 | Jetson Nano 8 GB host. 5 Intel NUCsJetson Nano 8 GB 主機。5 台 Intel NUC | 19 × 5 emulated19 × 5 模擬 | FSO ISLs, rate unspecifiedFSO ISL，速率待指定 | unspecified待指定 | unspecified待指定 | simulation + testbed模擬 + 測試平台 | numeric results pending in abstract摘要數值結果待補充 | CCNC 2026 workshop version verification pendingCCNC 2026 工作坊版本待驗證 |
| Constella ppConstella pp | AlexNet to EfficientNet-B0, split by layerAlexNet 至 EfficientNet-B0，逐層切割 | Jetson Orin Nano energy coefficientsJetson Orin Nano 能耗係數 | 8/2 to 8,000/2,000 processors/communicators8/2 至 8,000/2,000 處理者/通訊者 | "negligible". Cites up to 100 Gbps「可忽略」。引用最高 100 Gbps | fixed phases 5,100/300/900 s. Downlink 625 kbit/s固定階段 5,100/300/900 s。下行 625 kbit/s | model unspecified模型待指定 | simulation on BUPT-1 datasetBUPT-1 資料集上的模擬 | 1,541 vs 4,129 s. ≤2.7× latency, ≤74× energy. ≥81.9% success1,541 對 4,129 s。延遲 ≤2.7×、能耗 ≤74×。成功率 ≥81.9% | submission date vs arXiv id (item 52)投稿日期對 arXiv 編號（第 52 項） |
| Cached Model-as-a-Resource ppCached Model-as-a-Resource pp | LLM agents on ground edge/cloud地面邊緣/雲端上的 LLM 代理 | ground servers. Satellites relay地面伺服器。衛星中繼 | space-air-ground, unspecified空天地，待指明 | unspecified零 | "age of thought" freshness, with path delay separate「思考年齡」新鮮度，路徑延遲則另計 | unspecified零 | DQN auction simulationDQN 拍賣模擬 | +23% allocation efficiency配置效率 +23% | ToN DOI verification pendingToN DOI 待驗證 |
The flown record remains single-satellite. Hardware declarations in the distributed papers therefore require separate checks against sustained power, thermal behavior and the flight configuration.
飛行紀錄仍以單衛星為範圍，分散式論文的硬體假設因而需另外核對持續功率、熱行為與實際飛行配置。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | Set beside these simulations, the flown record is thin and single-satellite throughout. |
| Result and scope | 2026 | The table below gathers it with evidence tags. Its last column is the one that matters: multi-satellite distributed inference and training over ISLs still await flight demonstration as of October 2026. |
| Power and thermal conditions | 1, 2026, 1 kW, 15–20, 2027, 2025 | The Suncatcher MVP launched on 1 October 2026 with four TPUs, roughly 1 kW and 15–20 minute duty cycles. It uses ground connectivity only [100][104] company. Press. The first connected-ISL test awaits Planet’s planned 2027 pair (Planet, Nov 2025) company. |
| Power and thermal conditions | 7.28 GFLOPS, 50 W, 24 | SpaceMoE’s 7.28 GFLOPS board represents flight heritage. CollabLLM-LEO’s Orin class has flown on CubeSats, with sustained 50 W operation still awaiting demonstration. The H100 modeled by LLMSpace has flown once, while duty-cycle and thermal figures remain open (fact-base item 24, Part D). |
| 結果與適用範圍 | 定性範圍 | 放在這些模擬旁邊，飛行紀錄既單薄又全是單衛星。 |
| 結果與適用範圍 | 2026, 10 | 下表以證據標籤彙整，最後一欄才是重點：截至 2026 年 10 月，多衛星、經 ISL 的分散式推論與訓練仍待在軌示範。 |
| 電力與熱條件 | 2026, 10, 1, 1 kW, 15–20, 2027, 2025, 11 | 2026 年 10 月 1 日入軌的 Suncatcher MVP 載有四顆 TPU、約 1 kW，以 15–20 分鐘的工作週期運轉，太空船僅使用地面鏈路[100][104]company. Press，因此第一次 ISL 相連的測試要等到計畫於 2027 年發射、由 Planet 建造的兩顆衛星（Planet，2025 年 11 月）company。 |
| 電力與熱條件 | 7.28 GFLOPS, 50 W, 24 | 論文與實際發射清單之間的硬體差距同樣巨大：SpaceMoE 的 7.28 GFLOPS 板代表飛行履歷，CollabLLM-LEO 的 Orin 級硬體曾在 CubeSat 上飛過，持續 50 W 運轉仍待示範，LLMSpace 建模的 H100 只飛過一次，工作週期與熱學數字待公開（事實庫 Part D 第 24 項）。 |
| Flown demonstration已飛行的示範 | Hardware and model硬體與模型 | Published numbers公開數字 | Satellites involved涉及衛星數 | Evidence證據 |
|---|---|---|---|---|
| Φ-Sat-1 CloudScout (2020)Φ-Sat-1 CloudScout（2020） | Intel Movidius Myriad 2. 2.1 MB CNNIntel Movidius Myriad 2。2.1 MB CNN | 325 ms and 1.8 W per inference, 92% accuracy [87, Tabs. 3, 5]. In-orbit validation pending at publication每次推論 325 ms、1.8 W，準確度 92%[87, Tabs. 3, 5]。出版時在軌驗證尚待進行 | one一顆 | peer-reviewed |
| RaVAEn on D-Orbit ION SCV004 (2023)D-Orbit ION SCV004 上的 RaVAEn（2023） | Myriad X VPU. VAE encoder + 129-parameter classifierMyriad X VPU。VAE 編碼器 + 129 參數分類器 | 0.110 s per tile encode. 0.091 s per training epoch [95]每區塊編碼 0.110 s。每訓練 epoch 0.091 s[95] | one一顆 | preprint |
| BUPT-1 COTS compute (2023–24)BUPT-1 COTS 運算（2023–24） | 2 × Raspberry Pi 4B, 2 × Atlas 200 DK. YOLO workloads2 × Raspberry Pi 4B、2 × Atlas 200 DK。YOLO 工作負載 | Pi >80 °C within ~50 min at >75% load. Zero SEE errors in six months [85]Pi 在 >75% 負載下約 50 分鐘內超過 80 °C。六個月 SEE 錯誤計數為零[85] | one一顆 | peer-reviewed |
