Communication-Efficient Collaborative LLM Inference over LEO Satellite NetworksLEO 衛星網路上通訊高效的協同 LLM 推論
CollabLLM-LEO splits a transformer across satellites and compresses the tensors sent between layer blocks. Its comparison concerns processing satellite images in orbit or sending them to a ground GPU first. The reported reduction applies to that data path and baseline.CollabLLM-LEO 把 transformer 拆到多顆衛星上，並壓縮層區塊之間傳送的張量。其比較範圍是衛星影像先在軌處理，或先下傳到地面 GPU，報告的延遲改善須依此資料路徑及基準解讀。
ISL 速率為 0.5 Gbit/s[96, §VI-A, Tabs. II–III]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | CollabLLM-LEO connects one sensing satellite to K computing satellites. A shortest-path search over a DAG of candidate cuts splits the transformer into K sub-models. Boundary tensors pass through a learned Gumbel-Sigmoid mask, 8-bit quantization and entropy coding. Compression exceeds 20× at every boundary. The first boundary reaches 3.96× → 11.56× → 25.82× cumulatively. [96, §III, §V, Fig. 8] preprint |
| Model configuration | Vision transformers span 0.086–1.8 B parameters and process Earth-observation images. Compute timing uses four NVIDIA Jetson AGX Orin modules at 15, 30 and 50 W. Orbits and links are simulated. The ISL runs at 0.5 Gbit/s [96, §VI-A, Tabs. II–III] . |
| Delay comparison | At 800 Mbit/s ground-link rate, end-to-end delay is 42% below raw-image downlink to a ground GPU. Communication falls 71%, with accuracy loss under 1%. |
| Evaluation conditions | Table II gives a 6 Gbit/s ground link, while Fig. 4 sweeps 0.2–0.8 Gbit/s. One figure claims an impossible "103%" delay reduction. The baseline downlinks satellite-generated data. Ground-origin serving alternatives lie outside the comparison. |
| 星座配置 | CollabLLM-LEO 以星間鏈路串起一顆感測衛星與 K 顆運算衛星，把 transformer 逐層拆成 K 個子模型，拆分點由候選切點構成之 DAG 上的最短路徑搜尋決定，並以學習式 Gumbel-Sigmoid 遮罩、8 位元量化與熵編碼壓縮每個邊界張量，每個邊界都超過 20 倍，第一個邊界的累積比例為 3.96× → 11.56× → 25.82×[96, §III, §V, Fig. 8]預印本。 |
| 模型配置 | 模型是 0.086–1.8 B 參數、處理地球觀測影像的視覺 transformer。 |
| 評估條件 | 運算時間在四塊 NVIDIA Jetson AGX Orin 上以 15、30、50 W 實測，軌道與鏈路則為模擬。 |
| 延遲比較 | 頭條結果是：當地面鏈路為 800 Mbit/s 時，端到端延遲比把原始影像下傳到地面 GPU 低 42%，通訊量少 71%，準確度損失低於 1%。 |
| 星座配置 | 地面鏈路在 Table II 為 6 Gbit/s，Fig. 4 則掃描 0.2–0.8 Gbit/s，這項設定差異須先釐清。另有圖表宣稱「降低 103%」，會導致負延遲，引用前須先修正。其基準處理的是衛星產生的資料，起源於地面的服務請求則屬於另一種比較範圍。 |
1. Where this paper sits1. 這篇論文的位置
Chapter 12 compares three inference preprints with opposing resource assumptions. CollabLLM-LEO and SpaceMoE occupy opposite ends of that comparison. SpaceMoE assumes a slow rad-hard processor and "≥ 100 Gbps" ISLs. CollabLLM-LEO assumes a fast Orin-class processor and 0.5 Gbit/s ISLs.
Thus, its whole contribution is directed at shrinking what crosses the link. The paper is also the closest of the three to the orbital edge computing tradition of Chapter 8. This occurs because the data it processes is born on a satellite, an Earth-observation image from a sensing satellite.
The study compares processing satellite-generated images through an orbital chain with downlinking raw images for ground processing. That framing places it beside Kodan and the OEC work in spirit. Its pipeline-parallel split of a transformer places it with the language-model papers in method.
