[{"data":1,"prerenderedAt":2662},["ShallowReactive",2],{"site-info":3,"post-shared-trunk-four-modalities":99,"related-zh-TW-shared-trunk-four-modalities":1470},{"id":4,"company":5,"extension":16,"meta":17,"services":18,"social":96,"stem":97,"__hash__":98},"site\u002Fsite.yml",{"name":6,"nameEn":7,"email":8,"address":9,"addressEn":10,"lineUrl":11,"lineId":12,"duns":13,"youtube":14,"youtubeId":15},"銡策數位有限公司","Jecision Digital Co., Ltd.","contact@jecision.com","231 新北市新店區北新路二段 262 號 3 樓之 3","3F.-3, No. 262, Sec. 2, Beixin Rd., Xindian Dist., New Taipei City 231, Taiwan","https:\u002F\u002Fline.me\u002FR\u002Fti\u002Fp\u002F@374qnumi","@374qnumi","963905705","https:\u002F\u002Fwww.youtube.com\u002Fshorts\u002F6UE0WTmyjsg","6UE0WTmyjsg","yml",{},[19,43,69],{"key":20,"zh":21,"en":32},"audit",{"title":22,"tagline":23,"summary":24,"points":25,"deliverable":29,"detail":30,"duration":31},"企業 AI 導入健檢","先看清楚，再花錢","60 分鐘線上訪談，盤點最耗時、最依賴人力的營運環節；我們會告訴你哪些工作現在就能交給 AI、哪些還不成熟。",[26,27,28],"現況盤點：梳理日常營運中最耗時、最依賴人力的環節","痛點診斷：判斷哪些工作現在就能交給 AI，哪些建議先保留人力","可行性評估：依你的業務特性與預算，篩選適合的模型與工具","《企業 AI 導入建議報告書》（PDF）","含 1–2 個「開發成本最低、成效最明顯」的優先項目、技術棧建議、預期節省的人力時間與下一步","訪談 60 分鐘，報告書 3 個工作天內交付",{"title":33,"tagline":34,"summary":35,"points":36,"deliverable":40,"detail":41,"duration":42},"AI Readiness Audit","Look before you spend","A 60-minute remote audit of the workflows that eat the most time and people-power — and an honest answer on what AI can take over today, and what it cannot.",[37,38,39],"Audit the daily workflows that consume the most time and headcount","Triage which tasks AI can own today, and which should stay human","Shortlist the models and tools that fit your budget and constraints","AI Adoption Recommendation Report (PDF)","One or two highest-ROI entry points, a recommended tech stack, expected hours saved, and the concrete next step","60-minute call, report delivered within 3 business days",{"key":44,"zh":45,"en":56},"prompt",{"title":46,"tagline":47,"summary":48,"points":49,"deliverable":53,"detail":54,"duration":55},"提示詞工程","讓 AI 穩定產出，而不是碰運氣","針對你的實際任務設計提示詞與多步驟工作流，並做極限測試，確保換了資料、換了人操作也不會走鐘。",[50,51,52],"客製提示詞：ChatGPT、Claude、Gemini、Midjourney 等主流工具的專屬模板","除錯與優化：修掉幻覺、格式跑掉、時好時壞的問題","工作流建置：把複雜任務拆成多步驟提示詞與可重複的 SOP","提示詞文件或設定檔","附使用說明與範例，並提供 1 對 1 教學，確保你與團隊能自己維護","依任務複雜度報價，通常 3–7 個工作天",{"title":57,"tagline":58,"summary":59,"points":60,"deliverable":66,"detail":67,"duration":68},"Prompt Engineering","Make AI produce reliably, not by luck","Prompt sets and multi-step workflows designed for your actual tasks, stress-tested so they keep working when the data or the operator changes.",[61,62,64],"Custom prompts for ChatGPT, Claude, Gemini, Midjourney and friends",{"Debugging and tuning":63},"kill hallucinations, broken formats and flaky output",{"Workflow design":65},"break complex work into repeatable multi-step SOPs","Prompt library or exportable config","Ships with usage notes and examples, plus a 1:1 walkthrough so your team can maintain it","Quoted by complexity, typically 3–7 business days",{"key":70,"zh":71,"en":82},"poc",{"title":72,"tagline":73,"summary":74,"points":75,"deliverable":79,"detail":80,"duration":81},"極速 PoC","五天，看到能跑的東西","不談架構藍圖，直接做一個能動的雛形。五個工作天內交付線上可操作的智能體，用你真實的資料驗證。",[76,77,78],"Day 1：15 分鐘對焦，鎖定單一痛點與資料盤點","Day 2–4：模型串接、邏輯設計與雛形開發","Day 5：Demo 展示與 AI 擴展建議書","可線上操作的智能體雛形","交付範圍為概念驗證與展示，不含後續維運或大規模伺服器部署；要進階成完整系統會另行報價","5 個工作天",{"title":83,"tagline":84,"summary":85,"points":86,"deliverable":93,"detail":94,"duration":95},"Rapid PoC","Five days to something that runs","No architecture decks — a working prototype built in five business days, validated against your real data.",[87,89,91],{"Day 1":88},"15-minute scoping call, one pain point, data inventory",{"Day 2–4":90},"model integration, logic design, prototype build",{"Day 5":92},"live demo plus an AI expansion recommendation","A runnable online agent prototype","Proof of concept and demo scope only; ongoing operations and large-scale deployment are quoted separately","5 business days",[],"site","XTwARfVAOcrXNiKehsM6V6cdkd2HXy4d5yZVrSWrI3Q",{"zh":100,"en":824},{"id":101,"title":102,"author":103,"body":104,"cover":805,"coverAlt":806,"date":807,"description":808,"draft":809,"extension":810,"featured":809,"lang":811,"meta":812,"navigation":225,"ogImage":813,"path":814,"seo":815,"slug":816,"stem":817,"tags":818,"translationKey":816,"updated":822,"__hash__":823},"blog\u002Fblog\u002Fzh-TW\u002Fshared-trunk-four-modalities.md","共用維度：升維的新方向","Jecision",{"type":105,"value":106,"toc":785},"minimark",[107,114,118,121,124,127,143,149,152,155,161,167,173,179,183,189,200,203,206,211,214,231,234,237,242,246,249,254,257,264,270,274,277,283,286,290,293,300,303,306,317,320,323,329,332,335,339,342,345,351,358,361,364,370,373,377,380,383,386,392,396,399,465,468,634,637,640,646,652,658,664,667,670,673,679,685,691,701,712,718,722,725,731,734,747,753,759,762],[108,109,110],"blockquote",{},[111,112,113],"p",{},"這篇沒有我們的實測。數字都來自論文與模型卡，連結在文末；判斷是我們的，會標明。",[115,116,117],"h2",{"id":117},"先講一個場景",[111,119,120],{},"假設你要做一個「公司內部什麼都能搜」的系統。",[111,122,123],{},"合約是 PDF，發票是掃描件，會議是錄音，產品照是圖檔。",[111,125,126],{},"多模態過去有兩條路：一條是養四個模型——文字嵌入、視覺嵌入、語音轉文字、音訊分類，四套向量空間互不相通，最後再加一層對齊；另一條是把一個模型做得很大，大到能吞下所有模態。",[111,128,129,130,137,138,142],{},"2026 年 10 月，Google 的 ",[131,132,136],"a",{"href":133,"rel":134},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.37364",[135],"nofollow","EmbeddingGemma 2"," 示範了第三條：",[139,140,141],"strong",{},"共用語意 backbone，把不同模態的容量隔離在分支裡","。純文字跑起來只要 271M；需要視覺再打開視覺分支，需要聲音再打開音訊分支。",[111,144,145,146],{},"它真正值錢的地方不是「一個模型吃四種資料」，而是：",[139,147,148],{},"升維不必讓整個模型一起變胖。",[115,150,151],{"id":151},"升維的舊帳單",[111,153,154],{},"要理解這條新路，先看舊路的成本。",[111,156,157,160],{},[139,158,159],{},"第一是乘法。"," 稠密模型的帳單是參數量乘上 token 數。想多一種能力，就得把所有東西一起放大。",[111,162,163,166],{},[139,164,165],{},"第二是梯度打架。"," 同一組權重要同時滿足「這個像素是貓」和「這個詞是主詞」。兩邊梯度方向不同，互相拉扯。多任務學習裡叫負遷移。",[111,168,169,172],{},[139,170,171],{},"第三是輸入幾何不同。"," 文字是離散的 token、影像是像素網格、聲音是波形。它們連「什麼是一個單位」都不一樣。",[111,174,175,176],{},"這三個帳單不是要證明多模態不可行，而是說明為什麼舊路走到某個尺寸就開始不划算。",[139,177,178],{},"新方向的價值，就是換一種付帳方式。",[115,180,182],{"id":181},"新方向共用維度隔離容量","新方向：共用維度，隔離容量",[111,184,185,186],{},"EmbeddingGemma 2 的做法是：",[139,187,188],{},"拆在入口，合在語意。",[190,191,196],"pre",{"className":192,"code":194,"language":195},[193],"language-text","圖片\u002F影片\u002F聲音 ──[各自的編碼器]──> soft token ──[同一個 24 層 transformer]──> 向量\n                   ↑ 分開                              ↑ 共用\n","text",[197,198,194],"code",{"__ignoreMap":199},"",[111,201,202],{},"模態編碼器各自處理低階特徵（像素、波形），輸出變成「soft token」——連續向量，跟文字 token 一樣可以排進序列，但不需要對應任何一個字。高階語意則全部交給同一個 transformer。",[111,204,205],{},"參數這樣分：主幹（文字也走這裡）271M、視覺編碼器 167M、音訊編碼器 305M。",[111,207,208],{},[139,209,210],{},"注意那個 305M：音訊編碼器比整個文字主幹還大。",[111,212,213],{},"這正是新方向的關鍵。容量掛在分支上，於是它可以被關掉：",[215,216,219,220,219,227],"figure",{"className":217},[218],"article-figure","\n  ",[221,222],"video",{"src":223,"poster":224,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":226},"\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fsplit.mp4","\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fsplit-poster.jpg",true,"參數分配與實際載入記憶體",[228,229,230],"figcaption",{},"左邊是參數分配；右邊是實際載入。分支可以按需打開——這就是「升維不必變胖」的機制。",[111,232,233],{},"在 Pixel 11 Pro 上、Q4_C 量化下，峰值記憶體（含 runtime 編譯開銷）：純文字 191 MB、文字加影像 447 MB、全模態 567 MB。",[111,235,236],{},"而且文字品質沒有犧牲。它的 271M 文字路徑，多語 MTEB 平均 61.36；前一代 308M 是 61.15。參數更少、還多扛三種模態，文字沒有掉。",[111,238,239],{},[139,240,241],{},"這就是這條路最有希望的地方：升維可以是「選擇性打開」，不是「整體膨脹」。",[115,243,245],{"id":244},"兩種拆法兩張帳單","兩種拆法，兩張帳單",[111,247,248],{},"同樣是「共用主幹、隔離容量」，可以拆在不同地方。目前兩種主流答案：",[250,251,253],"h3",{"id":252},"moe拆在模型中間","MoE：拆在模型中間",[111,255,256],{},"同一層裡放很多專家，每個 token 只走其中幾個，由 router 決定。",[111,258,259,260,263],{},"它換到的是：",[139,261,262],{},"容量與「每 token 計算量」脫鉤。"