| Starcloud-1 (Nov–Dec 2025)Starcloud-1（2025 年 11–12 月） | one NVIDIA H100. Gemma-family model, nanoGPT training loop一顆 NVIDIA H100。Gemma 系列模型、nanoGPT 訓練迴圈 | throughput, power and thermal figures await publication (NVIDIA blog. DCD, 12 Dec 2025)吞吐量、功率與熱學數字待公開（NVIDIA 部落格。DCD，2025 年 12 月 12 日） | one一顆 | company. Press |
| Three-Body Computing Constellation (May 2025)三體計算星座（2025 年 5 月） | 12 satellites. Claimed 744 TOPS and an 8B model each12 顆衛星。宣稱每顆 744 TOPS 與一個 8B 模型 | claims only, mutually inconsistent (Live Science. Fact base Part D, item 28)僅為宣稱，彼此有差異（Live Science。事實庫 Part D 第 28 項） | one per model每個模型一顆 | company via press |
| Suncatcher MVP (1 Oct 2026)Suncatcher MVP（2026 年 10 月 1 日） | four TPUs, ≈1 kW. Workload undisclosed四顆 TPU，≈1 kW。工作負載待公開 | 15–20 min duty cycles reported. Heat-pipe orbital validation pending. Ground connectivity only (Google, Sept–Oct 2026. Payload)報導稱 15–20 分鐘工作週期。熱管在軌驗證待完成。僅地面鏈路（Google，2026 年 9–10 月。Payload） | one一顆 | company. Press |
| Any multi-satellite inference over ISLs任何經 ISL 的多衛星推論 | —— | flight demonstration pending as of October 2026. First ISL test planned for two Planet satellites in 2027截至 2026 年 10 月仍待示範。首次 ISL 測試規劃於 2027 年的兩顆 Planet 衛星 | zero零 | company plan |
12.5Sensitivity to the ISL assumption對 ISL 假設的敏感度
Recall
回顧
Section 12.1 gives 8 KB per decode boundary, 33.6 MB per 4,096-token prefill and 2.15 GB KV per 4,096-token context. Section 12.2 shows SpaceMoE’s compute-dominated token time. Section 2.1 converts hop distance to propagation delay at 3.336 µs/km.
The papers assume widely separated ISL rates. Holding the remaining inputs fixed separates the influence of transmission rate from compute and propagation.
各論文的 ISL 速率假設相差很大，固定其他輸入後，才能分開觀察傳輸速率、運算與傳播的影響。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Capacity and rate | 0.5 Gbit/s, 100, 200, 9.6 Tb/s | Stated ISL rates span more than four orders. CollabLLM-LEO uses 0.5 Gbit/s, SpaceMoE “≥ 100 Gbps”, LLMSpace up to 200 Gbps terminals and Suncatcher 9.6 Tb/s per aperture [36, pp. 3, 8]. Each paper draws conclusions at its selected rate. |
| Capacity and rate | Qualitative scope | Test rate sensitivity with the three-term time-per-token model. Vary ISL rate while holding computation, propagation and all remaining inputs fixed. |
| Latency comparison | Qualitative scope | For a single request stream the stages of a pipeline run one after another. Thus, the compute term is the whole model's work divided by the per-satellite rate and stays fixed as stages are added. In comparison, each added stage adds one boundary whose tensor must be transmitted and one hop whose light-time must be paid. |
| Result and scope | Qualitative scope | Let \(P\) be parameters exercised per token and \(T\) the tokens per step. Decode uses one token, while prefill uses prompt length. Let \(F\) denote per-satellite compute rate and \(S\) the stages in a ring. The sampled token returns to stage one over an extra hop. Hidden size is \(h\), compression ratio \(c\), ISL rate \(R\) and per-hop propagation delay \(\tau\). The model is as follows. |
| 容量與速率 | 0.5 Gbit/s, 100, 200, 9.6 Tb/s | 四篇有寫出 ISL 速率的論文，把它放在跨越四個數量級以上的四個點：從 CollabLLM-LEO 的 0.5 Gbit/s，經過 SpaceMoE 的「≥ 100 Gbps」與 LLMSpace 最高 200 Gbps 的終端，到 Suncatcher 設計中每口徑 9.6 Tb/s[36, pp. 3, 8]，而每篇論文都把結論畫得好像自己的那個點是唯一的點。 |
| 結果與適用範圍 | 定性範圍 | 檢驗這些結論的方法，是把每 token 時間寫成這些系統都包含的三項之和，再固定其他條件、只移動 ISL 速率。 |
| 結果與適用範圍 | 定性範圍 | 對單一請求串流而言，管線的各個階段是一個接一個執行的，因此運算項等於整個模型的工作量除以每顆衛星的運算速率，在階段增加時保持固定。 |
| 結果與適用範圍 | 定性範圍 | 而每多一個階段，就多一個要傳輸張量的邊界與一個要付出光行時間的跳。 |
| 結果與適用範圍 | 定性範圍 | 以 \(P\) 表示每 token 動用的參數，\(T\) 表示一步中的 token 數（解碼為一，預填為提示長度），\(F\) 表示每顆衛星的運算速率，\(S\) 表示排成環狀、使取樣出的 token 再經一跳回到第一階段的階段數，\(h\) 為隱藏維度，\(c\) 為壓縮比，\(R\) 為 ISL 速率，\(\tau\) 為一跳的傳播延遲，模型為 |
\[ t_{\text{step}} \;=\; \underbrace{\frac{2PT}{F}}_{\text{compute}} \;+\; \underbrace{(S-1)\,\frac{8\cdot 2hT}{c\,R}}_{\text{transmission}} \;+\; \underbrace{S\,\tau}_{\text{propagation}} \qquad (S \ge 2) \]
The panel’s presets use stated model shapes, processors and hop geometries. Each preset retains its source and interpretation as an assumption or reported hardware capability.