第 12 章把三篇推論預印本並列，是因為它們對兩項關鍵資源做了相反的假設，而 CollabLLM-LEO 正是 SpaceMoE 的鏡像。SpaceMoE 假設慢速的抗輻射處理器配上「≥ 100 Gbps」的雷射鏈路，CollabLLM-LEO 則假設快速的 Orin 等級處理器配上 0.5 Gbit/s 的鏈路，所以它的整個貢獻都在縮小跨越鏈路的東西。這篇論文也是三篇中最接近第 8 章在軌邊緣運算傳統的一篇，因為它處理的資料誕生在衛星上，是感測衛星拍下的地球觀測影像，而它回答的問題是：這張影像應該先在軌道上由一串鄰居處理再往下送，還是下傳原始影像後在地面處理。這個框架在精神上把它放在 Kodan 與 OEC 旁邊，而它對 transformer 的管線平行拆分，則在方法上把它歸入語言模型論文之列。
2. The problem, measured2. 問題，用量測說話
The problem is that a large model exceeds the available capacity on one satellite and a raw image is expensive to downlink. Thus, the authors split the model over several satellites and must then pay for the tensors that cross every split. Each computing satellite has 8 GB of memory [96, Tab. II] . Thus, the largest model exceeds one board's memory and therefore requires splitting.
This is the situation the paper is built for. "LLM" in the title refers to the transformer family and pipeline technique. Chapter 12 therefore treats the workload as image inference. This occurs because a vision transformer emits an activation row per image patch.
This makes its boundary tensors image-sized rather than token-sized.
問題在於大模型超出單顆衛星記憶體容量，而原始影像下傳的代價又高，所以作者把模型拆到數顆衛星上，接著就必須為跨越每個拆分點的張量付出代價。每顆運算衛星有 8 GB 記憶體[96, Tab. II]，所以最大的模型超出單板記憶體容量，必須拆分，這正是論文所設想的情境。因此標題中的「LLM」在本文指 transformer 家族與管線技術，工作負載則為視覺模型。第 12 章須保留此區分：視覺 transformer 對每個影像區塊輸出一列激活值，邊界張量依影像規模增加，語言模型的解碼張量則依 token 定義。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Model configuration | The models are vision transformers |
| ViT-B | from ViT-B at 0.086 B parameters and about 2 GB |
| ViT-L | through ViT-L at 0.307 B |
| ViT-H | and ViT-H at 0.632 B |
| ViT-G | to ViT-G at 1.8 B parameters and more than 12 GB |
| EuroSAT dataset | run on EuroSAT with 27,000 images in 10 classes |
| RESISC45 dataset | and RESISC45 with 31,500 images in 45 classes [96, Tab. III] preprint. |
| 模型配置 | 模型是視覺 transformer：從 0.086 B 參數、約 2 GB 的 ViT-B，經 0.307 B 的 ViT-L 與 0.632 B 的 ViT-H，到 1.8 B 參數、超過 12 GB 的 ViT-G，資料集為 27,000 張影像、10 類的 EuroSAT 與 31,500 張影像、45 類的 RESISC45[96, Tab. III]預印本。 |
3. The design3. 設計
The system has three parts that fit together. The sensing satellite \(S_0\) captures an image. The chain contains \(K\) computing satellites. Satellites \(S_1 \ldots S_K\) each run one contiguous layer block.
Candidate cuts form a DAG weighted by compute and transmission time. A shortest path selects the split points. Multithreading overlaps next-image computation with transmission of the previous tensor. [96, §III-A, §V] [96, §III] Every boundary tensor passes through activation compression. A Gumbel-Sigmoid mask relaxes the keep-or-drop decision.
