Digest · AI clusters and inference in orbit · read after Chapter 10
導讀 · 在軌 AI 叢集與推論 · 第 10 章後閱讀
The Cost and Network Limits of Space-Based AI Compute太空 AI 運算的成本與網路極限
At the evaluated scale, a laser torus has far less bisection bandwidth than a terrestrial Clos. Distributed training reaches this network limit. Inference confined to one satellite stays on a local roofline, while the comparison depends on stated topology and hardware assumptions.在評估規模下，雷射環面的對分頻寬遠低於地面 Clos，使分散式訓練受到網路上限限制。限於單一衛星的推論則維持本地 roofline，比較結果須依拓樸及硬體假設解讀。
If you read nothing else
只讀這一段也可以
van Berkel asks whether a nearest-neighbor laser-connected orbital cluster can train a frontier model. Roofline calculations evaluate N = 8,000 nodes. [101, §IV] preprintvan Berkel 問的是：一個以星座方式佈線、每顆衛星以雷射連到最近鄰居的軌道叢集，其前沿模型訓練能力為何。他以 N = 8,000 個節點的 roofline 算術回答[101, §IV]preprint。而且軌道幾何位於模型範圍之外，因此 8,000 顆衛星環面配置的飛行可行性仍待檢查。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Terrestrial Clos reference | The terrestrial Clos has 8,000 racks, 72 links per top-of-rack switch and 2:1 oversubscription. At 100 GB/s per link, bisection bandwidth is 28,800 TB/s. |
| 2-D torus | In comparison, a 2-D torus of 89 × 90 satellites has 2.25 TB/s |
| 3-D torus | and a 3-D torus of 20 × 20 × 20 has 10 TB/s when each laser link carries 100 Gb/s |
| Bandwidth ratios | ratios of 1 : 12,880 : 2,880 |
| Fabric latency | and latency across the fabric is 4 µs for the Clos against 1,800 µs and 600 µs for the two tori [101, Tab. 3]. |
| Training comparison | Data-parallel training needs about 500 kFLOP per byte crossing the bisection. Torus throughput falls two to three orders of magnitude, raising training cost "100+" times. Single-satellite inference retains the terrestrial roofline. [101, §IV-C, Figs. 7–8] |
| Training comparison | The verdict is that frontier training in space within 2–3 years is "not credible" and that a 10-year horizon "appears optimistic" [101, §V]. |
| Constellation configuration | Two cautions apply before quoting it: the text and Table 3 give different formulas for the 2-D torus bisection, a factor of two apart. Use the Table 3 values. Orbital geometry lies outside the model, leaving flight feasibility of an 8,000-satellite torus open. |
| 地面 Clos 基準 | 一個由 8,000 個機架組成、每個機架頂交換器 72 條鏈路、鏈路 100 GB/s、2:1 超額訂閱的地面 Clos，對分頻寬為 28,800 TB/s。 |
| 星座配置 | 而當每條雷射鏈路承載 100 Gb/s 時，89 × 90 顆衛星的二維環面只有 2.25 TB/s，20 × 20 × 20 的三維環面為 10 TB/s，比例為 1 : 12,880 : 2,880。 |
| 延遲比較 | 跨越整個架構的延遲，Clos 為 4 µs，兩種環面各自為 1,800 µs 與 600 µs[101, Tab. 3]。 |
| 訓練比較 | 因為資料平行訓練每跨越對分的一個位元組需要約 500 kFLOP 的運算，環面在有效吞吐量上損失二到三個數量級，訓練成本提高「100 倍以上」，而限於單一衛星的推論看到的則與地面 roofline 完全相同[101, §IV-C, Figs. 7–8]。 |
| 訓練比較 | 結論是 2–3 年內在太空進行前沿訓練「不可信」，10 年的時程「顯得樂觀」[101, §V]。 |
| 對分公式差異 | 引用前有兩點要注意：內文與 Table 3 對二維環面對分頻寬給出相差兩倍的各異公式，應採用 Table 3 的數值。 |