面板採用已列出的模型形狀、處理器與跳距幾何，每組預設都保留來源，以及模型假設或硬體能力的證據定位。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 12.1 | Panel 12.1 evaluates this expression with presets taken from the chapter's sources. |
| Model and compute | 1B, 2,048, 16, 7B, 4,096, 32, 12.1, 70B, 8,192, 80, 3.5B, 3.5 B | Decoder presets use two-byte activations. The 1B shape has hidden size 2,048 and 16 layers. The 7B shape uses 4,096 and 32 layers, as in Section 12.1. The 70B shape uses 8,192 and 80 layers. The 3.5B preset represents SpaceMoE’s 3.5 B active parameters with hidden size 4,096. |
| Compute presets | 7.28 GFLOPS, 275 TOPS, 918 TFLOPS | Compute presets are SpaceMoE’s 7.28 GFLOPS board, CollabLLM-LEO’s 275 TOPS Orin peak and TPU v6e’s 918 TFLOPS bf16 peak [99] company. TPU v6e is the Suncatcher prototype’s chip. Here, it represents the H100-class payloads modeled in LLMSpace. |
| Hop presets | 100 m, 659 km, 1,969 km, 24, 66, 72, 22, 12, 1,357 km | Hop presets include Suncatcher’s 100 m nearest-neighbor spacing and SpaceMoE’s 1,357 km. The Starlink filing presets are 659 km for 24 × 66 and 1,969 km for 72 × 22. These are intra-plane spacings (fact-base item 12, Part C). |
| Capacity and rate | 4.5 ms, 100 Gb/s | The initial SpaceMoE replica uses eight stages on the radiation-hardened board, with 4.5 ms hops at 100 Gb/s. Its result lies within a few percent of Table II. |
| 結果與適用範圍 | 12.1 | 面板 12.1 以取自本章來源的預設值計算這個式子。 |
| 結果與適用範圍 | 1B, 2,048, 16 | 模型預設是激活值為兩位元組的稠密 decoder 形狀：1B 的隱藏維度 2,048、16 層。 |
| 結果與適用範圍 | 7B, 4,096, 32, 12.1 | 7B 為 4,096、32 層，與第 12.1 節相同。 |
| 結果與適用範圍 | 70B, 8,192, 80 | 70B 為 8,192、80 層。 |
| 結果與適用範圍 | 3.5B, 3.5 B, 4,096 | 另有一個 3.5B 預設，代表 SpaceMoE 的 3.5 B 活躍參數，隱藏狀態同樣是 4,096 寬。 |
| 模型與運算 | 7.28 GFLOPS, 275 TOPS, 918 TFLOPS | 運算預設是 SpaceMoE 的 7.28 GFLOPS 抗輻射板、CollabLLM-LEO 的 275 TOPS Orin 峰值，以及 918 TFLOPS，也就是 Google 為單顆 TPU v6e Trillium 晶片列出的 bf16 速率[99]company。 |
| 結果與適用範圍 | 定性範圍 | 它代表 LLMSpace 所建模的 H100 級載荷，選它是因為 Suncatcher 原型機載的正是這顆晶片。 |
| 軌道配置 | 100 m, 24, 66, 72, 22, 659 km, 1,969 km, 12, 1,357 km | 跳距預設是 Suncatcher 編隊中 100 m 的最近鄰間距、Starlink 24 × 66 與 72 × 22 兩份申請中的同軌間距 659 km 與 1,969 km（事實庫 Part C 第 12 項），以及 SpaceMoE 的 1,357 km。 |
| 容量與速率 | 4.5 ms, 100 Gb/s | 面板一開始是 SpaceMoE 的複製版：八個階段、抗輻射板、每跳 4.5 ms、100 Gb/s，結果與論文 Table II 相差在百分之幾以內。 |
Stop and think停下來想一想
Start from the SpaceMoE replica and switch the compute to 275 TOPS. The time per token falls by a factor of about thirty, and the bottleneck tile changes, but to which term? Now sweep the ISL rate across all four presets and watch how little the total moves. Then switch to prefill and repeat the sweep. Identify the control that makes ISL rate the decode bottleneck, then find its threshold.
從 SpaceMoE 複製版開始，把運算切到 275 TOPS。每 token 時間大約降為三十分之一，瓶頸方塊也變了，但變成哪一項？接著把 ISL 速率掃過四個預設，觀察總時間變化極小。再切到預填，重做一次掃描。要讓 ISL 速率在解碼時成為瓶頸，你得改動哪一個控制項，又得改動多少？
At the selected decode rates, faster compute exposes propagation as the bottleneck. Prefill moves much larger tensors, so link rate has a stronger effect.
在所選解碼速率下，運算加快後會由傳播成為瓶頸。預填傳送的張量更大，鏈路速率因而更有影響。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Latency comparison | Qualitative scope | What the panel shows for decode is that the bottleneck moves from compute to propagation and stays below the transmission term’s bottleneck threshold. |
| Latency comparison | 0.96 s, 36 ms, 4.5 ms, 25 µs | The SpaceMoE replica spends 0.96 s computing and 36 ms traversing eight 4.5 ms hops. Replacing the board with Orin reduces compute to 25 µs. The same 36 ms propagation then dominates token time. |
| Reliability comparison | 0.5 Gb/s, 0.9 ms, 8, 131 µs, 100 Gb/s, 0.66 µs | Across four rate presets, total-time change is at most 0.9 ms, the transmission across seven boundaries at 0.5 Gb/s. An 8 KB state takes 131 µs at 0.5 Gb/s and 0.66 µs at 100 Gb/s. |