Training uses a straight-through estimator, sparsity loss and temperature annealing. Retained elements receive 8-bit quantization, followed by entropy coding. [96, Eqs. 1–5] [96, Eq. 6]
系統由三個互相配合的部分組成。第一是鏈：感測衛星 \(S_0\) 拍攝影像，\(K\) 顆運算衛星 \(S_1 \ldots S_K\) 各自執行一段連續的層，拆分點的選法是把每個候選切點建模為有向無環圖中的一條邊，以運算與傳輸時間為權重並取最短路徑[96, §III-A, §V]，並以多執行緒讓運算與傳輸重疊，使一顆衛星在送出上一個張量的同時處理下一張影像[96, §III]。第二是在每個邊界施加的激活值壓縮（activation compression）：先以 Gumbel-Sigmoid 遮罩將張量稀疏化，這是「保留或丟棄」決策的可微分鬆弛，以直通估計器、稀疏損失與溫度退火訓練[96, Eqs. 1–5]，再把保留下來的元素量化到 8 位元[96, Eq. 6]，最後做熵編碼。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | The third part is the operating point: Tables IV–V use 80% sparsity and 8-bit quantization. First-boundary compression reaches 3.96× after masking, 11.56× after quantization and 25.82× after entropy coding. Other boundaries reach 25.60×, 24.14× and 23.28×. The final two stages add factors of about 2.9 and 2.2. [96, Fig. 8] |
| 評估條件 | 第三是操作點：Tables IV–V 使用 80% 稀疏度與 8 位元量化，在第一個邊界上遮罩後為 3.96×、量化後 11.56×、熵編碼後 25.82×，其他三個邊界為 25.60×、24.14× 與 23.28×，所以每個階段各再貢獻約 2.9 與 2.2 倍[96, Fig. 8]。 |
The reason this design earns its gain is specific to its data and its baseline. Compression attacks the transmission term of a link, the bytes divided by the rate, and leaves the propagation term untouched. Thus, it pays off when tensors are large and links are slow. This is exactly the regime of image-sized vision-transformer activations on a 0.5 Gbit/s ISL.這個設計之所以有收益，與它的資料和基準密切相關。壓縮處理的是鏈路的傳輸項，也就是位元組數除以速率，傳播項則完全維持原水準，所以它在張量大、鏈路慢時才划算，而這正是影像大小的視覺 transformer 激活值跑在 0.5 Gbit/s ISL 上的情況。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | At 25.82× a 0.5 Gbit/s link carries the uncompressed equivalent of about 12.9 Gbit/s. Thus, a chain of slow links can beat a single downlink of the raw image at a few hundred Mbit/s. |
| Model configuration | For a token-sized decode tensor of a language model the same coder would shave microseconds off a hop whose propagation costs milliseconds. This is the point Chapter 12 makes when it reads this design against the inference traffic of Section 12.1. |
| 壓縮 | 在 25.82× 之下，0.5 Gbit/s 的鏈路等同承載約 12.9 Gbit/s 的原始精度資料，這就是一串慢鏈路能勝過以數百 Mbit/s 單次下傳原始影像的原因。 |
| 模型配置 | 對語言模型 token 大小的解碼張量而言，同樣的編碼器只能在傳播就要花上毫秒的一跳中省下微秒，這正是第 12 章把此設計與第 12.1 節的推論流量對照時所指出的重點。 |
Original figure note原始圖說出處
Drawn for this primer from Zhang et al., arXiv 2604.04654, §III, §VI-A, Table II and Fig. 8.4. Evaluation setup4. 實驗設定
The evaluation is a hybrid of measured compute and simulated network. The topology is a chain of five satellites drawn from a single-plane Walker Delta of twelve satellites. Thus, every ISL is intra-plane and the neighbors keep a fixed separation, which avoids re-pointing and handover by construction. The links enter the model through their rates.
Queueing, retransmission and link-failure models lie outside this paper. Among the three inference preprints, the fact base identifies link-failure modeling only in SpaceMoE. Flight hardware, thermal behavior and battery dynamics lie outside the evaluation.