1. Where this paper sits1. 這篇論文的位置
Pénot and Balakrishnan solve orbital geometry and map a Clos onto line-of-sight links, leaving rates open. The companion analysis supplies rates and latency, leaving geometry open. Together they frame Chapter 13 research direction B.本文正是把這個比較變成軌道架構數字的那一篇。它在第三部分的搭檔是 Pénot 與 Balakrishnan 的預印本，後者解決了軌道幾何並把 Clos 對映到視線鏈路上，速率則位於其範圍之外。本文則附上速率與延遲幾何則位於其範圍之外，兩者合起來框出了第 13 章研究方向 B 所處理的問題。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Path length and hops | Chapter 10 gives terrestrial references: 400 Gbps per GPU in Meta's Llama 3 cluster and 3.2 Tbps backend bandwidth per Alibaba HPN host. It also describes three-layer Clos fabrics with full pod bisection and multi-hop latency of "tens of microseconds". This paper is the one that turns that comparison into numbers for an orbital fabric. |
| Constellation configuration | The reference is Google's Suncatcher: 81 satellites inside 1 km radius at 650 km, spaced "a few 100m" apart. The analysis scales this to 8,000 nodes, a dedicated terrestrial training-cluster scale. [101, §I-B, §IV-A] |
| Evaluation conditions | It is a preprint released on 15 July 2026 and awaits peer review. |
| 地面網路基準 | 第 10 章設定了任何軌道叢集都必須拿來比較的地面基準：Meta Llama 3 叢集中每顆 GPU 400 Gbps、Alibaba HPN 每台主機 3.2 Tbps 的後端頻寬、pod 內全對分的三層 Clos 架構，以及「數十微秒」的多跳延遲。 |
| 星座配置 | 它以 Google 的 Suncatcher 叢集為參考設計，81 顆衛星位於 650 km、半徑 1 km 之內、相距「幾百公尺」，並把問題擴大到 8,000 個節點，也就是地面業者會為訓練建造專用叢集的規模[101, §I-B, §IV-A]。 |
| 評估條件 | 本文是 2026 年 7 月 15 日發布的預印本，同儕審查狀態待確認。 |
2. The problem, measured2. 問題，用量測說話
The problem is that satellite networks are wired as meshes. A laser terminal points to one partner, and a satellite has few terminals. Terrestrial training clusters instead use Clos fabrics. Its bisection grows with the number of nodes.
Bisection intensity measures training floating-point operations per byte crossing the bisection. It connects compute demand to network supply. The question becomes whether the torus can deliver bytes fast enough to keep the accelerators busy.
問題在於衛星網路是以網格方式佈線，因為一具雷射終端一次只能對準一個夥伴，而每顆衛星只搭載少數幾具。訓練叢集則以 Clos 架構佈線，其對分頻寬隨節點數成長。連接兩者的量是本文所稱的對分強度，也就是一個訓練步驟中每個必須跨越對分的位元組所執行的浮點運算數，問題於是變成：環面送出位元組的速率可支撐多少加速器運算。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | The paper fixes the satellite side at 4–6 laser ISLs per satellite, four for a 2-D torus and six for a 3-D torus |
| Per-link rate | each running at the up to 100 Gb/s claimed for Starlink's V2 lasers, which it enters in Table 3 as 12.5 GB/s |
| Beam-steering requirement | An 8,000-node torus needs adaptive steering of 32,000–48,000 beams. The authors call orbital wrap-around links "particularly challenging". [101, §II-D, §IV-A] |