| Latency comparison | 4.5 ms, 14.6 Mb/s, 2 | On a 4.5 ms hop, decode becomes link-bound below 14.6 Mb/s, slower than Chapter 2’s lowest RF crosslink rate. Within this range, token generation depends mainly on hop count and length, with little ISL-rate sensitivity. |
| Result and scope | 4,096 | Prefill reverses the picture, because the tensor at each boundary is 4,096 times larger. |
| Result and scope | 33.6 MB, 0.54 s, 0.5 Gb/s, 0.21 s, 7B, 100 Gb/s, 2.7 ms, 9.6 Tb/s, 28 µs | At 0.5 Gb/s, a 33.6 MB prefill tensor takes 0.54 s, exceeding Orin’s 0.21 s 7B-prefill compute time. Transfer takes 2.7 ms at 100 Gb/s, near one hop’s propagation. At 9.6 Tb/s it takes 28 µs. |
| Capacity and rate | Qualitative scope | The ISL rate therefore matters for whatever moves in bulk, which in inference means prompts and KV caches. It is irrelevant for what moves token by token. |
| 結果與適用範圍 | 定性範圍 | 面板在解碼時顯示的是：瓶頸從運算移到傳播，在所選速率範圍內，傳輸時間則較小。 |
| 結果與適用範圍 | 0.96 s, 4.5 ms, 36 ms | 在抗輻射板上，SpaceMoE 複製版花 0.96 s 運算、在八個 4.5 ms 的跳上花 36 ms。 |
| 結果與適用範圍 | 25 µs, 36 ms | 把板子換成 Orin，運算降到 25 µs，剩下的 36 ms 傳播幾乎就是全部的每 token 時間。 |
| 容量與速率 | 0.5 Gb/s, 0.9 ms, 8, 131 µs, 100 Gb/s, 0.66 µs | 接著把 ISL 速率掃過四個預設，總時間的變化至多是 0.5 Gb/s 下七個邊界的傳輸時間，也就是 0.9 ms，因為一個 8 KB 的隱藏狀態穿過 0.5 Gb/s 鏈路要 131 µs、穿過 100 Gb/s 鏈路只要 0.66 µs。 |
| 結果與適用範圍 | 4.5 ms, 14.6 Mb/s, 2 | 在 4.5 ms 的跳上，解碼張量要到 14.6 Mb/s 以下才會變成受鏈路限制，這比第 2 章中最慢的 RF 星間鏈路還慢，所以對逐 token 的生成而言，ISL 速率的影響較小，主要參數是跳數與跳距。 |
| 結果與適用範圍 | 4,096 | 預填則把局面反轉過來，因為每個邊界的張量大了 4,096 倍。 |
| 容量與速率 | 33.6 MB, 0.5 Gb/s, 0.54 s, 7B, 0.21 s | 33.6 MB 的預填張量在 0.5 Gb/s 鏈路上要 0.54 s，比 Orin 計算 7B 預填所需的 0.21 s 還久。 |
| 容量與速率 | 100 Gb/s, 2.7 ms | 在 100 Gb/s 下要 2.7 ms，接近一跳。 |
| 容量與速率 | 9.6 Tb/s, 28 µs | 在 9.6 Tb/s 下只要 28 µs。 |
| 結果與適用範圍 | 定性範圍 | 因此 ISL 速率對大量搬運的東西才重要，在推論中那就是提示與 KV 快取，而對逐 token 搬運的東西則影響較小。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Propagation of one hop at 100 m, 659 km, 1,357 km and 1,969 km100 m、659 km、1,357 km 與 1,969 km 一跳的傳播延遲 | \(d \times 3.336\,\mu\text{s/km}\) | 0.33 µs · 2.2 ms · 4.5 ms · 6.6 mscomputed. Spacings from Suncatcher (Joule p. 4), fact base Part C item 12 and Section 12.2計算值。間距取自 Suncatcher（Joule p. 4）、事實庫 Part C 第 12 項與第 12.2 節 |
| 8 KB decode tensor at 0.5 Gb/s, 100 Gb/s, 200 Gb/s and 9.6 Tb/s8 KB 解碼張量在 0.5 Gb/s、100 Gb/s、200 Gb/s 與 9.6 Tb/s 下 | \(65{,}536\,\text{bit} / R\) | 131 µs · 0.66 µs · 0.33 µs · 6.8 nscomputed. Rates from CollabLLM-LEO, SpaceMoE, LLMSpace and Suncatcher計算值。速率取自 CollabLLM-LEO、SpaceMoE、LLMSpace 與 Suncatcher |
| ISL rate at which the decode tensor's transmission equals one 4.5 ms hop解碼張量傳輸時間等於一跳 4.5 ms 時的 ISL 速率 | \(65{,}536\,\text{bit} / 4.5\,\text{ms}\) | 14.6 Mb/scomputed計算值 |
| ISL rate at which the 4,096-token prefill tensor's transmission equals one 4.5 ms hop4,096 token 預填張量傳輸時間等於一跳 4.5 ms 時的 ISL 速率 | \(4{,}096 \times 65{,}536\,\text{bit} / 4.5\,\text{ms}\) | 59.7 Gb/scomputed. Just below SpaceMoE's "≥ 100 Gbps"計算值。略低於 SpaceMoE 的「≥ 100 Gbps」 |
| Prefill tensor (33.6 MB) at 0.5 Gb/s, 100 Gb/s and 9.6 Tb/s預填張量（33.6 MB）在 0.5 Gb/s、100 Gb/s 與 9.6 Tb/s 下 | \(2.68\times10^{8}\,\text{bit} / R\) | 0.54 s · 2.7 ms · 28 µscomputed計算值 |
| ISL rate above which an original tensor beats CollabLLM-LEO's compressed one at 0.5 Gb/s原始張量勝過 CollabLLM-LEO 在 0.5 Gb/s 下壓縮張量的 ISL 速率 | \(0.5\,\text{Gb/s} \times 25.82\) | 12.9 Gb/scomputed. Ratio from Fig. 8計算值。比例取自 Fig. 8 |
| 7B decode compute at 7.28 GFLOPS, 275 TOPS and 918 TFLOPS7B 解碼運算在 7.28 GFLOPS、275 TOPS 與 918 TFLOPS 下 | \(14\times10^{9} / F\) | 1.92 s · 51 µs · 15 µscomputed. 918 TFLOPS bf16 from the TPU v6e documentation計算值。918 TFLOPS bf16 取自 TPU v6e 文件 |
| Time to read 7B bf16 weights once from Trillium HBM從 Trillium HBM 讀一次 7B bf16 權重的時間 | \(14\,\text{GB} / 1{,}638\,\text{GB/s}\) | 8.5 mscomputed. HBM bandwidth from the TPU v6e documentation (company)計算值。HBM 頻寬取自 TPU v6e 文件（company） |
Each paper’s conclusion responds to a separate input. Compression depends on transfer volume and rate, placement on compute and hop distance, and carbon accounting on terminal count.