," 參數可以很多，但每個 token 只碰一小塊。",[111,265,266,267],{},"代價是訓練變難：router 崩塌、專家負載不均、梯度稀疏。",[139,268,269],{},"MoE 用訓練的難度，換推論的成本。",[250,271,273],{"id":272},"embeddinggemma-2拆在入口","EmbeddingGemma 2：拆在入口",[111,275,276],{},"異質的模態交給設計好的編碼器，同質的語意交給共用主幹。",[111,278,259,279,282],{},[139,280,281],{},"容量與「常駐記憶體」脫鉤。"," 你不需要為了偶爾用一次影片檢索，讓整個模型常駐。",[111,284,285],{},"兩者解的不是同一道題，也不互斥。一個在計算上省，一個在記憶體上省。",[115,287,289],{"id":288},"早融合與晚融合在哪一層開始對齊","早融合與晚融合：在哪一層開始對齊",[111,291,292],{},"如果你聽過 CLIP，會覺得「這不就是多模態嗎」。",[111,294,295,296,299],{},"差別在",[139,297,298],{},"合併點","。",[111,301,302],{},"晚融合（CLIP、SigLIP）：兩個完整編碼器各自跑完，最後用對比損失對齊。兩塔之間沒有資訊交換。",[111,304,305],{},"早融合（EmbeddingGemma 2）：編碼完就變成 soft token，塞進同一個 transformer，所有模態在上面混在一起。",[215,307,219,309,219,314],{"className":308},[218],[221,310],{"src":311,"poster":312,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":313},"\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fflow.mp4","\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fflow-poster.jpg","晚融合與早融合的資料流",[228,315,316],{},"差別不在「有沒有對齊」，在「在哪一層開始對齊」——這決定能不能學到跨模態的交互。",[111,318,319],{},"晚融合的好處是兩個塔可以各自很強；代價是學不到跨模態的交互——圖片裡的文字、影片的原聲、簡報裡圖表跟標題的關係。",[111,321,322],{},"早融合的代價是所有模態要擠進同一個狹窄通道。它的模型維度只有 512，總共 24 層。（這是我們的解讀：刻意選窄。）",[111,324,325,326],{},"它的賭注是：",[139,327,328],{},"語意空間夠窄，才逼得出真正的共同表示。",[115,330,331],{"id":331},"三個支持這條路的證據",[111,333,334],{},"論文做了消融實驗，其中三個對「共用主幹」這個方向特別有說服力。",[250,336,338],{"id":337},"拿掉文字專屬的機制全部變好","拿掉文字專屬的機制，全部變好",[111,340,341],{},"Gemma 的 per-layer embeddings 原本多一個查找表，作用是把 token 資訊注入中間層。EmbeddingGemma 2 把查找表拿掉，只保留投影。",[111,343,344],{},"整體從 55.07 變成 55.46；影像 54.57 → 54.67、影片 47.13 → 47.82、視覺文件 61.76 → 62.37。這些消融是訓練過程中的中間配方，不等於最終模型（最終 MMEB-v2 整體是 59.01）。",[111,346,347,350],{},[139,348,349],{},"拿掉反而全部變好","，連純文字也是。論文給的理由是：查找表對非文字模態「沒有直接訊號」，還額外佔記憶體。",[111,352,353,354,357],{},"我們的解讀是：",[139,355,356],{},"共同空間不是靠加東西達成的，是靠拿掉東西。"," 任何讓某一模態「繞過」共同空間的機制，都會拉低整體。",[250,359,360],{"id":360},"視覺編碼器不能凍結",[111,362,363],{},"可訓練的整體 52.27、影像 50.10、影片 43.26；凍結後掉到 48.18、44.35、39.24，差 4.1 分。",[111,365,366,367],{},"很多多模態模型會凍結模態編碼器，省成本又保留原本能力。這篇的結果是：",[139,368,369],{},"編碼器要被訓練去適應共同空間。",[111,371,372],{},"對這條路來說這是好消息也是提醒——共用主幹不是免費的，它要求每個分支往同一個地方收斂，而收斂是可以被訓練出來的。",[250,374,376],{"id":375},"音訊編碼器是凍結的而音訊剛好只打平","音訊編碼器是凍結的，而音訊剛好只打平",[111,378,379],{},"論文寫得很清楚：音訊編碼器跟 Gemma 4 E2B\u002FE4B 完全相同，訓練期間凍結。",[111,381,382],{},"而音訊是它跟對手最接近的一項：MAEB 的 Mean(Task) 49.87，Jina Omni Nano 50.14，差 0.27 分。Mean(Type) 則是 48.65 對 55.31，沒有打平。對照之下，影像 57.28 對 55.43、影片 50.67 對 35.87，都是明顯領先。",[111,384,385],{},"作者自己也留了一句：「透過進一步訓練音訊編碼器，可能也會有類似的增益。」",[111,387,388,391],{},[139,389,390],{},"這是我們的推測，不是論文的結論","：把音訊解凍，這一項可能拉開。但他們沒做，所以只能說「有這個可能」——這正是這條路還沒走完的地方。",[115,393,395],{"id":394},"怎麼用選型不是排名","怎麼用：選型不是排名",[111,397,398],{},"選型要看的是「你的場景需要什麼」，不是「誰第一名」。",[400,401,402,418],"table",{},[403,404,405],"thead",{},[406,407,408,412,415],"tr",{},[409,410,411],"th",{},"你的用途",[409,413,414],{},"可以這樣選",[409,416,417],{},"理由",[419,420,421,433,444,454],"tbody",{},[406,422,423,427,430],{},[424,425,426],"td",{},"純文字檢索／RAG",[424,428,429],{},"Qwen3 或 Jina 的純文字模型",[424,431,432],{},"同尺寸下分數更高、模型更小",[406,434,435,438,441],{},[424,436,437],{},"文字加圖片搜尋",[424,439,440],{},"EmbeddingGemma 2 或 SigLIP 類專門模型",[424,442,443],{},"取決於要不要一份 checkpoint 打全部",[406,445,446,449,451],{},[424,447,448],{},"影片、聲音、混合模態",[424,450,136],{},[424,452,453],{},"影片領先 15–20 分，原生音訊比先轉文字好約 12 分",[406,455,456,459,462],{},[424,457,458],{},"手機／離線、記憶體吃緊",[424,460,461],{},"EmbeddingGemma 2 純文字版",[424,463,464],{},"191 MB，而且只載入需要的編碼器",[111,466,467],{},"幾個值得一起看的數字：",[400,469,470,498],{},[403,471,472],{},[406,473,474,477,480,483,486,489,492,495],{},[409,475,476],{},"模型",[409,478,479],{},"參數",[409,481,482],{},"多語文字",[409,484,485],{},"英文文字",[409,487,488],{},"程式碼",[409,490,491],{},"多模態整體",[409,493,494],{},"影像",[409,496,497],{},"影片",[419,499,500,535,559,588,611],{},[406,501,502,506,509,512,515,520,525,530],{},[424,503,504],{},[139,505,136],{},[424,507,508],{},"744M",[424,510,511],{},"61.36",[424,513,514],{},"68.46",[424,516,517],{},[139,518,519],{},"78.68",[424,521,522],{},[139,523,524],{},"59.01",[424,526,527],{},[139,528,529],{},"57.28",[424,531,532],{},[139,533,534],{},"50.67",[406,536,537,540,543,546,549,552,555,557],{},[424,538,539],{},"EmbeddingGemma 1",[424,541,542],{},"308M",[424,544,545],{},"61.15",[424,547,548],{},"69.67",[424,550,551],{},"68.76",[424,553,554],{},"—",[424,556,554],{},[424,558,554],{},[406,560,561,564,567,572,577,579,582,585],{},[424,562,563],{},"Jina Embed v5 Omni Nano",[424,565,566],{},"986M",[424,568,569],{},[139,570,571],{},"65.52",[424,573,574],{},[139,575,576],{},"71.11",[424,578,554],{},[424,580,581],{},"43.61",[424,583,584],{},"31.56",[424,586,587],{},"30.73",[406,589,590,593,596,599,602,605,607,609],{},[424,591,592],{},"Qwen3-Embedding-0.6B",[424,594,595],{},"596M",[424,597,598],{},"64.34",[424,600,601],{},"70.47",[424,603,604],{},"75.42",[424,606,554],{},[424,608,554],{},[424,610,554],{},[406,612,613,616,619,621,623,625,628,631],{},[424,614,615],{},"InterestFM UIR CAFe 0.5B",[424,617,618],{},"894M",[424,620,554],{},[424,622,554],{},[424,624,554],{},[424,626,627],{},"49.33",[424,629,630],{},"55.43",[424,632,633],{},"35.87",[111,635,636],{},"（— 代表論文該欄沒有分數；論文用 ∗ 標「不支援」、– 標「沒有分數」，兩者不是同一件事。數字皆來自論文 Table 6。）",[111,638,639],{},"三件事值得注意。",[111,641,642,645],{},[139,643,644],{},"第一，跟前代比，文字幾乎沒變，但多出了三種模態。"," 真正的差別不是分數，是「多出來的功能沒有把原本的能力吃掉」。",[111,647,648,651],{},[139,649,650],{},"第二，純文字它輸給專門模型。"," Jina 多語贏 4.2 分、英文贏 2.7 分；Qwen3 贏 3.0 分。而且 744M 是「全部載入」的數字，純文字實際只跑 271M。論文附錄也給了答案：268M 的 harrier-oss-v1-270m 多語 66.5、212M 的 jina-embed-v5-text-nano 65.5，都比它高。這兩個是純文字專家模型。",[111,653,654,655],{},"但這不是缺點，是分工。",[139,656,657],{},"你要一份能升維的 checkpoint，跟你要一個最強的文字模型，是兩個不同的需求。",[111,659,660,663],{},[139,661,662],{},"第三，多模態它領先很多。"," 影片 50.67，Jina 30.73、InterestFM 35.87，差距 15 到 20 分。影像 57.28 對 55.43，贏得不誇張但穩定。",[111,665,666],{},"聲音方面：同一個模型，原生音訊 69.54，先轉文字（ASR）只有 57.59——差約 12 分。聲音裡的非語言資訊，轉文字就沒了。（前代的 ASR 是 50.95；拿它跟這代的 69.54 比會得到 18.6 分，但那是跨世代比較，不是模態差距。）",[115,668,669],{"id":669},"還沒走完的地方",[111,671,672],{},"一條新路值不值得押注，要看它還缺什麼。這裡有四件事。",[111,674,675,678],{},[139,676,677],{},"純文字不是最強。"," 論文說的 Pareto frontier 是「品質與參數量」的效率前沿（贏過兩到十二倍大的模型），不是每個單項都第一。",[111,680,681,684],{},[139,682,683],{},"視覺文件還有空間。"," 在 280 token 預算下 VisDoc 67.84，Jina Omni Nano 71.34。掃描件、表格、排版複雜的文件靠細粒度視覺細節，而論文也說提高 token 預算對 VisDoc 還有增益——這比較像「還沒調到最好」，不是天花板。",[111,686,687,690],{},[139,688,689],{},"量化對聲音傷害最大。"," bfloat16 換到 int4，文字幾乎無感（61.36 → 60.84），聲音掉 5.6 分（69.54 → 63.92）。不過論文註明 float 與量化跑的是不同評測 harness，這個降幅含 harness 差異，不能全算在量化頭上。",[111,692,693,696,697,700],{},[139,694,695],{},"同一個 token 預算，對三種任務方向不同。"," 影像是每幀 280、影片是每幀 140 訓練出來的；把推論預算從 280 提到 560：視覺文件大幅改善（63.38 → 67.62）、影像飽和、",[139,698,699],{},"影片反而變差","（48.51 → 46.52）。論文說原因是過長的序列稀釋了注意力與 mean-pooled 表示。",[215,702,219,704,219,709],{"className":703},[218],[221,705],{"src":706,"poster":707,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":708},"\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fbudget.mp4","\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities\u002Fbudget-poster.jpg","視覺 token 預算對影像、影片、視覺文件的影響",[228,710,711],{},"同一個旋鈕，三種方向。影像的訓練值是 280、影片是 140；影片在 280 之後還在進步，掉的是 280 → 560。",[111,713,714,715],{},"所以部署時，這個旋鈕要按任務分開調——",[139,716,717],{},"這不是缺陷，是這條路給你的控制權。",[115,719,721],{"id":720},"結論升維的新方向","結論：升維的新方向",[111,723,724],{},"回到那個場景。你不再需要在「四個模型」與「一個大模型」之間二選一。",[111,726,727,730],{},[139,728,729],{},"共用維度，容量隔離在分支","：同一個語意空間，需要什麼模態就打開什麼分支。",[111,732,733],{},"兩種拆法，各換一張帳單：",[108,735,736,742],{},[111,737,738,741],{},[139,739,740],{},"MoE"," 拆在模型中間。同質的專家，用學出來的 router 決定誰上班。換到「容量 vs 每 token 計算」的脫鉤。",[111,743,744,746],{},[139,745,136],{}," 拆在入口。異質的模態，用設計好的編碼器決定誰處理。