| Evaluation conditions | The terrestrial side is a Clos of 8,000 racks with top-of-rack width W = 72, 100 GB/s links and 2:1 oversubscription. |
| 星座配置 | 本文把衛星端固定為每顆衛星 4–6 條雷射 ISL（二維環面 4 條、三維環面 6 條），每條以 Starlink V2 雷射所宣稱的最高 100 Gb/s 運作，在 Table 3 中記為 12.5 GB/s。 |
| 評估條件 | 並指出一個 8,000 節點的環面需要自適應操控 32,000–48,000 道光束，而環面的環繞鏈路在軌道上「尤其具挑戰性」[101, §II-D, §IV-A]。 |
| 評估條件 | 地面端是由 8,000 個機架組成的 Clos，機架頂寬度 W = 72、鏈路 100 GB/s、2:1 超額訂閱。 |
3. The design3. 設計
The method is a two-dimensional roofline. A conventional roofline limits throughput by peak compute or memory bandwidth times arithmetic intensity. The network version uses bisection bandwidth times bisection intensity as its second ceiling. [101, §IV-C]方法是二維的 roofline。在傳統 roofline 中，可達吞吐量是峰值運算與「記憶體頻寬乘以算術強度」兩者中較小者。這裡把同樣的建構套用到網路上，以「對分頻寬乘以對分強度」作為第二道天花板[101, §IV-C]。延遲是跳數乘以每跳延遲 D：Clos 為 log(N)·D、D = 1 µs。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Training comparison | Training costs k = 6 FLOP per parameter per token. Each parameter's all-reduce gradient occupies about c = 12 bytes. One gradient exchange serves a global batch of about one million tokens. This gives I_bisect = (k/c)·B_tokens ≈ 500 kFLOP/B. The calculation assumes only data-parallel all-reduce traffic crosses the bisection. [101, Eqs. 2–3] |
| Link or data rate | Table 3 gives Clos bisection ½·N·W·B_l. The 2-D torus gives 2·N^{1/2}·B_l, and the 3-D torus gives 2·N^{2/3}·B_l. Here B_l is per-link rate. |
| Latency comparison | Latency is hop count times per-hop delay D. Clos uses log(N)·D with D = 1 µs, giving 4 µs. The 2-D torus uses N^{1/2}·D, giving 1,800 µs at D = 20 µs. The 3-D torus uses (3/2)·N^{1/3}·D, giving 600 µs at the same delay. [101, Tab. 3] |
| 訓練比較 | 對分強度由訓練算術推得：每個參數每個 token 花 k = 6 FLOP，每個參數的梯度在 all-reduce 中佔約 c = 12 位元組，而約一百萬個 token 的全域批次共用一次梯度交換，所以 I_bisect = (k/c)·B_tokens ≈ 500 kFLOP/B，前提是作者明言的假設：資料平行的 all-reduce 是唯一跨越對分的流量[101, Eqs. 2–3]。 |
| 對分頻寬公式 | 對分頻寬來自 Table 3 的拓樸公式：Clos 為 ½·N·W·B_l、二維環面為 2·N^{1/2}·B_l、三維環面為 2·N^{2/3}·B_l，其中 B_l 為每條鏈路的速率。 |
| 評估條件 | 二維環面為 N^{1/2}·D、三維環面為 (3/2)·N^{1/3}·D，D = 20 µs。 |
| 評估條件 | 得到 4、1,800 與 600 µs[101, Tab. 3]。 |
Placed on these rooflines, the two workloads separate cleanly. Inference that fits inside one satellite keeps all traffic local to its satellite. Thus, its roofline is identical to the terrestrial one [101, §IV-D, Fig. 7] . The network ceiling limits torus-distributed training.
All-reduce crosses the bisection once per dimension. Bisection bandwidth falls two to three orders of magnitude, raising training cost "100+" times. Google's proposed 10 Tb/s links form a dotted roofline. Even these Suncatcher-class links leave a network gap. [101, §IV-E, §V, Fig. 8] The paper proposes a bi-planar "Clos satellite constellation" with communication satellites switching for compute satellites.