各論文的結論受到各自輸入影響，壓縮取決於傳輸量與速率，配置取決於運算及跳距，碳排計算則取決於終端數。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | 12.4 | Figure 12.4 places each paper on this ladder and marks where its conclusion would change. |
| Capacity and rate | 12.9 Gb/s, 25.82, 0.5 Gb/s | CollabLLM-LEO’s compression advantage depends on image-sized tensors and a slow link. Its 25.82× compression at 0.5 Gb/s loses its transmission advantage above 12.9 Gb/s. Above that rate, the original tensor crosses faster than the compressed baseline. |
| Result and scope | 4, 100%, 42%, 800 Mbit/s | Figure 4 shows gains falling as ground rate rises. The claim changes from “over 100%” at low rates to 42% at 800 Mbit/s. |
| Latency comparison | 14.6 Mb/s, 59.7 Gb/s, 100 | SpaceMoE’s roughly one-second decode result stays insensitive to ISL rate down to 14.6 Mb/s per hop. Prefill becomes link-bound below 59.7 Gb/s. Its “≥ 100 Gbps” assumption therefore sits just above the prefill threshold. |
| Result and scope | Qualitative scope | The about-a-second figure flips with compute rather than with the link. |
| Latency comparison | 625,000 bit/s, 100, 1, 3.1 | Constella calls ISL latency and energy negligible against its 625,000 bit/s downlink, citing modern links up to 100 Gbps (§1, §3.1). The ladder supports this relation for its CNN-sized tensors. |
| Terminal carbon model | 28 kg, 150 | LLMSpace's result is a carbon total in which the ISL enters as 28 kg and 150 kgCO2e per terminal. Thus, it flips with the number of terminals and with their rate excluded from the calculation. |
| Latency comparison | 100 m, 0.33 µs, 9.6 Tb/s, 6.8 ns | A 100 m Suncatcher hop propagates in 0.33 µs. Its 9.6 Tb/s link sends a decode tensor in 6.8 ns. |
| Result and scope | Qualitative scope | so a serving pipeline there would be bound entirely by the accelerator |
| Result and scope | 7B, 1,638 GB/s, 8.5 ms, 15 µs | On Trillium, weight reading dominates arithmetic. Reading 7B bf16 weights at 1,638 GB/s HBM takes 8.5 ms, compared with 15 µs of FLOPs. |
| Latency comparison | 1 ms, 10 ms, 19 | All five papers exclude queueing and per-hop processing. The literature supplies 1 ms and 10 ms node-delay assumptions (fact-base item 19, Part C). Either increases the propagation-related budget, widening the gap between dominant delay and ISL-rate effects. |
| 結果與適用範圍 | 12.4 | 圖 12.4 把每篇論文放到這條梯子上，並標出它的結論會在哪裡改變。 |
| 結果與適用範圍 | 定性範圍 | CollabLLM-LEO 的結論是壓縮讓軌道管線勝過下傳，之所以成立，是因為它的張量是影像大小、鏈路又慢。 |
| 容量與速率 | 12.9 Gb/s, 0.5 Gb/s, 25.82 | 當 ISL 速率達到 12.9 Gb/s，壓縮就收益降低，因為超過這個速率，原始的張量穿越鏈路比論文在 0.5 Gb/s 下壓縮 25.82× 的張量還快。 |
| 結果與適用範圍 | 4, 100%, 800 Mbit/s, 42% | 論文自己的 Fig. 4 也顯示，隨著地面鏈路變快，增益從低速率下的「超過 100%」縮小到 800 Mbit/s 下的 42%。 |
| 容量與速率 | 14.6 Mb/s | SpaceMoE 的結論是配置把每 token 時間壓到大約一秒，這在解碼時對 ISL 速率影響較小，每跳一直到 14.6 Mb/s 都一樣。 |
| 容量與速率 | 59.7 Gb/s, 100 | 而它的預填會在 59.7 Gb/s 以下變成受鏈路限制，使「≥ 100 Gbps」成為一個剛好位在它開始產生影響之點上方的假設。 |
| 結果與適用範圍 | 定性範圍 | 大約一秒的數字會隨運算條件翻轉，鏈路速率的影響較小。 |
| 容量與速率 | 100, 625,000 bit/s, 1, 3.1 | Constella 在引用現代 ISL 最高 100 Gbps 的同時，宣告 ISL 的延遲與能耗相對於其 625,000 bit/s 的下行鏈路可以忽略（§1、§3.1），對它 CNN 大小的張量而言，這條梯子證實了這一點。 |
| 結果與適用範圍 | 28 kg, 150 | LLMSpace 的結果是碳排總量，ISL 只以每個終端 28 kg 與 150 kgCO2e 的形式進入，所以它隨終端數量翻轉，速率則列於計算範圍之外。 |
| 結果與適用範圍 | 100 m, 0.33 µs, 9.6 Tb/s, 6.8 ns, 1,638 GB/s, 7B, 8.5 ms, 15 µs | 最後，在 Suncatcher 型叢集內，一跳 100 m 只要 0.33 µs，9.6 Tb/s 的鏈路搬一個解碼張量只要 6.8 ns，因此在那裡的服務管線會完全受加速器限制，而在 Trillium 上限制來自讀取權重，算術的耗時則較少：以 1,638 GB/s 的 HBM 讀一遍 7B bf16 權重要 8.5 ms，FLOPs 只要 15 µs。 |
| 結果與適用範圍 | 1 ms, 10 ms, 19 | 五篇論文的模型範圍之外都有排隊與每跳節點延遲，而文獻對此只提供 1 ms 與 10 ms 兩種建模選擇（事實庫 Part C 第 19 項）。 |
| 結果與適用範圍 | 定性範圍 | 兩種取值都，都會加到傳播項上，使真正影響結果的因素與 ISL 速率所能控制的因素之間的差距更大。 |
Original figure note原始圖說出處
Drawn for this primer from SpaceMoE §II-B and Table II, CollabLLM-LEO Table II and Figs. 4 and 8, LLMSpace §4.1 and Table 6, Agüera y Arcas et al., Joule 2026, pp. 3, 8, and the calc table above.12.6Ground and orbit as two serving sites地面與軌道作為兩個服務站點
Recall
回顧
Section 3.5 shows Starlink reassigning users at a 15-second reconfiguration interval. Section 9.3 shows that each satellite remains visible to a ground point for minutes per orbit.
Ground and orbital serving sites have separate access paths, energy supplies and state locations. Their comparison must include user-link variation as well as the path beyond the PoP.
地面與軌道服務站點各有接入路徑、能源供應及狀態位置，比較時需同時納入使用者鏈路變化與 PoP 之後的路徑。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Latency comparison | Qualitative scope | A model served for people on the ground can run in a terrestrial data center or on satellites. A system that uses both treats them as two serving sites with different latency paths, energy supplies and places to keep state. |
| Latency comparison | 40 ms, 40–50 ms, 15, 12, 27, 42, 57 | Orbital serving begins with the user access link. Measured idle RTT medians are about 40 ms through bent pipe and 40–50 ms across most countries [49, Figs. 5–7]. Latency changes at 15-second boundaries: seconds 12, 27, 42 and 57 each minute [13]. Terrestrial serving follows the same PoP with a fiber path. |
| Result and scope | Qualitative scope | Reading the evaluation sections of the inference papers for what each compares against and what it holds fixed yields three observations about this two-site picture. |
| Result and scope | Qualitative scope | These observations define the evaluations’ scope and the measurements needed for two-site serving. |
| 結果與適用範圍 | 定性範圍 | 為地面上的人服務的模型，可以在地面資料中心執行，也可以在衛星上執行。 |
| 結果與適用範圍 | 定性範圍 | 同時使用兩者的系統，就把它們當成兩個服務站點，各有相異的延遲路徑、能源供應與保存狀態的地方。 |
| 結果與適用範圍 | 40 ms, 40–50 ms, 15, 12, 27, 42, 57 | 通往軌道站點的路徑始於使用者鏈路，其量測到的閒置往返中位數經彎管約 40 ms，在多數國家為 40–50 ms[49, Figs. 5–7]，而且延遲在每個 15 秒邊界、即每分鐘的第 12、27、42、57 秒發生變化[13]。 |
| 結果與適用範圍 | 定性範圍 | 通往地面站點的路徑則從同一個 PoP 經光纖出發。 |
| 結果與適用範圍 | 定性範圍 | 把推論論文的評估章節讀成「各自拿什麼比較、固定了什麼」，就能對這個雙站點的圖像得出三個觀察。 |
| 結果與適用範圍 | 定性範圍 | 這些觀察界定了各項評估的範圍，以及雙站點服務仍需補充的量測。 |
A ground-versus-orbit comparison for the same ground-originated request remains open. Existing baselines compare placements, downlink strategies, energy sources or generic edge services.