換到「容量 vs 常駐記憶體」的脫鉤。",[111,748,749,750],{},"一個共同的教訓：",[139,751,752],{},"拆不是為了變大，是為了讓某些東西可以被關掉。",[111,754,755,756],{},"而它最反直覺、也最值得記住的一課是——",[139,757,758],{},"共同空間是拿掉專屬機制換來的，不是加更多模組換來的。",[760,761],"hr",{},[111,763,764],{},[765,766,767,768,772,773,778,779,784],"em",{},"我們沒有實測這顆模型的部署難度、工具鏈支援或實際延遲，所以這篇不建議你直接遷移。要做的話，先量三件事：你的文字主幹到底多大、分支能不能關、跨模態查詢佔比多少。原始資料：",[131,769,771],{"href":133,"rel":770},[135],"論文 arXiv 2610.37364","、",[131,774,777],{"href":775,"rel":776},"https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fdocs",[135],"Gemma 文件","。MoE 的相關工作見 ",[131,780,783],{"href":781,"rel":782},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.37386",[135],"arXiv 2610.37386","（channel selection，不是 MoE 綜述）。",{"title":199,"searchDepth":786,"depth":786,"links":787},2,[788,789,790,791,796,797,802,803,804],{"id":117,"depth":786,"text":117},{"id":151,"depth":786,"text":151},{"id":181,"depth":786,"text":182},{"id":244,"depth":786,"text":245,"children":792},[793,795],{"id":252,"depth":794,"text":253},3,{"id":272,"depth":794,"text":273},{"id":288,"depth":786,"text":289},{"id":331,"depth":786,"text":331,"children":798},[799,800,801],{"id":337,"depth":794,"text":338},{"id":360,"depth":794,"text":360},{"id":375,"depth":794,"text":376},{"id":394,"depth":786,"text":395},{"id":669,"depth":786,"text":669},{"id":720,"depth":786,"text":721},"\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities.svg","參數分配圖：主幹 271M、視覺 167M、音訊 305M，容量隔離在分支，需要時才載入","2026-11-01","多模態過去只有兩條路：養四個模型，或把一個模型做大。EmbeddingGemma 2 示範第三條——共用語意 backbone，把容量隔離在分支。純文字只載入 271M，需要時再打開視覺與音訊。這篇講這條路怎麼走、代價在哪、什麼時候值得用。",false,"md","zh-TW",{},"\u002Fimages\u002Fblog\u002Fshared-trunk-four-modalities.jpg","\u002Fblog\u002Fzh-tw\u002Fshared-trunk-four-modalities",{"title":102,"description":808},"shared-trunk-four-modalities","blog\u002Fzh-TW\u002Fshared-trunk-four-modalities",[819,820,821],"multimodal","embedding","model-design",null,"pJvRGZyuC8339uRXXK3bnnd35lxoy_07t8wNLSWB1Is",{"id":825,"title":826,"author":103,"body":827,"cover":1459,"coverAlt":1460,"date":807,"description":1461,"draft":809,"extension":810,"featured":809,"lang":1462,"meta":1463,"navigation":225,"ogImage":1464,"path":1465,"seo":1466,"slug":816,"stem":1467,"tags":1468,"translationKey":816,"updated":822,"__hash__":1469},"blog\u002Fblog\u002Fen\u002Fshared-trunk-four-modalities.md","Shared dimensions: a new direction for going multimodal",{"type":105,"value":828,"toc":1441},[829,834,838,841,844,847,858,864,868,871,877,883,889,895,899,902,908,911,914,919,922,931,934,937,942,946,949,953,956,963,969,973,976,982,985,989,992,999,1002,1005,1014,1017,1020,1026,1030,1033,1037,1040,1043,1049,1056,1060,1063,1069,1072,1076,1079,1082,1085,1091,1095,1098,1158,1161,1297,1300,1303,1309,1315,1321,1327,1330,1334,1337,1343,1349,1355,1365,1374,1380,1384,1387,1393,1396,1408,1414,1420,1422],[108,830,831],{},[111,832,833],{},"No benchmarks of our own here. The numbers come from the paper and the model card, linked at the end; the judgments are ours and labelled.",[115,835,837],{"id":836},"start-with-a-scene","Start with a scene",[111,839,840],{},"Suppose you want a system that can search everything inside a company.",[111,842,843],{},"Contracts are PDFs. Invoices are scans. Meetings are audio. Product photos are images.",[111,845,846],{},"Multimodal used to offer two paths. One was to run four models — text embeddings, image embeddings, speech-to-text, audio classification — with four vector spaces that do not talk to each other, plus an alignment layer on top. The other was to make one model big enough to swallow every modality.",[111,848,849,850,853,854,857],{},"In October 2026, Google's ",[131,851,136],{"href":133,"rel":852},[135]," demonstrated a third path: ",[139,855,856],{},"share the semantic backbone, and isolate capacity for each modality in branches",". Text-only runs at 271M; turn on vision when you need images, turn on audio when you need sound.",[111,859,860,861],{},"What is actually valuable here is not \"one model handles four data types\". It is that ",[139,862,863],{},"going multimodal does not require making the whole model fatter.",[115,865,867],{"id":866},"the-old-bill-for-scaling-up","The old bill for scaling up",[111,869,870],{},"To see why the third path matters, look at what the old paths cost.",[111,872,873,876],{},[139,874,875],{},"The first is multiplication."," A dense model's bill is parameters times tokens. Want one more capability, and you scale everything up together.",[111,878,879,882],{},[139,880,881],{},"The second is gradient conflict."," The same weights must satisfy \"this pixel is a cat\" and \"this word is a subject\". The gradients pull in different directions — negative transfer, in multitask learning terms.",[111,884,885,888],{},[139,886,887],{},"The third is input geometry."," Text is discrete tokens, images are pixel grids, audio is waveforms. They do not even agree on what one unit is.",[111,890,891,892],{},"These three costs are not an argument that multimodal is impossible. They explain why the old path stops paying off past a certain size. ",[139,893,894],{},"The value of the new direction is a different way to pay.",[115,896,898],{"id":897},"the-new-direction-shared-dimensions-isolated-capacity","The new direction: shared dimensions, isolated capacity",[111,900,901],{},"EmbeddingGemma 2 splits at the entrance and merges in semantics.",[190,903,906],{"className":904,"code":905,"language":195},[193],"image\u002Fvideo\u002Faudio ──[per-modality encoders]──> soft tokens ──[one 24-layer transformer]──> vector\n                        ↑ split                                    ↑ shared\n",[197,907,905],{"__ignoreMap":199},[111,909,910],{},"Each modality encoder handles low-level features (pixels, waveforms) and turns them into \"soft tokens\" — continuous vectors that sit in the sequence like text tokens but need not correspond to any word. High-level meaning goes through one shared transformer.",[111,912,913],{},"The parameters split like this: a 271M trunk (text also runs here), a 167M vision encoder, and a 305M audio encoder.",[111,915,916],{},[139,917,918],{},"Note that 305M: the audio encoder is bigger than the entire text trunk.",[111,920,921],{},"That is the point of the new direction. Capacity lives in branches, so it can be switched off:",[215,923,219,925,219,928],{"className":924},[218],[221,926],{"src":223,"poster":224,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":927},"Parameter split and actual loaded memory",[228,929,930],{},"Parameters on the left, actual load on the right. Branches open on demand — that is how scale stops meaning bloat.",[111,932,933],{},"On a Pixel 11 Pro at Q4_C, peak memory (including runtime compilation overhead): 191 MB for text, 447 MB for text plus images, 567 MB for everything.",[111,935,936],{},"And text quality does not pay for it. Its 271M text path scores 61.36 on multilingual MTEB; the previous generation at 308M scored 61.15. Fewer parameters, three extra modalities, no text regression.",[111,938,939],{},[139,940,941],{},"That is the most promising part: scale can be opened selectively instead of expanded universally.",[115,943,945],{"id":944},"two-ways-to-split-two-bills","Two ways to split, two bills",[111,947,948],{},"\"Share the trunk, isolate capacity\" can be done at different places. Two current answers:",[250,950,952],{"id":951},"moe-split-in-the-middle-of-the-model","MoE: split in the middle of the model",[111,954,955],{},"Many experts sit inside a layer; each token visits a few, chosen by a router.",[111,957,958,959,962],{},"What it buys: ",[139,960,961],{},"capacity decouples from compute per token."," Parameters can be plentiful while each token touches a small slice.",[111,964,965,966],{},"The cost is training difficulty: router collapse, uneven expert load, sparse gradients. ",[139,967,968],{},"MoE trades training difficulty for inference cost.",[250,970,972],{"id":971},"embeddinggemma-2-split-at-the-entrance","EmbeddingGemma 2: split at the entrance",[111,974,975],{},"Heterogeneous modalities go to designed encoders; homogeneous meaning goes to the shared trunk.",[111,977,958,978,981],{},[139,979,980],{},"capacity decouples from resident memory."," You do not keep the whole model resident for the occasional video search.",[111,983,984],{},"These are not the same problem, and they are not mutually exclusive. One saves compute; the other saves memory.",[115,986,988],{"id":987},"early-fusion-late-fusion-where-alignment-starts","Early fusion, late fusion: where alignment starts",[111,990,991],{},"If you know CLIP, this looks like \"multimodal, again\".",[111,993,994,995,998],{},"The difference is ",[139,996,997],{},"where the merge happens",".",[111,1000,1001],{},"Late fusion (CLIP, SigLIP): two full encoders run independently, then a contrastive loss aligns them. The towers never exchange information.",[111,1003,1004],{},"Early fusion (EmbeddingGemma 2): encoders produce soft tokens that enter one transformer, where all modalities mix.",[215,1006,219,1008,219,1011],{"className":1007},[218],[221,1009],{"src":311,"poster":312,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":1010},"Late fusion versus early fusion data flow",[228,1012,1013],{},"The difference is not whether alignment exists — it is which layer starts aligning. That decides whether cross-modal interaction can be learned.",[111,1015,1016],{},"Late fusion lets each tower be strong on its own; the cost is that it cannot learn cross-modal interaction — text inside images, the original audio of a video, the relationship between a chart and its heading.",[111,1018,1019],{},"Early fusion costs something too: every modality must fit through one narrow channel. The model is 512-dimensional and 24 layers deep. (Our reading: deliberately narrow.)",[111,1021,1022,1023],{},"Its bet: ",[139,1024,1025],{},"only a narrow semantic space forces a genuinely shared representation.",[115,1027,1029],{"id":1028},"three-pieces-of-evidence-for-this-direction","Three pieces of evidence for this direction",[111,1031,1032],{},"The paper ran ablations; three of them speak directly to the \"shared trunk\" idea.",[250,1034,1036],{"id":1035},"removing-a-text-only-mechanism-improves-everything","Removing a text-only mechanism improves everything",[111,1038,1039],{},"Gemma's per-layer embeddings used an extra lookup table to inject token information into intermediate layers. EmbeddingGemma 2 drops the lookup table and keeps only the projection.",[111,1041,1042],{},"Overall goes from 55.07 to 55.46; image 54.57 → 54.67, video 47.13 → 47.82, visual documents 61.76 → 62.37. These ablations are intermediate training recipes, not the final model (which reaches 59.01 MMEB-v2 overall).",[111,1044,1045,1048],{},[139,1046,1047],{},"Removing it made everything better",", text included. The paper's explanation: the lookup table carries \"no direct signal\" for non-text modalities, and costs memory.",[111,1050,1051,1052,1055],{},"Our reading: ",[139,1053,1054],{},"a shared space is achieved by removing things, not adding them."," Any mechanism that lets one modality bypass the shared space drags the whole down.",[250,1057,1059],{"id":1058},"the-vision-encoder-cannot-be-frozen","The vision encoder cannot be frozen",[111,1061,1062],{},"Trainable: 52.27 overall, 50.10 image, 43.26 video. Frozen: 48.18, 44.35, 39.24 — a 4.1 point gap.",[111,1064,1065,1066],{},"Many multimodal models freeze modality encoders to save cost and preserve prior ability. This result says otherwise: ",[139,1067,1068],{},"encoders have to be trained to adapt to the shared space.",[111,1070,1071],{},"For this direction that is both good news and a reminder — the shared trunk is not free; it asks every branch to converge somewhere, and that convergence can be trained.",[250,1073,1075],{"id":1074},"the-audio-encoder-is-frozen-and-audio-merely-ties","The audio encoder is frozen, and audio merely ties",[111,1077,1078],{},"The paper is explicit: the audio encoder is identical to Gemma 4 E2B\u002FE4B and frozen during training.",[111,1080,1081],{},"Audio is also where it is closest to its rival: MAEB Mean(Task) 49.87 versus Jina Omni Nano 50.14, a 0.27 gap. Mean(Type) is 48.65 versus 55.31 — not a tie. By contrast, image is 57.28 versus 55.43 and video 50.67 versus 35.87, both clearly ahead.",[111,1083,1084],{},"The authors leave one line: further training of the audio encoder \"may also yield similar gains\".",[111,1086,1087,1090],{},[139,1088,1089],{},"This is our speculation, not the paper's conclusion",": unfreezing audio could open that gap. They did not do it, so we can only say it is possible — and that is exactly the part of this path that is not finished.",[115,1092,1094],{"id":1093},"how-to-use-it-selection-is-not-a-ranking","How to use it: selection is not a ranking",[111,1096,1097],{},"What matters is what your use case needs, not who is first.",[400,1099,1100,1113],{},[403,1101,1102],{},[406,1103,1104,1107,1110],{},[409,1105,1106],{},"Your use case",[409,1108,1109],{},"A reasonable choice",[409,1111,1112],{},"Why",[419,1114,1115,1126,1137,1147],{},[406,1116,1117,1120,1123],{},[424,1118,1119],{},"Text-only retrieval \u002F RAG",[424,1121,1122],{},"Qwen3 or Jina text models",[424,1124,1125],{},"Higher scores at the same size, smaller models",[406,1127,1128,1131,1134],{},[424,1129,1130],{},"Text plus image search",[424,1132,1133],{},"EmbeddingGemma 2 or a SigLIP-style specialist",[424,1135,1136],{},"Depends on whether you want one checkpoint for everything",[406,1138,1139,1142,1144],{},[424,1140,1141],{},"Video, audio, mixed modality",[424,1143,136],{},[424,1145,1146],{},"15–20 points ahead on video; native audio beats ASR by about 12 points",[406,1148,1149,1152,1155],{},[424,1150,1151],{},"Phone \u002F offline, tight memory",[424,1153,1154],{},"EmbeddingGemma 2 text-only",[424,1156,1157],{},"191 MB, and only the encoders you need",[111,1159,1160],{},"Some numbers worth reading together:",[400,1162,1163,1191],{},[403,1164,1165],{},[406,1166,1167,1170,1173,1176,1179,1182,1185,1188],{},[409,1168,1169],{},"Model",[409,1171,1172],{},"Params",[409,1174,1175],{},"Multilingual text",[409,1177,1178],{},"English text",[409,1180,1181],{},"Code",[409,1183,1184],{},"Multimodal overall",[409,1186,1187],{},"Image",[409,1189,1190],{},"Video",[419,1192,1193,1221,1239,1261,1279],{},[406,1194,1195,1199,1201,1203,1205,1209,1213,1217],{},[424,1196,1197],{},[139,1198,136],{},[424,1200,508],{},[424,1202,511],{},[424,1204,514],{},[424,1206,1207],{},[139,1208,519],{},[424,1210,1211],{},[139,1212,524],{},[424,1214,1215],{},[139,1216,529],{},[424,1218,1219],{},[139,1220,534],{},[406,1222,1223,1225,1227,1229,1231,1233,1235,1237],{},[424,1224,539],{},[424,1226,542],{},[424,1228,545],{},[424,1230,548],{},[424,1232,551],{},[424,1234,554],{},[424,1236,554],{},[424,1238,554],{},[406,1240,1241,1243,1245,1249,1253,1255,1257,1259],{},[424,1242,563],{},[424,1244,566],{},[424,1246,1247],{},[139,1248,571],{},[424,1250,1251],{},[139,1252,576],{},[424,1254,554],{},[424,1256,581],{},[424,1258,584],{},[424,1260,587],{},[406,1262,1263,1265,1267,1269,1271,1273,1275,1277],{},[424,1264,592],{},[424,1266,595],{},[424,1268,598],{},[424,1270,601],{},[424,1272,604],{},[424,1274,554],{},[424,1276,554],{},[424,1278,554],{},[406,1280,1281,1283,1285,1287,1289,1291,1293,1295],{},[424,1282,615],{},[424,1284,618],{},[424,1286,554],{},[424,1288,554],{},[424,1290,554],{},[424,1292,627],{},[424,1294,630],{},[424,1296,633],{},[111,1298,1299],{},"(— means