It describes the idea and leaves analysis for later work. [101, §V]
放到這些 roofline 上，兩種工作負載清楚分開。能放進單一衛星的推論所有流量皆留在單一衛星內，所以它的 roofline 與地面完全相同[101, §IV-D, Fig. 7]。分散在環面上的訓練則受網路天花板限制，而且因為環面 all-reduce 每個維度都要跨越一次對分，結果是對分頻寬損失二到三個數量級、訓練成本提高「100 倍以上」。Google 所設想的每條自由空間光學鏈路 10 Tb/s 被畫成虛線 roofline，用來顯示即使是 Suncatcher 等級的鏈路也只能縮小、差距仍持續存在[101, §IV-E, §V, Fig. 8]。本文最後提出一種雙平面的「Clos 衛星星座」，由獨立的通訊衛星擔任運算衛星的交換節點，這個構想只被描述分析則仍待完成[101, §V]。
Original figure note原始圖說出處
Drawn for this primer from van Berkel, arXiv 2607.14172, Table 3, §IV-A–B and Fig. 8. Llama 3 reference from Chapter 10. Dashed markers computed in the calc table.4. Evaluation setup4. 實驗設定
This is a purely analytical paper: its evidence consists of analytical formulas, with simulation, emulation, bench measurements and flight data outside its scope. Every result comes from closed-form topology formulas placed on roofline plots. The qualitative failure model covers permanent satellite failures, rebalancing, rerouting, debris and Kessler cascades. Direction-dependent capacity complicates repair, while availability figures remain open. [101, §IV-A]這是一篇純分析的論文：證據範圍為解析公式，模擬、模擬器、實驗台量測及飛行資料皆位於範圍之外，每個結果都來自放在 roofline 圖上的封閉形式拓樸公式。失效模型是定性的：衛星可能永久失效，迫使在各方向容量各異的網路上重新平衡與重新路由，並點名太空碎片與 Kessler 連鎖效應為風險，但可用度數字仍待補充[101, §IV-A]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Link or data rate | Each ISL carries 100 Gb/s, entered as 12.5 GB/s. The 2-D torus has 4 links per satellite, and the 3-D torus has 6. Google's proposed 10 Tb/s appears only as a dotted roofline. [101, §II-D, §IV-A, Fig. 8] |
| Latency comparison | Latency multiplies hop count by an assumed delay: 1 µs for Clos and 20 µs for tori. Separate propagation, queueing and acquisition terms lie outside the model. |
| Evaluation conditions | The 2-D torus has 89 × 90 nodes, and the 3-D torus has 20 × 20 × 20. The baseline Clos has 8,000 racks, W = 72, 100 GB/s links and 2:1 oversubscription. |
| Evaluation conditions | Economic inputs assume Starship launch at $200/kg and about 10,000 t for a 1 GW constellation. Launch cost is therefore about $2 B. Radiation is about 150 rad/yr, totaling 750 rad over five years. [101, §III-A, §III-D] |
| 鏈路或資料速率 | ISL 速率是每條雷射鏈路 100 Gb/s，記為 12.5 GB/s，二維環面每顆衛星 4 條、三維環面 6 條，Google 所設想的每條 10 Tb/s 只以虛線 roofline 出現[101, §II-D, §IV-A, Fig. 8]。 |
| 延遲比較 | 延遲模型是跳數乘以假設的每跳延遲，Clos 中為 1 µs、環面中為 20 µs，獨立的傳播、排隊及擷取項皆位於模型範圍之外。 |
| 評估條件 | 拓樸是 89 × 90 的二維環面、20 × 20 × 20 的三維環面，以及作為地面基準、由 8,000 個機架組成、W = 72、鏈路 100 GB/s、2:1 超額訂閱的 Clos。 |
| 評估條件 | 經濟輸入是 Starship 每公斤 200 美元的發射價格、1 GW 星座約 10,000 公噸，因此約 20 億美元的發射費用，以及每年約 150 rad、五年 750 rad 的總劑量[101, §III-A, §III-D]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Bisection 28,800 : 2.25 : 10 TB/s for Clos : 2-D torus : 3-D torus at N = 8,000. Ratios 1 : 12,880 : 2,880 [101, Tab. 3].N = 8,000 時對分頻寬 Clos : 二維環面 : 三維環面 = 28,800 : 2.25 : 10 TB/s。比例 1 : 12,880 : 2,880[101, Tab. 3]。 | The text says "about three orders of magnitude". The 2-D formula differs by 2× between text and Table 3, and Table 3 is used here.內文說「約三個數量級」。二維公式在內文與 Table 3 之間相差 2 倍，此處採用 Table 3。 |