評估是實測運算與模擬網路的混合。拓樸是從十二顆衛星的單一軌道面 Walker Delta 中取出的五顆衛星鏈，因此每條 ISL 都是同軌鏈路，相鄰衛星保持固定間距，在設計上就避開了重新指向與換手。鏈路透過其速率進入模型，事實庫將排隊、重傳與鏈路故障列於本文模型範圍之外，這與「三篇推論預印本中只有 SpaceMoE 模擬鏈路故障」的發現一致。飛行硬體、熱行為與電池動態皆位於評估範圍之外。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Power budget | Four NVIDIA Jetson AGX Orin modules represent the computing satellites. Each offers up to 275 TOPS and is timed at 15, 30 and 50 W. The ground server uses an RTX 4070 Ti. Orbits, visibility and links are simulated. [96, §VI-A, Tab. II] preprint |
| Link or data rate | The ISL rate is 0.5 Gbit/s, described in the text as a 0.5 GHz free-space-optical bandwidth. The satellite-to-ground rate is 6 Gbit/s in Table II while Fig. 4 sweeps it over 0.2–0.8 Gbit/s. |
| 功率預算 | 運算時間在四塊各達 275 TOPS 的 NVIDIA Jetson AGX Orin 上以 15、30、50 W 功率模式實測，代表四顆運算衛星，地面伺服器則是一張 RTX 4070 Ti，而軌道、可見性與鏈路都是模擬[96, §VI-A, Tab. II]預印本。 |
| 鏈路或資料速率 | ISL 速率為 0.5 Gbit/s，內文描述為 0.5 GHz 的自由空間光學頻寬。 |
| 衛星對地速率 | 衛星對地速率在 Table II 為 6 Gbit/s，而 Fig. 4 則在 0.2–0.8 Gbit/s 之間掃描。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| More than 20× activation compression on every boundary. 3.96× → 11.56× → 25.82× on boundary 1 [96, Fig. 8].每個邊界的激活值壓縮都超過 20 倍。第 1 個邊界為 3.96× → 11.56× → 25.82×[96, Fig. 8]。 | 80% sparsity, 8-bit. Measured on ViT activations, which are image-sized. A safe quotation per fact base Part C, item 39.80% 稀疏度、8 位元。在影像大小的 ViT 激活值上量得。依事實庫 Part C 第 39 項為可安全引用的數字。 |
| 42% lower end-to-end delay than ground-only processing at a ground-link rate of 800 Mbit/s [96, Fig. 4].在地面鏈路 800 Mbit/s 時，端到端延遲比只在地面處理低 42%[96, Fig. 4]。 | Baseline downlinks the raw image to an RTX 4070 Ti. At Table II's 6 Gbit/s the comparison would be different and is unreported.基準把原始影像下傳到 RTX 4070 Ti。若以 Table II 的 6 Gbit/s，比較結果會各異，相關結果仍待論文報告。 |
| "Over 100%" improvement at low ground-link rates [96, Fig. 4]. "103% reduction" against a heuristic split [96, Fig. 12].在低地面鏈路速率下「超過 100%」的改善[96, Fig. 4]。相對啟發式拆分「降低 103%」[96, Fig. 12]。 | A reduction above 100% implies negative delay. The figure therefore requires correction and falls outside the results quoted here.降低超過 100% 會導致負延遲，因此這個數字須先修正，本導讀將其列於有效結果範圍之外。 |
| 71% less communication and under 1% accuracy loss (abstract. Introduction gives 71.7%).通訊量少 71%，準確度損失低於 1%（摘要。導論寫 71.7%）。 | The 71% is end-to-end against the baseline and has a separate definition from the per-boundary 20×. The text gives a Top-k loss of 4.1% where Table IV shows about 0.3%.71% 是相對基準的端到端數字，與每個邊界的 20 倍各有定義。內文給出 Top-k 損失 4.1%，Table IV 卻約為 0.3%。 |
| More than 194 of 200 split strategies stay within 1.0% of baseline accuracy [96, Fig. 10].200 種拆分策略中超過 194 種的準確度落在基準的 1.0% 以內[96, Fig. 10]。 | Shows the compression is robust to where the cuts fall, on two remote-sensing classification datasets.顯示在兩個遙測分類資料集上，壓縮對切點位置具穩健性。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| What does each stage add on boundary 1?第 1 個邊界上每個階段各貢獻多少？ | \( \dfrac{11.56}{3.96},\qquad \dfrac{25.82}{11.56} \) | 2.9× (quantise) · 2.2× (entropy)computed from Fig. 8由 Fig. 8 計算 |
| What uncompressed rate does a 0.5 Gbit/s ISL carry at 25.82×?0.5 Gbit/s 的 ISL 在 25.82× 下等同承載多少原始精度速率？ | \( 0.5\ \text{Gbit/s} \times 25.82 \) | ≈ 12.9 Gbit/scomputed. ISL rate from Table II, ratio from Fig. 8計算值。ISL 速率取自 Table II，比例取自 Fig. 8 |