the paper reports no score for that cell; the paper uses ∗ for \"not supported\" and – for \"score unavailable\", which are not the same thing. All numbers from Table 6.)",[111,1301,1302],{},"Three things stand out.",[111,1304,1305,1308],{},[139,1306,1307],{},"First, versus the previous generation, text barely moved — but three modalities appeared."," The real difference is not the score; it is that the extra capability did not eat the existing one.",[111,1310,1311,1314],{},[139,1312,1313],{},"Second, text-only it loses to specialists."," Jina leads by 4.2 on multilingual and 2.7 on English; Qwen3 by 3.0. And 744M is the \"everything loaded\" number — text-only runs at 271M. The appendix gives the fair comparison: harrier-oss-v1-270m at 268M scores 66.5 and jina-embed-v5-text-nano at 212M scores 65.5, both higher. Both are text specialists.",[111,1316,1317,1318],{},"That is not a flaw, it is a division of labour. ",[139,1319,1320],{},"Wanting one checkpoint that can grow into other modalities, and wanting the strongest text model, are two different requirements.",[111,1322,1323,1326],{},[139,1324,1325],{},"Third, multimodal it leads clearly."," Video 50.67 versus Jina 30.73 and InterestFM 35.87 — a 15 to 20 point gap. Image 57.28 versus 55.43 is modest but consistent.",[111,1328,1329],{},"On audio: same model, native audio 69.54 versus 57.59 for ASR — about 12 points. What is lost in transcription is the non-linguistic part of sound. (The previous generation's ASR scored 50.95; comparing that to 69.54 gives 18.6, but that is a generational comparison, not a modality gap.)",[115,1331,1333],{"id":1332},"what-is-not-finished","What is not finished",[111,1335,1336],{},"To judge whether a new path is worth betting on, look at what it still lacks. Four things.",[111,1338,1339,1342],{},[139,1340,1341],{},"Text is not the strongest."," The Pareto frontier the paper claims is a quality-versus-parameters efficiency frontier (beating models 2–12× larger), not a first place in every column.",[111,1344,1345,1348],{},[139,1346,1347],{},"Visual documents still have room."," At a 280-token budget, VisDoc is 67.84 versus Jina Omni Nano's 71.34. Scans, tables and dense layouts depend on fine visual detail, and the paper notes that raising the token budget still improves VisDoc — this looks more like \"not tuned yet\" than a ceiling.",[111,1350,1351,1354],{},[139,1352,1353],{},"Quantization hurts audio most."," Moving from bfloat16 to int4, text barely notices (61.36 → 60.84) while audio drops 5.6 points (69.54 → 63.92). The paper notes, however, that float and quantized runs use different evaluation harnesses, so the drop includes harness differences and cannot be blamed on quantization alone.",[111,1356,1357,1360,1361,1364],{},[139,1358,1359],{},"The same token budget moves three tasks in three directions."," Images are trained at 280 tokens per frame, video at 140; raising inference from 280 to 560 improves visual documents sharply (63.38 → 67.62), saturates images, and ",[139,1362,1363],{},"makes video worse"," (48.51 → 46.52). The paper attributes this to excessive sequence length diluting attention and mean-pooled representations.",[215,1366,219,1368,219,1371],{"className":1367},[218],[221,1369],{"src":706,"poster":707,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":1370},"Vision token budget effects on image, video and visual documents",[228,1372,1373],{},"One knob, three directions. Images train at 280, video at 140; video is still improving at 280 and drops from 280 to 560.",[111,1375,1376,1377],{},"So the knob has to be tuned per task — ",[139,1378,1379],{},"not a defect, but the control this direction gives you.",[115,1381,1383],{"id":1382},"conclusion-a-new-direction-for-scaling-up","Conclusion: a new direction for scaling up",[111,1385,1386],{},"Back to the scene. You no longer have to choose between \"four models\" and \"one big model\".",[111,1388,1389,1392],{},[139,1390,1391],{},"Shared dimensions, capacity isolated in branches",": one semantic space, open the branch you need.",[111,1394,1395],{},"Two ways to split, two different bills:",[108,1397,1398,1403],{},[111,1399,1400,1402],{},[139,1401,740],{}," splits inside the model. Homogeneous experts, a learned router decides who works. It buys decoupling of capacity from compute per token.",[111,1404,1405,1407],{},[139,1406,136],{}," splits at the entrance. Heterogeneous modalities, designed encoders decide who handles what. It buys decoupling of capacity from resident memory.",[111,1409,1410,1411],{},"The shared lesson: ",[139,1412,1413],{},"splitting is not about getting bigger; it is about making some parts switchable off.",[111,1415,1416,1417],{},"And the most counter-intuitive lesson of all — ",[139,1418,1419],{},"a shared space comes from removing specialised machinery, not from adding more modules.",[760,1421],{},[111,1423,1424],{},[765,1425,1426,1427,1431,1432,1436,1437,1440],{},"We have not measured this model's deployment difficulty, tooling support or real latency, so this is not a recommendation to migrate. If you do, measure three things first: how big your text trunk actually is, whether the branches can be switched off, and what share of your queries are cross-modal. Sources: ",[131,1428,1430],{"href":133,"rel":1429},[135],"arXiv 2610.37364",", ",[131,1433,1435],{"href":775,"rel":1434},[135],"Gemma docs",". Related MoE work: ",[131,1438,783],{"href":781,"rel":1439},[135]," (channel selection, not a MoE survey).",{"title":199,"searchDepth":786,"depth":786,"links":1442},[1443,1444,1445,1446,1450,1451,1456,1457,1458],{"id":836,"depth":786,"text":837},{"id":866,"depth":786,"text":867},{"id":897,"depth":786,"text":898},{"id":944,"depth":786,"text":945,"children":1447},[1448,1449],{"id":951,"depth":794,"text":952},{"id":971,"depth":794,"text":972},{"id":987,"depth":786,"text":988},{"id":1028,"depth":786,"text":1029,"children":1452},[1453,1454,1455],{"id":1035,"depth":794,"text":1036},{"id":1058,"depth":794,"text":1059},{"id":1074,"depth":794,"text":1075},{"id":1093,"depth":786,"text":1094},{"id":1332,"depth":786,"text":1333},{"id":1382,"depth":786,"text":1383},"\u002Fimages\u002Fblog\u002Fen\u002Fshared-trunk-four-modalities.svg","Parameter split showing a 271M text trunk, 167M vision encoder and 305M audio encoder isolated in branches that can be loaded on demand","Multimodal used to mean four models, or one big enough to swallow everything. EmbeddingGemma 2 shows a third path: share the semantic backbone, isolate capacity in branches, load only what you need.","en",{},"\u002Fimages\u002Fblog\u002Fen\u002Fshared-trunk-four-modalities.jpg","\u002Fblog\u002Fen\u002Fshared-trunk-four-modalities",{"title":826,"description":1461},"blog\u002Fen\u002Fshared-trunk-four-modalities",[819,820,821],"wi1YnHGolk3e0hWo1T4Qw9oxzCu8vI22647doO01M-Q",[1471,1796],{"id":1472,"title":1473,"author":103,"body":1474,"cover":1781,"coverAlt":1782,"date":1783,"description":1784,"draft":809,"extension":810,"featured":809,"lang":811,"meta":1785,"navigation":225,"ogImage":1786,"path":1787,"seo":1788,"slug":1789,"stem":1790,"tags":1791,"translationKey":1789,"updated":822,"__hash__":1795},"blog\u002Fblog\u002Fzh-TW\u002Feth-refuted-not-p-np.md","朋友問我：P=NP 被攻破了嗎？",{"type":105,"value":1475,"toc":1772},[1476,1479,1482,1485,1488,1491,1494,1497,1500,1503,1509,1512,1518,1527,1530,1533,1536,1539,1542,1548,1551,1558,1561,1564,1567,1574,1577,1588,1592,1599,1602,1605,1608,1611,1614,1617,1620,1623,1626,1629,1640,1643,1646,1649,1652,1655,1658,1661,1664,1667,1670,1673,1684,1687,1690,1693,1702,1705,1708,1711,1714,1717,1720,1723,1726,1729,1732,1735,1738,1741,1744,1747,1750,1753,1756,1758],[111,1477,1478],{},"星期二晚上，朋友在群組丟來一張新聞截圖，問我：「這是說 P=NP 被攻破了嗎？」",[111,1480,1481],{},"我說：「不是 P=NP，是 ETH。」",[111,1483,1484],{},"他問：「那不是一樣嗎？」",[111,1486,1487],{},"那個當下我其實沒有好答案。