| Latency 4 : 1,800 : 600 µs. Ratios 1 : 450 : 150 [101, Tab. 3].延遲 4 : 1,800 : 600 µs。比例 1 : 450 : 150[101, Tab. 3]。 | Driven by the assumed per-hop delays (1 versus 20 µs) as much as by hop counts.由假設的每跳延遲（1 對 20 µs）驅動的與跳數同樣重要。 |
| Bisection intensity ≈500 kFLOP/B for data-parallel training [101, §IV-C, Eqs. 2–3].資料平行訓練的對分強度約 500 kFLOP/B[101, §IV-C, Eqs. 2–3]。 | Assumes the gradient all-reduce is the only bisection traffic, with k = 6, c ≈ 12 B/param and a 1 M-token batch.假設梯度 all-reduce 是唯一的對分流量，k = 6、c ≈ 12 B/參數、批次 100 萬 token。 |
| Inference confined to one satellite has the terrestrial roofline [101, §IV-D, Fig. 7].限於單一衛星的推論具有與地面相同的 roofline[101, §IV-D, Fig. 7]。 | True by construction, since all traffic remains local. Multi-satellite inference requires a separate model.依構造成立，因為流量皆留在單一衛星內。多衛星推論則須另行建模。 |
| Torus training loses 2–3 orders of magnitude of bisection. "100+" times training cost [101, §IV-E, §V, Fig. 8].環面訓練損失二到三個數量級的對分頻寬。訓練成本「100 倍以上」[101, §IV-E, §V, Fig. 8]。 | The 3-D torus Table 3 ratio of 2,880 is above three orders of magnitude, so the "2–3" phrasing is the conservative reading.三維環面 Table 3 的比例 2,880 已超過三個數量級，所以「二到三」是保守的說法。 |
| Frontier training in space within 2–3 years "not credible". 10 years "appears optimistic" [101, §V].2–3 年內在太空進行前沿訓練「不可信」。10 年「顯得樂觀」[101, §V]。 | A judgement drawn from the rooflines and the economics, with a scope distinct from a calculated quantity.由 roofline 與經濟分析得出的判斷，其性質為作者判斷，與計算量各有定義。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| Does Table 3's 2-D torus value follow from its formula, and what does the text's formula give?Table 3 的二維環面數值能由其公式推得？內文公式又得出多少？ | \( 2\sqrt{8000}\times 12.5\ \text{GB/s} \quad\text{vs}\quad \sqrt{8000}\times 12.5\ \text{GB/s} \) | 2.24 vs 1.12 TB/scomputed. Table 3's 2.25 matches its own formula, the text formula halves it計算值。Table 3 的 2.25 與其自身公式相符，內文公式使其減半 |
| Do the 3-D torus and Clos values follow?三維環面與 Clos 的數值與公式相符嗎？ | \( 2\cdot 8000^{2/3}\times 12.5\ \text{GB/s},\qquad \tfrac12 \cdot 8000 \cdot 72 \cdot 100\ \text{GB/s} \) | 10 TB/s · 28,800 TB/scomputed. Inputs from Table 3 and §IV-A計算值。輸入來自 Table 3 與 §IV-A |
| Do the latencies follow?延遲數值與公式相符嗎？ | \( \log_{10}8000 \times 1\ \mu\text{s},\quad \sqrt{8000}\times 20\ \mu\text{s},\quad \tfrac32\, 8000^{1/3}\times 20\ \mu\text{s} \) | ≈4 · ≈1,790 · 600 µscomputed. The 4 µs implies a base-10 logarithm, which the paper leaves unspecified計算值。4 µs 意味以 10 為底的對數，論文說明資料待補 |
| Where does 500 kFLOP/B come from?500 kFLOP/B 從何而來？ | \( I_{\text{bisect}} = \dfrac{k}{c}\,B_{\text{tokens}} = \dfrac{6}{12}\times 10^{6} \) | 5 × 10⁵ FLOP/Bcomputed from Eqs. 2–3由 Eqs. 2–3 計算 |
| What would the tori reach with 10 Tb/s Suncatcher-class links?換成 10 Tb/s 的 Suncatcher 等級鏈路，環面能達到多少？ | \( 2\sqrt{8000}\times 1.25\ \text{TB/s},\qquad 2\cdot 8000^{2/3}\times 1.25\ \text{TB/s} \) | ≈224 · 1,000 TB/scomputed with Table 3 formulas. Still 129× and 29× below the Clos以 Table 3 公式計算。仍比 Clos 低 129 倍與 29 倍 |