| How does the ground link compare with the ISL in each setting?在各設定中地面鏈路與 ISL 相比如何？ | \( \dfrac{6}{0.5},\qquad \dfrac{0.2}{0.5}\;\ldots\;\dfrac{0.8}{0.5} \) | 12× [96, Tab. II] · 0.4–1.6× [96, Fig. 4]computed. The two settings disagree by an order of magnitude計算值。兩種設定相差一個數量級 |
| How far apart are the two papers' processors?兩篇論文的處理器相差多遠？ | \( \dfrac{275\times10^{12}}{7.28\times10^{9}} \) | ≈ 3.8 × 10⁴computed. Orin peak TOPS [96, Tab. II] vs SpaceMoE's effective GFLOPS. Integer TOPS and float FLOPS are distinct measurement quantities, so read it as an order of magnitude計算值。Orin 峰值 TOPS[96, Tab. II]對 SpaceMoE 的有效 GFLOPS。整數 TOPS 與浮點 FLOPS 並各異價，應視為數量級 |
| What does the four-satellite chain draw at full power?四顆衛星的鏈在全功率下耗電多少？ | \( 4 \times 50\ \text{W} \) | 200 Wcomputed. 50 W mode from Table II. Excludes ISL terminals and bus計算值。50 W 模式取自 Table II。計算範圍排除 ISL 終端與平台 |
6. Limits and open questions6. 限制與待解的問題
- The baseline sends satellite data to the ground for processing. The comparison is between processing satellite-born data in orbit and downlinking it first. Ground-origin requests lie outside the comparison. Their inputs start in terrestrial data centers, requiring a separate comparison with a ground-side serving design.基準是先下傳衛星資料，地面端服務方案則屬於另一種比較範圍。比較的是「在軌處理衛星產生的資料」與「先下傳」，所以起源於地面之請求的地面端服務替代方案，位於論文比較範圍之外。在那種情況下輸入從地面資料中心開始，軌道必須和它競爭才能證明自己的價值。
- Energy is measured as power modes, outside the model with the orbit. The 15/30/50 W timings come from laboratory Jetsons. Eclipse, battery depth of discharge and thermal throttling lie outside the model. Xing et al. observed throttling within 50 minutes on a 5 W board in orbit. The evaluation leaves joint energy and orbital dynamics open. Chapter 12 records Orin-class CubeSat flights, while sustained 50 W operation still requires flight evidence.能源評估依固定功率模式量測，與軌道動態的聯合模型則仍待建立。15/30/50 W 模式是在實驗室的 Jetson 上計時，日蝕、電池放電深度，以及 Xing 等人在軌量到 5 W 板子 50 分鐘內就降頻的熱效應皆位於範圍之外，所以評估將能源與軌道動態的聯合模型留待後續。第 12 章也指出 Orin 等級模組曾在 CubeSat 上飛行，持續 50 W 的工作週期則仍待飛行證據。
- A gentle network. A single-plane chain keeps neighbor distances fixed. Thus, the model holds pointing and links fixed, sets setup delay to zero and excludes handovers, failures and queueing. The 0.5 Gbit/s rate is conservative against the "≥ 100 Gbps" of SpaceMoE. This makes the three papers disagree by more than 200× on what an ISL carries.溫和的網路。單一軌道面的鏈讓相鄰距離固定，所以重新指向、建立延遲、換手、鏈路故障及排隊皆位於模型範圍之外。0.5 Gbit/s 的速率相對 SpaceMoE 的「≥ 100 Gbps」相當保守，使三篇論文對 ISL 承載量的假設相差超過 200 倍。
- Internal inconsistencies. Table II gives a 6 Gbit/s ground link, while Fig. 4 uses 0.2–0.8 Gbit/s. A delay reduction of "103%" in Fig. 12 exceeds the mathematical limit. The model names switch between ViT-B/L/H/G and ViT-Tiny/Base/Large. And the Top-k loss is 4.1% in the text against about 0.3% in Table IV (fact base Part C, item 39). The safe quotations are the 42% at 800 Mbit/s and the greater-than-20× compression.內部差異。地面鏈路在 Table II 為 6 Gbit/s、在 Fig. 4 為 0.2–0.8 Gbit/s。Fig. 12 的「降低 103%」會導致負延遲。模型名稱在 ViT-B/L/H/G 與 ViT-Tiny/Base/Large 之間切換。Top-k 損失在內文為 4.1%、在 Table IV 約為 0.3%（事實庫 Part C，第 39 項）。可安全引用的是 800 Mbit/s 時的 42% 與超過 20 倍的壓縮。