隔天中午我們碰面，我花了半小時把這件事從頭講一次。這篇就是那次講的內容，加上後來我自己去讀原始手稿補上的細節。",[115,1489,1490],{"id":1490},"那天到底發生了什麼",[111,1492,1493],{},"OpenAI 在十月六日先把七百多篇數學手稿放上 GitHub。十月二十六日又加了兩篇，其中一篇是目錄上的第 378 號：〈Disproof of the Exponential Time Hypothesis〉，論文自己的標題則是〈The Exponential Time Hypothesis is False〉。",[111,1495,1496],{},"整個庫現在是 721 篇手稿、373 個家族。家族編號有跳號，所以編號到 378，不代表真的有 378 個家族。",[111,1498,1499],{},"幾小時後，Anthropic 也說他們的模型在 10 月 16 日獨立找到核心想法，並稱 OpenAI 的模型「獨立且同時」證明了類似的結果。",[111,1501,1502],{},"朋友打斷我：「所以是誰先的？」",[111,1504,1505,1506,299],{},"「不知道，也不重要。」我跟他說，真正值得看的地方，是",[139,1507,1508],{},"這些東西還不是論文",[111,1510,1511],{},"OpenAI 自己在 README 裡寫得很白：這些是「原始研究資料」，不是打磨過的學術論文，而且「部分沒有形式化的結果可能有問題」。整個庫大約 42% 的頂層結果有 Lean 形式化，這篇第 378 號是其中之一，核心的時間上界有被機器檢查過。",[111,1513,1514,1515,299],{},"學界沒有急著下結論。Impagliazzo、Paturi（1999 年提出 ETH，期刊版 2001）跟 MIT 的 Williams 組了工作組、發公開信，特別強調一件事：",[139,1516,1517],{},"參與，不等於認證",[111,1519,1520,1521,1526],{},"到 10 月 28 日，",[131,1522,1525],{"href":1523,"rel":1524},"https:\u002F\u002Fmathworld.wolfram.com\u002FExponentialTimeHypothesis.html",[135],"MathWorld"," 的紀錄寫得很清楚：還沒有獨立驗證。",[115,1528,1529],{"id":1529},"先從數獨開始",[111,1531,1532],{},"要講清楚這件事，得先講一個看起來完全不相關的東西：數獨。",[111,1534,1535],{},"朋友點頭，他玩數獨。",[111,1537,1538],{},"「你先想像一個填好的盤面。我拿給你看，你大概十秒就能確認對不對。但如果我給你一張空白盤，你要找到那組解，可能要半小時。」",[111,1540,1541],{},"這就是 P 與 NP 的差別。P 是「可以快速解」的問題，NP 是「答案可以快速驗證」的問題。",[111,1543,1544,1545],{},"而 P 等於 NP 這句話的意思是：",[139,1546,1547],{},"所有能快速驗證的答案，都能快速找到。",[111,1549,1550],{},"學界幾乎所有人都覺得這句話是錯的，但沒有人證明過。",[111,1552,1553,1554,1557],{},"ETH 比這句話更強。它不只是說「3SAT 不在 P 裡」，而是說：",[139,1555,1556],{},"存在一個常數 c，使得 3SAT 至少要 2 的 cn 次方步。"," 3SAT 是這一類問題裡最標準的那一個，跟數獨同樣是 NP-完全。",[111,1559,1560],{},"打個比方。P 不等於 NP 只說「兩者不是同一個位置」；ETH 說的是「兩者至少相距兩千公里」——後者資訊多得多。",[115,1562,1563],{"id":1563},"為什麼有人要假設這種東西",[111,1565,1566],{},"「假設一個沒人證明的東西，有什麼用？」朋友問。",[111,1568,1569,1570,1573],{},"用處在於：",[139,1571,1572],{},"如果你知道兩地至少相距兩千公里，就可以推論沒有人能在一小時內開車到。"," 距離的下界，直接排除掉一整類「很快」的解法。",[111,1575,1576],{},"過去二十年，「細粒度複雜度」這個領域就是這樣運作的：先假設 ETH 成立，再證明「如果這個問題有快速演算法，那 3SAT 就有快速演算法，矛盾」。所以 ETH 不是課本上的名詞，它是一整批論文的地基。",[215,1578,219,1580,219,1585],{"className":1579},[218],[221,1581],{"src":1582,"poster":1583,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":1584},"\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F03-ladder.mp4","\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F03-ladder-poster.jpg","複雜度階梯：多項式、次指數、指數三層",[228,1586,1587],{},"複雜度階梯。ETH 主張最上層不能更低；這次的結果把「上界」壓到中間那層。",[115,1589,1591],{"id":1590},"那-openai-到底做了什麼","那 OpenAI 到底做了什麼",[111,1593,1594,1595,1598],{},"他們的模型找到一個演算法，可以",[139,1596,1597],{},"算出","一個邏輯公式有幾組解——是「算出有幾組」，不是「判斷有沒有解」，比原本的要求更強。時間是 2 的 n 除以根號（自然對數 n）次方，空間則是多項式。",[111,1600,1601],{},"朋友指著我在紙上寫的那串式子：「這是什麼意思？」",[111,1603,1604],{},"「意思是指數還在，只是除以一個會慢慢變大的數字。n 越大，除得越多。」",[111,1606,1607],{},"「那不是變快了？」",[111,1609,1610],{},"「是。但要看跟誰比。」",[111,1612,1613],{},"目前最好的 3SAT 演算法（隨機化）大約是 1.307 的 n 次方。",[111,1615,1616],{},"假設 n 等於一千。那個等級的演算法，大約是 2 的 386 次方；這次的結果，大約是 2 的 380 次方。",[111,1618,1619],{},"只快了大約 60 倍。",[111,1621,1622],{},"60 倍聽起來不少，但 2 的 380 次方還是 10 的 114 次方。宇宙年齡以秒計，也只有 4 乘 10 的 17 次方。",[111,1624,1625],{},"真正的差別在 n 很大的時候。n 等於一百萬，一個是 2 的 269000 次方，一個是 2 的 386000 次方——差距拉開，但兩個都還是天文數字。",[111,1627,1628],{},"所以這次的重點不是「快多少」，而是「快過那條下界」。",[215,1630,219,1632,219,1637],{"className":1631},[218],[221,1633],{"src":1634,"poster":1635,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":1636},"\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F05-magnitude.mp4","\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F05-magnitude-poster.jpg","n=1000 時的量級比較：已知最快 2^386，這次 2^380",[228,1638,1639],{},"n = 1000 的對照。這次不是從 2^n 降到 2^380，而是從 2^386 降到 2^380——快 64 倍，仍然不可行。",[115,1641,1642],{"id":1642},"上下界要分清楚",[111,1644,1645],{},"這裡有個容易搞混的地方。",[111,1647,1648],{},"演算法是一種「上界」：它告訴我們，這件事最多需要多少時間就能做到。",[111,1650,1651],{},"ETH 是一種「下界」：它說，這件事至少需要多少時間。",[111,1653,1654],{},"這次的結果，把上界壓到次指數。對足夠大的 n，這個上界會低於 ETH 宣稱的下界——所以 ETH 被推翻。",[111,1656,1657],{},"朋友問：「所以反過來也成立嗎？」",[111,1659,1660],{},"「你說的『反過來』有兩種。」",[111,1662,1663],{},"P 等於 NP 確實會讓 ETH 不成立——但那是同一個箭頭的逆否命題，不是新的推論。",[111,1665,1666],{},"真正不成立的是另一種反推：手上只有「ETH 被推翻」，推不出「P 等於 NP」。這個箭頭，不成立。",[111,1668,1669],{},"用回距離的比喻：原本大家以為至少兩千公里，現在證明只有八百公里。距離變短了，但你手上那台車還是太慢，一小時內到不了。",[111,1671,1672],{},"所以 P 是否等於 NP，到今天為止，完全開放。",[215,1674,219,1676,219,1681],{"className":1675},[218],[221,1677],{"src":1678,"poster":1679,"autoPlay":225,"loop":225,"muted":225,"playsInline":225,"ariaLabel":1680},"\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F04-implication.mp4","\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np\u002F04-implication-poster.jpg","兩個箭頭：上面成立、下面打叉",[228,1682,1683],{},"上面那個箭頭成立，下面那個不成立。這是整件事最常被誤讀的地方。",[115,1685,1686],{"id":1686},"那為什麼還是大事",[111,1688,1689],{},"「既然 P 等於 NP 沒有被動到，那大家在興奮什麼？」",[111,1691,1692],{},"「因為它拆掉了很多人的直覺。」",[111,1694,1695,1696,1701],{},"Ryan Williams ",[131,1697,1700],{"href":1698,"rel":1699},"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F10\u002F27\u002Ffrontier-ai-labs-say-their-models-refuted-the-exponential-time-hypothesis-what-does-that-mean\u002F",[135],"對媒體說","，證明的有些部分也許算例行公事，但確實有兩三個「看起來很外星人」的部分，正是這些部分讓他覺得，這是整個社群要一起處理的課題。",[111,1703,1704],{},"他也說了一句很誠實的話：「每一位電腦科學家大概都該把 P 等於 NP 的可能性權重調高一點，但調整多少因人而異。」",[111,1706,1707],{},"「所以有人檢查過了嗎？」朋友問。",[111,1709,1710],{},"「有，但還沒檢查完。」",[111,1712,1713],{},"手稿裡大約 42% 的頂層結果有 Lean 形式化。Lean 是一套能讓電腦逐步檢查證明的工具。可是電腦檢查過，不等於結論成立——它確認推導沒有漏洞，卻不能保證題目定義是大家想的那一題，也不能保證這個定理有用。",[111,1715,1716],{},"實際影響大概三塊。",[111,1718,1719],{},"第一，細粒度複雜度要重看一批論文。那些「假設 ETH 成立，則某問題不能低於某個複雜度」的條件句，前提沒了；結論不一定變假，但不能再拿來推無條件下界。",[111,1721,1722],{},"第二，密碼學沒有因此被破。現行系統依賴的是因數分解、離散對數、格問題這幾類特定問題，不是 3SAT 這個家族；而且這次的演算法還是指數時間。",[111,1724,1725],{},"第三，工程上短期完全無感。任何拿這件事說「AI 推論要變快了」的說法，都是把理論上界當成執行時間。",[115,1727,1728],{"id":1728},"回到那個群組",[111,1730,1731],{},"隔天吃完午飯，朋友問我最後一個問題：「所以這是假新聞嗎？」",[111,1733,1734],{},"「不是。」我說，「它有真東西。錯的是把它講成『攻破 P=NP』。」",[111,1736,1737],{},"「那我以後要怎麼看這種新聞？」",[111,1739,1740],{},"我想了一下，伸出三根手指。",[111,1742,1743],{},"「第一，看它是宣稱還是證明。發在 GitHub 上的手稿，跟期刊論文不是同一個等級的東西——更何況 OpenAI 自己就寫了『這是原始研究資料』。",[111,1745,1746],{},"第二，看它形式化到哪裡。Lean 能檢查推導，但不能保證題目是大家想的那一題。",[111,1748,1749],{},"第三，看它推翻了什麼、又沒推翻什麼。這次推翻的是『3SAT 至少要 2 的 cn 次方』這個下界的宣稱；它沒有推翻 P 不等於 NP，也沒有給出任何能用的演算法。」",[111,1751,1752],{},"他聽完，把手機收起來，說了一句：「所以那則新聞，是把『距離沒那麼遠』寫成『已經到了』。」",[111,1754,1755],{},"對。就是這樣。",[760,1757],{},[111,1759,1760],{},[765,1761,1762,1763,1768,1769,299],{},"這篇寫在 2026-10-29，學界審查還在進行中。如果後續出現獨立驗證、或證明被找出問題，我會回來更新。原始資料在 ",[131,1764,1767],{"href":1765,"rel":1766},"https:\u002F\u002Fgithub.com\u002Fopenai\u002Fmath",[135],"openai\u002Fmath"," 的第 378 號家族；ETH 的定義與推論可以看 ",[131,1770,1525],{"href":1523,"rel":1771},[135],{"title":199,"searchDepth":786,"depth":786,"links":1773},[1774,1775,1776,1777,1778,1779,1780],{"id":1490,"depth":786,"text":1490},{"id":1529,"depth":786,"text":1529},{"id":1563,"depth":786,"text":1563},{"id":1590,"depth":786,"text":1591},{"id":1642,"depth":786,"text":1642},{"id":1686,"depth":786,"text":1686},{"id":1728,"depth":786,"text":1728},"\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np.svg","複雜度階梯圖，標出這次把上界從指數推到次指數，但還沒到多項式","2026-10-29","朋友在群組丟來一則新聞，問我「P=NP 是不是被攻破了」。這則新聞是真的有東西，但朋友把兩件事混在一起了：被推翻的是 ETH，不是 P=NP。而且這次改善的是「上界」，不是「我們終於有快演算法」。