| How large a cluster would 8,000 satellites at 100 m spacing need, by Pénot's fits?依 Pénot 的擬合，8,000 顆、間距 100 m 的衛星需要多大的叢集？ | \( x = (8000/3.63)^{1/2},\qquad x = (8000/0.27)^{1/2.99} \) | R_max ≈ 4.7 km (planar) · ≈ 3.1 km (3-D)computed by extrapolating Pénot & Balakrishnan Table 1 beyond its sweep, R_min = 100 m由 Pénot 與 Balakrishnan Table 1 外推到其掃描範圍之外計算，R_min = 100 m |
| What does the 1 GW launch bill come to?1 GW 的發射費用是多少？ | \( 10^{7}\ \text{kg} \times 200\ \$/\text{kg} \) | ≈ $2 Bcomputed. Reproduces §III-A計算值。重現 §III-A |
6. Limits and open questions6. 限制與待解的問題
- The 2-D torus formula is inconsistent. The text gives 2-D bisection as N^{1/2}·B_l. Table 3 gives 2·N^{1/2}·B_l, a factor of two higher (factbase Part C, item 41). The table's 2.25 TB/s follows from its own formula. Thus, this primer uses Table 3 and notes the factor.二維環面公式差異。內文把二維環面對分頻寬寫成 N^{1/2}·B_l，而 Table 3 使用 2·N^{1/2}·B_l，相差兩倍（事實庫 Part C 第 41 項）。表中的 2.25 TB/s 由其自身公式推得，所以本教材採用 Table 3 並註明這個倍數。
- The node is a rack on one side and a satellite on the other. The 12,880× gap compares 8,000 racks with 72 links each against 8,000 satellites with 4–6. Thus, most of the ratio is the port count. How many accelerators a satellite holds, and therefore whether 8,000 satellites equal 8,000 racks, remains unestablished.一邊的節點是機架，另一邊是衛星。12,880 倍的差距比較的是每個 72 條鏈路的 8,000 個機架，與每顆 4–6 條鏈路的 8,000 顆衛星，所以比例大多來自埠數。一顆衛星容納多少加速器、因而 8,000 顆衛星與 8,000 個機架的對應關係，仍待論證。
- The per-hop delay is an assumption. Per-hop node delay still requires satellite-network measurements (factbase Part C, item 19). The 20 µs assumption drives the 1,800 and 600 µs results. Light-time across a 2 km cluster is only about 6.7 µs.每跳延遲是假設。衛星網路的每跳節點延遲仍待公開量測證據（事實庫 Part C 第 19 項），所以驅動 1,800 與 600 µs 數字的 20 µs 是建模選擇，而光跨越 2 km 叢集只需約 6.7 µs。
- Orbital geometry and a Clos alternative remain open. Flight feasibility remains open for 89 × 90 and 20 × 20 × 20 tori. Stable wrap-around line of sight is "particularly challenging". Its bi-planar "Clos satellite constellation" is proposed with quantification still pending, whereas Pénot and Balakrishnan show a Clos is topologically feasible at L ≥ 3.軌道幾何與 Clos 替代方案仍待分析。本文仍待檢查 89 × 90 或 20 × 20 × 20 的環面維持穩定視線飛行的可行性，包括它自己稱為「尤其具挑戰性」的環繞鏈路。而它提出的雙平面「Clos 衛星星座」量化結果亦仍待補充，Pénot 與 Balakrishnan 則證明 Clos 在 L ≥ 3 時拓樸上可行。
- Only data-parallel traffic crosses the bisection. Tensor- and pipeline-parallel traffic is assumed to stay local. Terrestrial tensor parallelism stays within a host. Applying this to orbit requires evidence that a satellite carries enough accelerators.只有資料平行流量跨越對分。假設張量平行與管線平行流量留在本地，這是地面上把張量平行留在主機內的做法，但仍須證據確認一顆衛星容納的加速器足以讓這個假設成立。
- Outdated or unstated link assumptions. Four to six ISLs exceed the 3 terminals on current Starlink satellites. The 100 Gb/s laser rate is a company claim relayed by journalists. The 2–30 s acquisition delay of modern terminals and the re-pointing it implies for 32,000–48,000 beams are outside the model. The 15-second Starlink schedule lies outside the model. And failure rates remain unspecified, although Chapter 10 contrasts terrestrial clusters' 0.057% monthly link failures.過時或說明資料待補的鏈路假設。每顆衛星 4–6 條 ISL 超過目前 Starlink 衛星的 3 具終端。每具雷射 100 Gb/s 是記者轉述的公司宣稱 company。現代終端 2–30 秒的擷取延遲，以及它對 32,000–48,000 道光束重新指向的含意皆位於模型範圍之外。Starlink 的 15 秒排程位於模型範圍之外。也失效率仍待補充，雖然第 10 章對比了地面叢集每月 0.057% 的鏈路失效。