- Vision-transformer image classification. The workload is single-pass image classification. Autoregressive decoding, KV-cache behavior and per-token latency lie outside its scope. Chat requests require separate evaluation because their smaller decode tensors make propagation dominate each hop.視覺 transformer 的影像分類。工作負載是單次通過的影像分類，所以自迴歸解碼、KV 快取及每 token 延遲皆位於範圍之外，結果須另行評估才能移轉到回應聊天請求的服務。後者的解碼張量小到每一跳都主要由傳播主導。
7. How it connects7. 與其他內容的關聯
- Section 12.3: the chapter's treatment of CollabLLM-LEO with Figure 12.3, and the observation that compression attacks only the transmission term.第 12.3 節：該章對 CollabLLM-LEO 的討論與圖 12.3，以及「壓縮只處理傳輸項」的觀察。
- Section 12.2: SpaceMoE, the mirror-image assumption of a slow processor and a fast link.第 12.2 節：SpaceMoE，慢處理器配快鏈路的鏡像假設。
- Section 12.4: LLMSpace, SkyMemory, Constella and the cached-model framework, the rest of the inference set.第 12.4 節：LLMSpace、SkyMemory、Constella 與快取模型框架，推論論文組的其餘部分。
- Section 12.5: Panel 12.1 places CollabLLM-LEO at the link-bound corner, where activation bytes per boundary and the ISL rate decide the time.第 12.5 節：面板 12.1 把 CollabLLM-LEO 放在受鏈路限制的那一角，時間由每個邊界的激活值位元組數與 ISL 速率決定。
- Section 12.6: why a downlink baseline for satellite-born data differs from a ground-side alternative for ground-born requests.第 12.6 節：為什麼衛星產生之資料的下傳基準，與地面產生之請求的地面端替代方案各異。
- Section 10.6: a 0.5 Gbit/s chain is the constellation-wide regime at its slowest, far from the Tbps formation regime.第 10.6 節：0.5 Gbit/s 的鏈是星座尺度區間中最慢的情況，與 Tbps 的編隊區間相距甚遠。
- Section 13.6: research direction E asks for the ground baseline and the joint time-and-energy model this paper lacks.第 13.6 節：研究方向 E 要求的正是本文缺少的地面基準與時間能源聯合模型。
- SpaceMoE [78] and LLMSpace [97] are the sibling digests. Kodan [83] and OEC [60] frame the same process-before-downlink question for smaller models. Xing et al. [85] measures the thermal limit the Jetson emulation leaves out.SpaceMoE[78]與 LLMSpace[97]是姊妹導讀。Kodan[83]與 OEC[60]針對較小的模型提出了同樣的「先處理再下傳」問題。Xing 等人[85]量測了 Jetson 模擬所遺漏的熱限制。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 1 + K · 5 of 12 | One sensing and K computing satellites. A five-satellite chain from a 12-satellite single plane [96, Tab. II] | 一顆感測與 K 顆運算衛星。從 12 顆單一軌道面中取出的五顆鏈[96, Tab. II] | |
| 0.086–1.8 B | ViT-B to ViT-G parameters, about 2 GB to more than 12 GB [96, Tab. III] | ViT-B 到 ViT-G 的參數量，約 2 GB 到超過 12 GB[96, Tab. III] | |
| 4 × Orin · 275 TOPS · 15/30/50 W | Emulated computing satellites, 8 GB each [96, Tab. II] | 模擬的運算衛星，各 8 GB[96, Tab. II] | |
| 0.5 Gbit/s | ISL rate [96, Tab. II] | ISL 速率[96, Tab. II] | |
| 6 vs 0.2–0.8 Gbit/s | Ground link in Table II vs the Fig. 4 sweep, unreconciled | Table II 與 Fig. 4 掃描中的地面鏈路，仍待釐清 | |
| > 20× · 3.96 → 11.56 → 25.82× | Activation compression on every boundary. Cumulative on boundary 1 [96, Fig. 8] | 每個邊界的激活值壓縮。第 1 個邊界的累積值[96, Fig. 8] | |
| −42% at 800 Mbit/s | End-to-end delay vs downlinking the raw image [96, Fig. 4] | 相對下傳原始影像的端到端延遲[96, Fig. 4] | |
| −71% · < 1% | Communication reduction. Accuracy loss (abstract) | 通訊量減少。準確度損失（摘要） |