這篇是把話講清楚的版本。",{},"\u002Fimages\u002Fblog\u002Feth-refuted-not-p-np.jpg","\u002Fblog\u002Fzh-tw\u002Feth-refuted-not-p-np",{"title":1473,"description":1784},"eth-refuted-not-p-np","blog\u002Fzh-TW\u002Feth-refuted-not-p-np",[1792,1793,1794],"complexity","ai-research","mathematics","NMrlp3rpEa3JWDfQNgon-vgHLEx-iVcVc5fiZjvqrAk",{"id":1797,"title":1798,"author":103,"body":1799,"cover":2647,"coverAlt":2648,"date":2649,"description":2650,"draft":809,"extension":810,"featured":809,"lang":811,"meta":2651,"navigation":225,"ogImage":2652,"path":2653,"seo":2654,"slug":2655,"stem":2656,"tags":2657,"translationKey":2655,"updated":822,"__hash__":2661},"blog\u002Fblog\u002Fzh-TW\u002Flogic-folding-fpga-cgra.md","FPGA 的死胡同，邏輯折疊可能是出口",{"type":105,"value":1800,"toc":2634},[1801,1806,1809,1812,1815,1821,1824,1828,1831,1853,1856,1859,1862,1915,1918,1961,1970,1973,1977,1980,1997,2007,2019,2031,2035,2038,2044,2050,2056,2062,2069,2073,2076,2082,2088,2094,2106,2109,2113,2167,2171,2174,2381,2384,2433,2436,2439,2442,2514,2517,2520,2572,2575,2578,2584,2590,2596,2602,2605,2613,2619,2625,2631],[108,1802,1803],{},[111,1804,1805],{},"這篇沒有我們的實測。已發生的事實都有連結；預測與機率是我們的判斷，會標明。我們對這個方向是樂觀的，但會把「要看到什麼才算成立」寫清楚。",[115,1807,1808],{"id":1808},"先講結論",[111,1810,1811],{},"2026 年 9 月，麒麟 2026 出貨。它是第一顆把邏輯折疊帶進量產的 SoC。拆機之後的數字是：密度從 1.55 億顆\u002Fmm² 升到 2.38 億（+55%），同效能下 NPU 功耗降 66%、CPU 大核降 41%。",[111,1813,1814],{},"這不是路線圖，是已經在手上的手機晶片。",[111,1816,1817,1818],{},"所以真正的問題不是「華為行不行」。是：",[139,1819,1820],{},"同一個槓桿，能不能把 FPGA 與 CGRA 從二十年的死胡同裡拉出來？",[111,1822,1823],{},"我的答案：有望。但理由不是「數字變大」，而是它同時鬆開三個卡點，而這三個卡點剛好是 AI 時代最痛的地方。",[115,1825,1827],{"id":1826},"死胡同fpga-卡在什麼地方","死胡同：FPGA 卡在什麼地方",[111,1829,1830],{},"FPGA 不是不強，是三個體質問題讓它一直停在「膠合與原型」：",[1832,1833,1834,1841,1847],"ol",{},[1835,1836,1837,1840],"li",{},[139,1838,1839],{},"面積。"," 配置 SRAM 與可程式開關吃掉一大塊面積，同樣功能比 ASIC 大一個數量級。",[1835,1842,1843,1846],{},[139,1844,1845],{},"記憶體牆。"," 可程式織體換不到 HBM 級的頻寬，只能外掛 DRAM；推論一進入大模型，就先輸在搬資料。",[1835,1848,1849,1852],{},[139,1850,1851],{},"靜態性。"," 權重與路由在合成時就固定，模型一改就得重新走一次工具鏈。這件事在模型每季改版的年代，變成最大的成本。",[111,1854,1855],{},"第三點最少被討論，但它其實是 FPGA 在 AI 時代最應該贏的地方。只是過去的硬體做不到「改得夠快、改得夠便宜」。",[115,1857,1858],{"id":1858},"折疊給了三把鑰匙",[111,1860,1861],{},"邏輯折疊在 2026 年指兩件不同的事，而它們解的是不同的鎖：",[400,1863,1864,1874],{},[403,1865,1866],{},[406,1867,1868,1871],{},[409,1869,1870],{},"華為 LogicFolding（折 Z 軸）",[409,1872,1873],{},"學術 temporal logic folding（折 T 軸）",[419,1875,1876,1889,1902],{},[406,1877,1878,1884],{},[424,1879,1880,1883],{},[139,1881,1882],{},"做法","：一個邏輯區塊拆成上下兩層、面對面鍵合",[424,1885,1886,1888],{},[139,1887,1882],{},"：同一組 LUT 每隔幾個 cycle 重配置一次",[406,1890,1891,1897],{},[424,1892,1893,1896],{},[139,1894,1895],{},"解鎖","：密度、頻率、更短的路徑",[424,1898,1899,1901],{},[139,1900,1895],{},"：用時間換面積，重配置變成常態",[406,1903,1904,1910],{},[424,1905,1906,1909],{},[139,1907,1908],{},"實證","：麒麟 2026（已量產）",[424,1911,1912,1914],{},[139,1913,1908],{},"：NATURE／NanoMap、Tabula Spacetime",[111,1916,1917],{},"三把鑰匙分別對應三個卡點：",[1919,1920,1921,1933,1949],"ul",{},[1835,1922,1923,1926,1927,1932],{},[139,1924,1925],{},"密度","：把配置記憶體搬出邏輯層（BEOL 配置記憶體，",[131,1928,1931],{"href":1929,"rel":1930},"https:\u002F\u002Fdoi.org\u002F10.1109\u002FDAC63849.2025.11132615",[135],"M3D FPGA，DAC 2025","），面積稅就有機會打折。",[1835,1934,1935,1938,1939,1944,1945,299],{},[139,1936,1937],{},"頻寬","：把 SRAM 或 DRAM 貼上去。28nm 3D logic-on-SRAM CGRA 已經做出 ",[131,1940,1943],{"href":1941,"rel":1942},"https:\u002F\u002Fieeexplore.ieee.org\u002Fdocument\u002F11509617",[135],"4TB\u002Fs 垂直頻寬","，51.19mm²、",[131,1946,1948],{"href":1947},"mailto:1GHz@0.9V","1GHz@0.9V",[1835,1950,1951,1954,1955,1960],{},[139,1952,1953],{},"重配置速度","：混合鍵合把重配置延遲壓到 sub-10ns（",[131,1956,1959],{"href":1957,"rel":1958},"https:\u002F\u002Fdoi.org\u002F10.1145\u002F3748173.3779206",[135],"vFPGA，2026","），時間折疊第一次在成本上說得通。",[111,1962,1963,1964,1969],{},"再加一項：面對面鍵合讓線變短。",[131,1965,1968],{"href":1966,"rel":1967},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2505.05579",[135],"LaZagna（ICCAD 2025）","在 VTR 模擬裡量到臨界路徑最多縮短 6.7%、總線長 22.9%。",[111,1971,1972],{},"單看每一項都不算革命。放在一起，是 FPGA 二十年來第一次有機會把三個稅同時往下壓。",[115,1974,1976],{"id":1975},"已經發生的事不是預測","已經發生的事，不是預測",[111,1978,1979],{},"這一節全部是可以查證的。它重要，因為它把「折疊」從論文搬到商品。",[111,1981,1982,1985,1986,299,1991,1996],{},[139,1983,1984],{},"密度與功耗。"," 麒麟 2026 對比前代平面架構的麒麟 9030 Pro：密度 +55%，同效能下 NPU 功耗 −66%、CPU 大核 −41%，",[131,1987,1990],{"href":1988,"rel":1989},"https:\u002F\u002Fwww.ithome.com\u002F0\u002F998\u002F598.htm",[135],"GPU 的數字各家報導落在 45–58%",[131,1992,1995],{"href":1993,"rel":1994},"https:\u002F\u002Fm.cena.com.cn\u002Fsemi\u002F20261021\u002F129796.html",[135],"華為的說法","是：晶片能耗大半花在金屬走線搬資料。折疊縮短路徑、降低互連電容，所以同效能可以用更低的頻率與電壓。",[111,1998,1999,2002,2003,299],{},[139,2000,2001],{},"做得比預期細。"," 華為強調這是「選擇性折疊」：只折關鍵路徑與關鍵時序。再配合熱感知布局，讓冷熱電路交錯、避免熱點垂直重疊。垂直互連已經超過 1 億根，",[131,2004,2006],{"href":1988,"rel":2005},[135],"目標三年內做到 720nm 頂層金屬節距、2 億根以上",[111,2008,2009,2012,2013,2018],{},[139,2010,2011],{},"製程也不再是藉口。"," SMIC 的 N+3 已經進入 5nm 級量產（",[131,2014,2017],{"href":2015,"rel":2016},"https:\u002F\u002Fwww.techpowerup.com\u002F344000\u002Fchinese-smic-achieves-5-nm-production-on-n-3-node-without-euv-tools",[135],"TechPowerUp","）。也就是說，折疊不是在「沒有先進製程」時的替代品，而是「有製程之後仍然值得做」的架構選項。這一點改變了整件事的性質。",[111,2020,2021,2024,2025,2030],{},[139,2022,2023],{},"官方路線圖把它排進 AI 加速器。"," 依韜定律論文 V2，2025 年的昇騰 910C、2026 年的 950 仍用 Chiplet 與 2.5D；",[131,2026,2029],{"href":2027,"rel":2028},"https:\u002F\u002Fwww.trendforce.com\u002Fnews\u002F2026\u002F07\u002F06\u002Fnews-huawei-expands-tau-scaling-law-v2-paper-highlighting-logicfolding-path-to-ascend-ai-chips-around-2030\u002F",[135],"約 2030 年的昇騰 990 會把邏輯折疊引進 AI 加速器","，此後 3D 折疊成為主要載體。",[115,2032,2034],{"id":2033},"往前看折疊真正解鎖的是什麼","往前看：折疊真正解鎖的是什麼",[111,2036,2037],{},"如果折疊成立，FPGA／CGRA 在 AI 推論的位置會從「膠合」變成「可重構的推理引擎」。有四件事會因此變得可行：",[111,2039,2040,2043],{},[139,2041,2042],{},"一、持續學習。"," 權重與路由可以隨資料更新，不必為每一版模型重投片。ASIC 的結構弱點正好在這裡：模型一改，它就過期。",[111,2045,2046,2049],{},[139,2047,2048],{},"二、動態路由。"," MoE 的 expert routing 本質上是「每個 token 走不同的路徑」。這件事在硬體層可變，比在軟體層用快取模擬更省。FPGA 與 CGRA 天生就在做路由。",[111,2051,2052,2055],{},[139,2053,2054],{},"三、動態權重與精度。"," 同一顆晶片可以今天跑 INT4、明天跑 FP8，或在長上下文時切換 KV 的存放層級。GPU 靠軟體排程，可重構硬體可以直接改電路。",[111,2057,2058,2061],{},[139,2059,2060],{},"四、邊緣與現場。"," 不能連網、不能大聲散熱、不能被斷料的場景，需要的是「買得到、改得動、十年供貨」。這一段 GPU 進不去，ASIC 的出貨量養不活。",[111,2063,2064,2065,2068],{},"把這四件事放在一起，就會看到一個和過去不同的結論：",[139,2066,2067],{},"FPGA 賣的不是算力，是面對不確定性的能力。"," 而 AI 的下一段，剛好充滿不確定性。",[115,2070,2072],{"id":2071},"但要小心求證四個還沒到","但要小心求證：四個還沒到",[111,2074,2075],{},"樂觀不等於盲目。以下四件事如果沒解決，折疊就只是漂亮的原型：",[111,2077,2078,2081],{},[139,2079,2080],{},"一、2026 的實作是保守版。"," 麒麟用的是微凸塊與標準間距鍵合，不是最細的間距；TSV 的禁置區還會吃掉標準元件的位置。真正的密度紅利要等間距往下走。",[111,2083,2084,2087],{},[139,2085,2086],{},"二、工具鏈。"," 3D 要跨 die 做 P&R、時序、熱、SI 簽核，把多顆晶粒當成一個連續的設計實體。對 FPGA 來說，這等於整個編譯流程要重寫。這件事比晶圓鍵合難。",[111,2089,2090,2093],{},[139,2091,2092],{},"三、散熱與良率。"," 兩層主動邏輯把功率密度翻倍。華為用手機證明「密度高不等於更熱」。但 FPGA 長期跑在無風扇、工業、基地台環境，考驗的是十年可靠度，不是跑分。",[111,2095,2096,2099,2100,2105],{},[139,2097,2098],{},"四、成本。"," 混合鍵合連 HBM4 都還在",[131,2101,2104],{"href":2102,"rel":2103},"https:\u002F\u002Ftechnews.tw\u002Fai-agent\u002Fhuaweis-chip-law-looks-less-like-moore-and-more-like-marketing\u002F11265\u002F",[135],"觀望","。