7. How it connects7. 與其他內容的關聯
- Section 10.4 and Panel 10.1 compare 28,800 : 2.25 : 10 TB/s bandwidth and 4 : 1,800 : 600 µs latency. They use 500 kFLOP/B bisection intensity.第 10.4 節：28,800 : 2.25 : 10 TB/s 與 4 : 1,800 : 600 µs 的比較以及 500 kFLOP/B 的強度，是該節與面板 10.1 的核心。
- Section 11.5: torus versus Clos next to Pénot's packing. Section 11.6: what from the terrestrial stack survives once bisection is this scarce. Section 11.3: the 10 Tb/s link drawn as the dotted roofline.第 11.5 節：環面對 Clos，與 Pénot 的堆積並列。第 11.6 節：對分頻寬如此稀缺時，地面堆疊中有哪些能保留。第 11.3 節：被畫成虛線 roofline 的 10 Tb/s 鏈路。
- Section 12.1: why inference confined to one satellite escapes the fabric problem entirely.第 12.1 節：為什麼限於單一衛星的推論完全避開了架構問題。
- Section 13.2 considers collectives on predictable fabrics. Section 13.3 compares bisection per launched kilogram under real geometry and terminal counts. Both use this paper's values and its open geometry question.第 13.2 節（善用可預測架構的集合通訊）與第 13.3 節（在真實終端數與幾何下每發射公斤的對分頻寬）都從本文的數字與它缺少的幾何出發。
- Suncatcher [36] is the reference cluster. Pénot and Balakrishnan [102] supply the geometry and Clos feasibility this paper lacks. SpaceMoE is the multi-satellite inference case outside the single-satellite roofline's scope.Suncatcher[36]是參考叢集。Pénot 與 Balakrishnan[102]提供本文缺少的幾何與 Clos 可行性。SpaceMoE 是單一衛星 roofline 涵蓋低於的多衛星推論案例。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| N = 8,000 | Nodes compared: 2-D torus 89 × 90, 3-D torus 20 × 20 × 20, Clos of 8,000 racks [101, §IV-A] | 比較的節點數：二維環面 89 × 90、三維環面 20 × 20 × 20、8,000 個機架的 Clos[101, §IV-A] | |
| 100 Gb/s · 4–6 ISLs | Laser link rate (12.5 GB/s in Table 3) and links per satellite [101, §II-D, §IV-A] | 雷射鏈路速率[101, Tab. 3]與每顆衛星鏈路數[101, §II-D, §IV-A] | |
| W = 72 · 100 GB/s · 2:1 | Terrestrial Clos ToR width, link rate, oversubscription [101, Tab. 3] | 地面 Clos 的機架頂寬度、鏈路速率、超額訂閱[101, Tab. 3] | |
| 28,800 : 2.25 : 10 TB/s | Bisection, Clos : 2-D torus : 3-D torus. Ratios 1 : 12,880 : 2,880 [101, Tab. 3] | 對分頻寬，Clos : 二維環面 : 三維環面。比例 1 : 12,880 : 2,880[101, Tab. 3] | |
| 4 : 1,800 : 600 µs | Latency with D = 1 µs (Clos) and 20 µs (tori). Ratios 1 : 450 : 150 [101, Tab. 3] | 延遲，D = 1 µs（Clos）與 20 µs（環面）。比例 1 : 450 : 150[101, Tab. 3] | |
| ≈500 kFLOP/B | Bisection intensity of data-parallel training [101, Eqs. 2–3] | 資料平行訓練的對分強度[101, Eqs. 2–3] | |
| 32,000–48,000 | Beams to steer in an 8,000-node torus [101, §IV-A] | 8,000 節點環面需操控的光束數[101, §IV-A] | |
| "100+"× | Training-cost increase on a torus [101, §IV-E, §V] | 環面上訓練成本的增幅[101, §IV-E, §V] | |
| $200/kg · 10,000 t · ≈$2 B | Starship launch price, 1 GW constellation mass, launch bill [101, §III-A] | Starship 發射價格、1 GW 星座質量、發射費用[101, §III-A] | |
| 2–3 yr / 10 yr | Frontier training "not credible" / "appears optimistic" [101, §V] | 前沿訓練「不可信」／「顯得樂觀」[101, §V] |