針對同一個源自地面的請求，地面與軌道服務的比較仍待完成。現有基準比較的是配置、下傳策略、能源來源或泛用邊緣服務。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Orbital configuration | Qualitative scope | The first open evaluation is a ground-versus-orbit comparison for the same ground-originated request. |
| Latency comparison | Qualitative scope | SpaceMoE's baselines are three other placements of the same experts in the same constellation. Its measured latency begins and ends inside the constellation with the user link excluded (§VII-B, Table II). |
| Result and scope | 4, 4.1, 1 | CollabLLM-LEO’s ground server processes satellite-originated images after downlink (§VI-A, Fig. 4). Constella’s naive and bent-pipe Earth-observation baselines use the same structure (§4.1, Table 1). |
| Carbon and service models | 20, 380, 6, 2 | LLMSpace compares orbital operational carbon with grids at 20 and 380 gCO2e/kWh (§6, Fig. 2a). This evaluates energy sources, with serving latency outside its scope. Cached Model-as-a-Resource retains models on ground servers and uses satellites as relays. |
| Latency comparison | 46, 16 ms | Earlier ground–orbit comparisons concern generic edge servers. Bhattacherjee et al. estimate one server per satellite at roughly three times terrestrial ownership cost. A West African user’s latency falls from 46 to 16 ms [2]. |
| Latency comparison | 22–183 ms, 13–90 ms | and Celestial's emulation moved an edge service from ground to the LEO edge and cut latency from 22–183 ms to 13–90 ms [79]. |
| Result and scope | Qualitative scope | Both studies leave models with token-growing state outside their scope. |
| 結果與適用範圍 | 定性範圍 | 第一項待完成的評估，是針對同一個源自地面的請求，比較地面與軌道服務。 |
| 結果與適用範圍 | 定性範圍 | SpaceMoE 的基準是同一批專家在同一個星座中的另外三種配置，它量測的延遲在星座內開始、在星座內結束，使用者鏈路列於範圍之外（§VII-B，Table II）。 |
| 結果與適用範圍 | 4, 4.1, 1 | CollabLLM-LEO 確實與地面伺服器比較，但對象是誕生在感測衛星上、必須先下傳才能在任何地方處理的影像（§VI-A，Fig. 4），而 Constella 的天真方案與彎管基準對地球觀測資料也是同樣的結構（§4.1，Table 1）。 |
| 結果與適用範圍 | 20, 380, 6, 2 | LLMSpace 把軌道載荷的營運碳排與 20 及 380 gCO2e/kWh 的地面電網比較（§6，Fig. 2a），比的是能源來源，服務站點延遲則列於範圍之外。 |
| 模型與運算 | 定性範圍 | Cached Model-as-a-Resource 框架則把模型留在地面伺服器、把衛星當中繼。 |
| 結果與適用範圍 | 46 ms, 16 ms | 確實存在的地面對軌道比較都早於 LLM 服務，對象是泛用的邊緣伺服器：Bhattacherjee 等估計每顆衛星一台伺服器的總擁有成本約為地面的三倍，並把西非使用者的延遲從 46 ms 降到 16 ms[2]。 |
| 結果與適用範圍 | 22–183 ms, 13–90 ms | Celestial 的模擬把一個邊緣服務從地面搬到 LEO 邊緣，延遲從 22–183 ms 降到 13–90 ms[79]。 |
| 結果與適用範圍 | 定性範圍 | 狀態隨 token 增長的模型列於這兩項研究的範圍之外。 |
Time-varying energy and orbital motion remain separate in these evaluations. A serving scheduler must combine sunlight, battery state, grid carbon and latency to shift requests between sites.
這些評估仍分開處理時變能源與軌道運動。跨站點服務排程需結合日照、電池狀態、電網碳排與延遲，才能安排請求移轉。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Power and thermal conditions | Qualitative scope | The second open evaluation is joint modeling of time-varying energy and the orbit that causes it. |
| Eclipse and battery model | 90–100, 30–35, 80%, 4.1 | LLMSpace assumes a 90–100 min orbit, 30–35 min eclipse and batteries at 80% discharge depth. These inputs size hardware for embodied-carbon accounting. Latency and sunlight-based work schedules remain outside scope (§4.1). |
| Result and scope | 200, 7.28 GFLOPS | SpaceMoE varies topology over 200 slots while keeping compute at 7.28 GFLOPS, with power outside scope (§VII-A1). CollabLLM-LEO fixes one of three Orin power modes per run while simulating orbits (§VI-A). |
| Orbital configuration | 6,300 s, 5,100 s, 300 s, 900 s, 3.1 | Constella fixes a 6,300 s orbit into 5,100 s compute, 300 s communication and 900 s idle (§3.1). These allocations remain constant across sunlight and battery conditions. |
| Result and scope | Qualitative scope | On the ground side, LLMSpace's two grid intensities are constants. |
| Dataset and workload | Qualitative scope | Moving requests between sites as satellites enter eclipse or grid carbon changes remains outside all these evaluations. |
| 結果與適用範圍 | 定性範圍 | 第二項待完成的評估，是將隨時間變化的能源與造成變化的軌道一起建模。 |
| 結果與適用範圍 | 90–100, 30–35, 80% | LLMSpace 具備這些成分：90–100 分鐘的軌道、30–35 分鐘的地影、以 80% 放電深度設計的電池。 |
| 結果與適用範圍 | 4.1 | 但它用這些來決定硬體規模並加總其體現碳排，延遲模型與按日照安排的工作時程則待建立（§4.1）。 |
| 結果與適用範圍 | 200, 7.28 GFLOPS | SpaceMoE 在 200 個時槽中改變拓樸，而運算始終是固定的 7.28 GFLOPS、功率模型待建立（§VII-A1）。 |
| 結果與適用範圍 | 定性範圍 | CollabLLM-LEO 每次執行固定三種 Orin 功率模式中的一種，軌道則是模擬的（§VI-A）。 |
| 結果與適用範圍 | 6,300 s, 5,100 s, 300 s, 900 s, 3.1 | Constella 把 6,300 s 的軌道分成固定比例：5,100 s 運算、300 s 通訊、900 s 閒置，在日照與電池狀態改變時仍維持固定（§3.1）。 |
| 結果與適用範圍 | 定性範圍 | 在地面這一側，LLMSpace 的兩個電網碳強度都是常數。 |
| 結果與適用範圍 | 定性範圍 | 因此，一個在衛星進入地影時、或在電網碳強度隨一天變化時於站點之間轉移請求的服務系統，列於這組論文的評估範圍之外。 |
Evaluated configurations keep KV state local. Serving a moving satellite’s conversation therefore requires either return hops to that satellite or a bulk KV transfer, whose cost depends on link rate.