良率相乘、成本相加，誰來付溢價，目前沒有答案。",[111,2107,2108],{},"這四點不是反對理由，是「什麼時候知道它成立」的清單。",[115,2110,2112],{"id":2111},"檢查點未來三年要看什麼","檢查點：未來三年要看什麼",[400,2114,2115,2125],{},[403,2116,2117],{},[406,2118,2119,2122],{},[409,2120,2121],{},"訊號",[409,2123,2124],{},"為什麼重要",[419,2126,2127,2135,2143,2151,2159],{},[406,2128,2129,2132],{},[424,2130,2131],{},"鍵合節距往 720nm 以下走、互連數破 2 億",[424,2133,2134],{},"密度紅利能不能延續",[406,2136,2137,2140],{},[424,2138,2139],{},"昇騰 990 如期在 2030 前導入折疊",[424,2141,2142],{},"AI 加速器真的採用，不只是手機",[406,2144,2145,2148],{},[424,2146,2147],{},"出現第一顆商用的 3D logic-on-logic FPGA",[424,2149,2150],{},"從「華為的 SoC」變成「產業的選項」",[406,2152,2153,2156],{},[424,2154,2155],{},"3D-aware 的 EDA 與編譯器能把利用率拉高",[424,2157,2158],{},"沒有工具鏈，硬體只是樣品",[406,2160,2161,2164],{},[424,2162,2163],{},"邊緣裝置的 tokens\u002FJ 明顯優於 2026 年的 GPU",[424,2165,2166],{},"商業價值最終要落在每瓦推論",[115,2168,2170],{"id":2169},"_2030-的樂觀情境","2030 的樂觀情境",[111,2172,2173],{},"推論速度的關係很簡單：",[2175,2176,2179],"span",{"className":2177},[2178],"katex-display",[2175,2180,2183,2228],{"className":2181},[2182],"katex",[2175,2184,2187],{"className":2185},[2186],"katex-mathml",[2188,2189,2192],"math",{"xmlns":2190,"display":2191},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML","block",[2193,2194,2195,2223],"semantics",{},[2196,2197,2198,2202,2206],"mrow",{},[2199,2200,2201],"mtext",{},"tok\u002Fs",[2203,2204,2205],"mo",{},"≈",[2207,2208,2209,2220],"mfrac",{},[2196,2210,2211,2214,2217],{},[2199,2212,2213],{},"記憶體頻寬",[2203,2215,2216],{},"×",[2199,2218,2219],{},"頻寬效率",[2199,2221,2222],{},"每 token 要讀的權重",[2224,2225,2227],"annotation",{"encoding":2226},"application\u002Fx-tex","\\text{tok\u002Fs} \\approx \\frac{\\text{記憶體頻寬} \\times \\text{頻寬效率}}{\\text{每 token 要讀的權重}}",[2175,2229,2233,2261],{"className":2230,"ariaHidden":2232},[2231],"katex-html","true",[2175,2234,2237,2242,2249,2254,2258],{"className":2235},[2236],"base",[2175,2238],{"className":2239,"style":2241},[2240],"strut","height:1em;vertical-align:-0.25em;",[2175,2243,2246],{"className":2244},[2245,195],"mord",[2175,2247,2201],{"className":2248},[2245],[2175,2250],{"className":2251,"style":2253},[2252],"mspace","margin-right:0.2778em;",[2175,2255,2205],{"className":2256},[2257],"mrel",[2175,2259],{"className":2260,"style":2253},[2252],[2175,2262,2264,2268],{"className":2263},[2236],[2175,2265],{"className":2266,"style":2267},[2240],"height:2.0463em;vertical-align:-0.686em;",[2175,2269,2271,2276,2377],{"className":2270},[2245],[2175,2272],{"className":2273},[2274,2275],"mopen","nulldelimiter",[2175,2277,2279],{"className":2278},[2207],[2175,2280,2284,2368],{"className":2281},[2282,2283],"vlist-t","vlist-t2",[2175,2285,2288,2363],{"className":2286},[2287],"vlist-r",[2175,2289,2293,2320,2331],{"className":2290,"style":2292},[2291],"vlist","height:1.3603em;",[2175,2294,2296,2301],{"style":2295},"top:-2.314em;",[2175,2297],{"className":2298,"style":2300},[2299],"pstrut","height:3em;",[2175,2302,2304],{"className":2303},[2245],[2175,2305,2307,2312,2316],{"className":2306},[2245,195],[2175,2308,2311],{"className":2309},[2245,2310],"cjk_fallback","每",[2175,2313,2315],{"className":2314},[2245]," token ",[2175,2317,2319],{"className":2318},[2245,2310],"要讀的權重",[2175,2321,2323,2326],{"style":2322},"top:-3.23em;",[2175,2324],{"className":2325,"style":2300},[2299],[2175,2327],{"className":2328,"style":2330},[2329],"frac-line","border-bottom-width:0.04em;",[2175,2332,2334,2337],{"style":2333},"top:-3.677em;",[2175,2335],{"className":2336,"style":2300},[2299],[2175,2338,2340,2346,2350,2354,2357],{"className":2339},[2245],[2175,2341,2343],{"className":2342},[2245,195],[2175,2344,2213],{"className":2345},[2245,2310],[2175,2347],{"className":2348,"style":2349},[2252],"margin-right:0.2222em;",[2175,2351,2216],{"className":2352},[2353],"mbin",[2175,2355],{"className":2356,"style":2349},[2252],[2175,2358,2360],{"className":2359},[2245,195],[2175,2361,2219],{"className":2362},[2245,2310],[2175,2364,2367],{"className":2365},[2366],"vlist-s","​",[2175,2369,2371],{"className":2370},[2287],[2175,2372,2375],{"className":2373,"style":2374},[2291],"height:0.686em;",[2175,2376],{},[2175,2378],{"className":2379},[2380,2275],"mclose",[111,2382,2383],{},"以 7B 模型、INT4、batch 1、頻寬效率 0.6 估算（權重約 3.5 GB）：",[400,2385,2386,2398],{},[403,2387,2388],{},[406,2389,2390,2393,2395],{},[409,2391,2392],{},"機器",[409,2394,1937],{},[409,2396,2397],{},"推論速度",[419,2399,2400,2411,2422],{},[406,2401,2402,2405,2408],{},[424,2403,2404],{},"2026 邊緣 FPGA（LPDDR5）",[424,2406,2407],{},"25 GB\u002Fs",[424,2409,2410],{},"約 4 tok\u002Fs",[406,2412,2413,2416,2419],{},[424,2414,2415],{},"2030 3D FPGA（3D DRAM）",[424,2417,2418],{},"2 TB\u002Fs",[424,2420,2421],{},"約 340 tok\u002Fs",[406,2423,2424,2427,2430],{},[424,2425,2426],{},"2030 GPU（HBM4）",[424,2428,2429],{},"8 TB\u002Fs",[424,2431,2432],{},"約 1,370 tok\u002Fs",[111,2434,2435],{},"從 4 到 340 是 85 倍。這 85 倍跨了兩個產品級別，不能直接當成「同級比較」。",[111,2437,2438],{},"跟同年的 GPU 比，3D FPGA 仍然慢約四倍。但這個比較只在「大模型、批次、頻寬受限」時才決定勝負。在延遲敏感、模型碎片化、每瓦 token 被寫進規格書的場景，基準換了，答案也會換。",[111,2440,2441],{},"相對 2026 年同級 FPGA，我們的推估分佈是：",[400,2443,2444,2457],{},[403,2445,2446],{},[406,2447,2448,2451,2454],{},[409,2449,2450],{},"倍數",[409,2452,2453],{},"機率",[409,2455,2456],{},"情境",[419,2458,2459,2470,2481,2492,2503],{},[406,2460,2461,2464,2467],{},[424,2462,2463],{},"\u003C3×",[424,2465,2466],{},"15%",[424,2468,2469],{},"只做到 2-tier 邏輯，記憶體照舊",[406,2471,2472,2475,2478],{},[424,2473,2474],{},"3–10×",[424,2476,2477],{},"25%",[424,2479,2480],{},"3D DRAM 到位，編譯器仍弱",[406,2482,2483,2486,2489],{},[424,2484,2485],{},"10–30×",[424,2487,2488],{},"35%",[424,2490,2491],{},"邏輯、記憶體、工具鏈三者同時到位",[406,2493,2494,2497,2500],{},[424,2495,2496],{},"30–80×",[424,2498,2499],{},"20%",[424,2501,2502],{},"窄門：3B 以下模型、全 SRAM、邊緣對邊緣",[406,2504,2505,2508,2511],{},[424,2506,2507],{},">80×",[424,2509,2510],{},"5%",[424,2512,2513],{},"需要 CIM 或類比，或者基準作弊",[111,2515,2516],{},"中位數大約 15 倍，期望值大約 22 倍。這組數字偏樂觀，理由是麒麟已經把密度與功耗同時做出來。那是三把鑰匙裡最難的一把。",[115,2518,2519],{"id":2519},"我的判斷",[400,2521,2522,2531],{},[403,2523,2524],{},[406,2525,2526,2529],{},[409,2527,2528],{},"情境（未來 5 年）",[409,2530,2453],{},[419,2532,2533,2541,2549,2557,2564],{},[406,2534,2535,2538],{},[424,2536,2537],{},"折疊成為某一家 FPGA 旗艦的標準微架構",[424,2539,2540],{},"60%",[406,2542,2543,2546],{},[424,2544,2545],{},"中國 3D 路線在 AI 加速器落地（昇騰 990 如期）",[424,2547,2548],{},"65%",[406,2550,2551,2554],{},[424,2552,2553],{},"FPGA／CGRA 在邊緣推論拿下可觀市佔",[424,2555,2556],{},"40%",[406,2558,2559,2562],{},[424,2560,2561],{},"出現以折疊為核心賣點、年營收逾 10 億美元的新公司",[424,2563,2466],{},[406,2565,2566,2569],{},[424,2567,2568],{},"折疊讓 FPGA 在資料中心推論取代 GPU 的一部分",[424,2570,2571],{},"10%",[111,2573,2574],{},"這些是判斷，不是預測模型。它們的共同前提是：工具鏈要在 2028 年前追上硬體。追不上，上面全部往下修。",[115,2576,2577],{"id":2577},"這篇的限制",[111,2579,2580,2583],{},[139,2581,2582],{},"一、我們沒有實測。"," 所有數字來自公開報導、論文與廠商說法；麒麟那組功耗數字是華為自行量測。",[111,2585,2586,2589],{},[139,2587,2588],{},"二、樂觀有來源，也有偏誤。"," 麒麟是個案，而且是華為在特定條件下（選擇性折疊、手機功耗預算）調出來的結果。換到 600mm² 的 FPGA，散熱與良率的帳要重算。",[111,2591,2592,2595],{},[139,2593,2594],{},"三、機率是判斷。"," 沒有統計模型，請當成「我們的看法」。",[111,2597,2598,2601],{},[139,2599,2600],{},"四、未來的推估建立在三個假設上。"," 混合鍵合的良率與成本在 2030 年前達標、3D-aware 工具鏈成熟、2030 年的 GPU 沒有進步得比預期快。",[115,2603,2604],{"id":2604},"如果你要下注",[111,2606,2607,2608,2612],{},"三個問題，比記住任何數字有用。這是我們做",[131,2609,2611],{"href":2610},"\u002Fservices","概念驗證","時的前三題。",[111,2614,2615,2618],{},[139,2616,2617],{},"一、你的模型還會不會變？"," 會變，可重構的價值就成立；已經定型、出貨量又大，去開 ASIC。",[111,2620,2621,2624],{},[139,2622,2623],{},"二、你的瓶頸是頻寬還是延遲？"," 大模型批次推論，頻寬決定一切，FPGA 不會贏；延遲、離線、每瓦 token，這些是它的地盤。",[111,2626,2627,2630],{},[139,2628,2629],{},"三、你的比較基準是 die，還是板子？"," 只要競品是「2–4 顆離散元件加一塊板子」，折疊就有帳可算。",[111,2632,2633],{},"至於要不要現在押注——我們的經驗是，硬體路線的賭注通常輸在「太早下注」，而不是「太晚跟上」。但這一次，值得把它放進 2027 年的選項清單裡。",{"title":199,"searchDepth":786,"depth":786,"links":2635},[2636,2637,2638,2639,2640,2641,2642,2643,2644,2645,2646],{"id":1808,"depth":786,"text":1808},{"id":1826,"depth":786,"text":1827},{"id":1858,"depth":786,"text":1858},{"id":1975,"depth":786,"text":1976},{"id":2033,"depth":786,"text":2034},{"id":2071,"depth":786,"text":2072},{"id":2111,"depth":786,"text":2112},{"id":2169,"depth":786,"text":2170},{"id":2519,"depth":786,"text":2519},{"id":2577,"depth":786,"text":2577},{"id":2604,"depth":786,"text":2604},"\u002Fimages\u002Fblog\u002Flogic-folding-fpga-cgra.svg","邏輯折疊把邏輯區塊拆成上下兩層堆疊的示意圖","2026-10-25","麒麟 2026 已經把邏輯折疊帶進量產，密度 +55%、同效能功耗大降。同一個槓桿，能不能把 FPGA／CGRA 從二十年的死胡同裡拉出來？這篇看它解鎖什麼、還缺什麼、以及 2030 年前要盯哪些檢查點。",{},"\u002Fimages\u002Fblog\u002Flogic-folding-fpga-cgra.jpg","\u002Fblog\u002Fzh-tw\u002Flogic-folding-fpga-cgra",{"title":1798,"description":2650},"logic-folding-fpga-cgra","blog\u002Fzh-TW\u002Flogic-folding-fpga-cgra",[2658,2659,2660],"fpga","cgra","hardware","ZLO0bnOuzLOnC9C4wZH-KP3lBPxf7NcvEuwfvwTfbhc",1793531345994]