已評估配置將 KV 留在本地，對話隨衛星移動時便需經 ISL 返回原衛星，或整批搬移 KV，後者成本取決於鏈路速率。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Result and scope | Qualitative scope | The third open evaluation is cross-ISL KV transfer. |
| Result and scope | Qualitative scope | All evaluated configurations keep KV local. |
| Payload and cache models | 6, 4.1 | SpaceMoE transfers hidden states of size \(MQ_B\) (Eq. 6). CollabLLM-LEO’s vision transformers make one image forward pass with local, KV-free processing (§III). LLMSpace models one satellite with local traffic only (§4.1). SkyMemory alone proposes KV-sized traffic, while ISL rate, cache size and hit rate remain unspecified in its abstract. |
| Result and scope | 12.1 | The consequence for ground users follows from the arithmetic of Section 12.1 and from the motion of the serving satellite. |
| Result and scope | 164 s, 1,000, 1.02–1.07 s | Bhattacherjee et al. calculate 164 s between handoffs for a ground user [2]. A 1,000-token answer at SpaceMoE’s 1.02–1.07 s per token spans about six such intervals. |
| Latency comparison | 12.5, 4,096, 2.15 GB, 34 s, 0.5 Gb/s, 172 ms, 100 Gb/s, 1.8 ms, 9.6 Tb/s | A conversation can return to its original KV-hosting satellite through ISLs each turn, paying Section 12.5’s propagation cost. Alternatively, it migrates KV state. A 4,096-token context has 2.15 GB KV. Transfer takes 34 s at 0.5 Gb/s, 172 ms at 100 Gb/s and 1.8 ms at 9.6 Tb/s. |
| Capacity and rate | 12.4 | That transfer is the one place in inference where the ISL rate of Figure 12.4 decides the outcome. It is also a transfer that still awaits evaluation. |
| Result and scope | Qualitative scope | On the ground the same KV sits in a stationary data center. |
| 結果與適用範圍 | 定性範圍 | 第三項待完成的評估是跨 ISL 的 KV 傳輸，各項已評估配置皆將 KV 留在本地。 |
| 結果與適用範圍 | 6, 4.1 | SpaceMoE 的鏈路項承載大小為 \(MQ_B\) 的隱藏狀態（Eq. 6），CollabLLM-LEO 的視覺 transformer 每張影像只做一次前向傳遞，僅執行單次前向傳遞（§III），LLMSpace 建模的是僅有本地流量的單一衛星（§4.1），而 SkyMemory 作為唯一一個流量會是 KV 大小的提案，摘要中的 ISL 速率、快取大小與命中率待補充。 |
| 結果與適用範圍 | 12.1 | 這對地面使用者的後果，來自第 12.1 節的算術與服務衛星的運動。 |
| 結果與適用範圍 | 164 s, 1.02–1.07 s, 1,000 | Bhattacherjee 等算出地面使用者與頭頂衛星之間每 164 s 換手一次[2]，而以 SpaceMoE 每 token 1.02–1.07 s 生成的 1,000 個 token 回答，大約持續六個這樣的間隔。 |
| 結果與適用範圍 | 12.5 | 因此一段 KV 誕生在某顆衛星上的對話，要麼每一輪都經 ISL 跳回那顆衛星、照第 12.5 節的算法付出傳播延遲，要麼搬動它的 KV。 |
| 容量與速率 | 4,096, 2.15 GB, 0.5 Gb/s, 34 s, 100 Gb/s, 172 ms, 9.6 Tb/s, 1.8 ms | 對 4,096 個 token 的上下文而言那是 2.15 GB，在 0.5 Gb/s 下要 34 s、100 Gb/s 下要 172 ms、9.6 Tb/s 下要 1.8 ms。 |
| 結果與適用範圍 | 12.4 | 這筆傳輸是推論中唯一由圖 12.4 的 ISL 速率決定結果的地方，也是仍待論文評估的傳輸。 |
| 結果與適用範圍 | 定性範圍 | 在地面上，同樣的 KV 待在固定位置的資料中心裡。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Duration of a 1,000-token answer at SpaceMoE's time per token, in handoff intervals以 SpaceMoE 每 token 時間生成 1,000 token 回答的時長，以換手間隔計 | \(1{,}000 \times (1.02\ldots1.07)\,\text{s} / 164\,\text{s}\) | 1,020–1,070 s ≈ 6.2–6.5 intervalscomputed. 164 s from [2]計算值。164 s 取自 [2] |
| Moving a 4,096-token KV cache (2.15 GB, 7B-class) at 0.5 Gb/s在 0.5 Gb/s 下搬動 4,096 token 的 KV 快取（2.15 GB，7B 級） | \(1.72\times10^{10}\,\text{bit} / 5\times10^{8}\,\text{bit/s}\) | 34 scomputed. KV size from Section 12.1計算值。KV 大小取自第 12.1 節 |
| Same at 100 Gb/s and at 9.6 Tb/s在 100 Gb/s 與 9.6 Tb/s 下 | \(1.72\times10^{10} / 10^{11}\) . \(1.72\times10^{10} / 9.6\times10^{12}\) | 172 ms · 1.8 mscomputed計算值 |
The open question is which site should serve each request and when. User access changes every fifteen seconds, while energy follows orbital sunlight or terrestrial grid conditions. KV remains on a satellite that leaves view within minutes. Section 12.5 makes ISL rate relevant through prefill and KV transfer. Token-sized work depends on hop count, compute and node delay. Section 13.6 develops direction E against all-ground and all-orbit baselines.
綜合起來，這三個觀察界定了推論論文列為待解的問題：在使用者鏈路每十五秒重新洗牌、能源一邊跟隨軌道一邊跟隨電網、KV 快取被釘在幾分鐘內就離開視野的衛星上的情況下，哪個站點應該在什麼時候服務哪個請求。第 12.5 節的敏感度分析補充了一點：ISL 速率只透過預填與 KV 搬移影響這個問題，而 token 大小的一切都由跳數、運算與節點延遲決定。第 13.6 節把這些觀察轉化為研究方向 E，並以全地面與全軌道兩種部署作為它的兩個基準。
Chapter summary
本章摘要
Decode traffic is small relative to prefill and KV migration. The open serving problem combines request placement, state locality, propagation and orbit-dependent energy across ground and orbital sites.
相較於預填及 KV 搬移，解碼流量較小。待解的跨站點服務問題需同時考量請求配置、狀態位置、傳播及隨軌道變化的能源。
| Item項目 | Values數值 | Conditions and source條件與來源 |
|---|---|---|
| Model and compute | 8, 7B, 28 GB | For a 7B model, decode transfers one 8 KB hidden vector per layer boundary. Data-parallel training transfers roughly 28 GB per worker per step. |
| Result and scope | 33.6 MB, 2.15 GB, 4,096 | The larger ISL demands are a 33.6 MB prefill tensor and a 2.15 GB KV cache for 4,096 tokens. Token-stream traffic is much smaller. |
| Model and compute | 3.5 B, 1,056, 1.02–1.07 s, 7.28 GFLOPS, 90–94% | SpaceMoE distributes a 3.5 B-active MoE across 1,056 satellites and reports 1.02–1.07 s per token. Its 7.28 GFLOPS board accounts for 90–94% of that time. |
| Capacity and rate | 20, 42%, 800 Mbit/s, 0.5 Gbit/s | while CollabLLM-LEO compresses ViT activations by more than 20× to beat a downlink by 42% at 800 Mbit/s over 0.5 Gbit/s ISLs |
| Model and compute | Qualitative scope | and LLMSpace, SkyMemory, Constella and the cached-model framework address carbon, KV caching, satellite roles and relaying with per-token paths still to be specified. |
| Result and scope | 0.5 Gb/s, 9.6 Tb/s, 59.7 Gb/s, 4.5 ms, 12.9 Gb/s | Sweeping 0.5 Gb/s to 9.6 Tb/s changes decode’s bottleneck from compute to propagation, while transmission remains smaller. On 4.5 ms hops, prefill becomes link-bound below 59.7 Gb/s. CollabLLM-LEO compression loses its advantage above 12.9 Gb/s. |
| Evaluation gaps | Qualitative scope | Common ground-request comparisons, orbit-dependent energy schedules and cross-ISL KV transfers remain open evaluation questions. A joint serving comparison between both sites therefore remains open. |
| 結果與適用範圍 | 7B, 8, 28 GB, 33.6 MB, 4,096, 2.15 GB | 推論跨越網路時用的是與訓練相異的貨幣：對 7B 級模型而言，一個解碼步驟每個層邊界只搬一個 8 KB 的隱藏向量，而資料平行的訓練步驟每個工作節點約搬 28 GB，所以能讓 ISL 吃緊的是 33.6 MB 的預填張量與 4,096 token 上下文的 2.15 GB KV 快取，token 串流的傳輸量則小得多。 |
| 結果與適用範圍 | 3.5 B, 1,056, 1.02–1.07 s, 90–94%, 7.28 GFLOPS | SpaceMoE 把一個 3.5 B 活躍參數的 MoE 鋪到 1,056 顆衛星上，報告每 token 1.02–1.07 s，而這個數字的 90–94% 可以由它自己 7.28 GFLOPS 的抗輻射運算解釋。 |
| 容量與速率 | 0.5 Gbit/s, 20, 800 Mbit/s, 42% | CollabLLM-LEO 在 0.5 Gbit/s 的 ISL 上把 ViT 激活值壓縮超過 20×，在 800 Mbit/s 時比下傳快 42%。 |
| 結果與適用範圍 | 定性範圍 | LLMSpace、SkyMemory、Constella 與快取模型框架則處理碳排、KV 快取、衛星角色與中繼，每 token 路徑待指定。 |
| 結果與適用範圍 | 0.5 Gb/s, 9.6 Tb/s | 把 ISL 速率在論文涵蓋的 0.5 Gb/s 到 9.6 Tb/s 之間移動，可以看到解碼的瓶頸從運算移到傳播、傳輸時間在所選範圍內較小。 |
| 容量與速率 | 4.5 ms, 59.7 Gb/s | 預填在 4.5 ms 的跳上低於 59.7 Gb/s 才會受鏈路限制。 |
| 容量與速率 | 12.9 Gb/s | CollabLLM-LEO 的壓縮在 12.9 Gb/s 以上收益降低。 |
| 結果與適用範圍 | 定性範圍 | 把這些論文讀成「拿什麼比較」，會發現其中源自地面請求的地面基準、隨軌道變化的能源模型與跨 ISL 的 KV 傳輸都待評估，雙站點服務的問題因此仍然開放。 |
Where these ideas go next
這些概念接下來用在哪裡
Chapter 13 organizes the primer’s research questions. Its Section 13.6 develops direction E from Section 12.6’s three observations. The serving design combines KV-local request routing, a latency budget incorporating the 15-second schedule and Panel 12.1’s ISL-rate sensitivity. It compares against all-ground and all-orbit baselines.
第 13 章把整份入門教材的待解問題整理成一份研究計畫，其中第 13.6 節把第 12.6 節的三個觀察作為研究方向 E：橫跨地面與軌道的服務，包括尊重 KV 位置的請求路由、建立在 15 秒排程上的延遲預算，以及面板 12.1 所量測的對 ISL 速率的敏感度，並與全地面及全軌道兩種基準比較。
Papers behind this chapter本章所依據的論文
arXiv 2605.00515SpaceMoE: Realizing Distributed Mixture-of-Experts Inference over Space Networks
arXiv 2604.04654Communication-Efficient Collaborative LLM Inference over LEO Satellite Networks
arXiv 2605.05615LLMSpace: Carbon Footprint Modeling for Large Language Model Inference on LEO Satellites
ACM HotNets 2020In-orbit Computing: An Outlandish Thought Experiment?
ACM/IFIP Middleware 2022Celestial: Virtual Software System Testbeds for the LEO Edge