{ "text": "我最近開始比較大量串接 API 跑 Agent，\n後面又開始碰圖片跟影片生成之後，\n有一個問題變得越來越明顯。\n就是這些東西真的用多了之後，\n費用消耗其實非常快。\n所以我那時候很自然就開始想：\n現在開源模型都已經做到這種程度了，\n那我乾脆把模型放在自己的電腦裡面跑，不就好了？\n語言模型現在有千問、\nGemma 這些選擇，\n圖片跟影片也有越來越多模型可以自己部署。\n而且 Agent 又會一直反覆呼叫模型，\n在這樣的消耗底下，\n還不如本地的模型來去使用，\n結果真的開始研究之後，\n我第一個卡住的地方根本不是模型怎麼裝。\n而是：我要買什麼電腦？\n這件事情比我原本想像中麻煩很多。\n你開始查就會看到有人推四零九零、\n有人說直接買五零九零，\n有人覺得 AMD 同樣價格可以拿到更多記憶體，\n也有人說 Mac 的 Unified Memory\n才適合跑大模型。\n如果你本來就不是硬體玩家，\n看完這些東西其實只會更亂。\n但我後來真的開始看不同模型怎麼跑之後，\n反而發現一件事情：\n如果今天是為了本地 AI 買電腦，\n新手其實不用先研究顯卡型號。\n因為第一個真正會卡住你的東西，\n通常不是這張卡快不快是你的模型到\n底裝不裝得進去。\n而這裡第一個要看的，\n就是 VRAM。\nVRAM 簡單來說就是顯示卡自己的記憶體。\n這跟我們平常買遊戲電腦的思考方式有一點不一樣。\n像玩遊戲的時候，\n我們很常先看這張卡 FPS 多高、\n新一代快了多少。\n但是 AI 第一關會比較直接。\n假設今天一個模型跟整套 workflow\n需要的記憶體已經超過你的 VRAM，\n那這張卡算力再強，\n你還是得開始做各種妥協。\n可能把模型量化得更小、\n把一部分東西丟到系統 RAM，\n或者乾脆只能換小一點的模型，\n等於你用太小的 VRAM 時，\n你的限制就會越多。所以我現在覺得，\n買本地 AI 電腦第一個問題根本不是：\n「4090 還是 5090？」\n而是：「我到底需要多少 VRAM？」\n這個問題先確定，\n後面其實簡單很多。\n但這裡也很容易出現另外一種誤會。\n不是說今天一定要 24GB，\n低於 24GB 就完全不能跑 AI。\n如果你只是想跑比較小的語言模型，8GB、12GB、\n16GB 其實都有很多東西可以玩。\n甚至現在各種量化做得很好，\n一些原本很大的模型也可以壓到一般電腦上。\n問題是你今天想做的事情有多廣。\n如果我只想在本地跑聊天模型，\n需求其實沒有這麼誇張。\n但如果跟我一樣，\n我希望這台機器除了語言模型，\n還要跑 Coding Agent、\n圖片、ComfyUI，\n甚至連影片模型都想碰，\n那需求就會開始完全不一樣。\n以我自己來說，\n我會把 24GB 左右當成比較實際的全用途起點。\n不是因為 24GB 有什麼魔法。\n而是從這裡開始，\n你會比較少遇到「模型明明可以跑，\n但是為了塞進去，\n每一個 workflow 都要重新折\n騰一次」這種狀況。\n像 3090、4090 都是 24GB，\n而現在 5090 已經來到 32GB。\n對 AI 來說，5090 多出來的 8GB 有時候比單純\n效能提升還有感。\n因為它可能直接讓原本卡在邊界的模型或\n更長的 Context 放得進去。\n所以這也是為什麼我覺得，\n本地 AI 看顯卡不能只看快多少，\n重要是能不能運行需要的模型。\n這時，記憶體容量就是你能不能運行這類\n模型的第一個關鍵。\n這件事情到了影片模型會更明顯。\n因為一個影片 workflow 很多時候\n不只有主要生成模型。\n你可能還有 Text Encoder、\nVAE、LoRA，\n甚至其他前後處理。\n像 Wan 2.2 官方自己的 720p 單卡推理範例，\n就直接把 24GB VRAM 當成可以執行的級距。\n但這也不是代表所有影片模型一定需要 24GB。\n現在很多模型透過量化、\nCPU offload，\n甚至社群特別做的版本，\n16GB 一樣可以硬塞進去。\n像 MiniMax H3 現在甚至已經有人\n把消費級 16GB、\n24GB 的部署方式做出來。\n所以我覺得這裡真正要理解的不是某個固定數字。\n而是：能跑，跟跑得舒服，是兩件事情。\n16GB 可能透過很多優化真的跑成功。\n但是如果我今天每天都要用它工作，\n每一個新模型都要研究半天怎麼省 VRAM，\n那我自己就不會把這個叫做很舒服的配置。\nWan 2.2 官方的 720p TI2V\n範例就是至少 24GB VRAM；\nMiniMax H3 官方原始部署\n甚至以多 GPU 為主，\n消費卡之所以能跑，\n很多是後來靠量化與 offload 把門檻壓下來。\n更重要的是，\n像是量化模型通常會有犧牲掉的品質或者能力，\n這未必是一件好事。所以如果有人跟你說：\n「我 16GB 就什麼都能跑。」他不一定錯。\n但你要先問的是：\n他到底怎麼跑？\n跑多快？\n為了跑起來犧牲了多少東西？\n這比單純看到「成功運行」重要很多。\n那等到 VRAM 大概確定之後，\n我第二個會看的，\n才是這張卡後面的軟體生態。這也是為什麼\n如果今天是一個完全的新手，\n要我推薦一條比較少踩坑的路，\n我現在還是會比較偏 NVIDIA。\n原因不是 AMD 的卡不能跑，\n而是 NVIDIA 有一套叫做 CUDA 的東西。\n你可以先把它理解成：\n很多 AI 軟體跟顯示卡溝通時，\n最常用的一套環境。\n現在很多 AI 工具、新模型，尤其圖片、\n影片這一類，開發者第一個測試的通常就是\nNVIDIA 加 CUDA。\n所以你實際使用的時候，\n差別不一定只是速度。\n而是同一個模型，\n別人用 NVIDIA 可能照教學裝一裝就可以跑，\n你換成別的平台，\n有時候還要自己處理相容性、\n版本，甚至某些功能根本還沒支援。\n這個差距在單純跑語言模型時可能沒那麼明顯，\n但開始碰 ComfyUI、圖片、影片，\n尤其剛出的新模型，\n就會比較有感。\n所以我覺得 NVIDIA 很大一部分的優勢，\n其實不是顯卡本身，\n而是它後面那整套 CUDA 生態。\n至於 AMD 當然不是不能用。\nAMD 也有自己對應的環境，\n叫 ROCm，現在支援已經比以前完整很多，\n很多主流 AI 工具也都能跑。\n只是如果今天是新手，\n我會比較建議你先確認自己固定要用哪些模型、\n哪些工具，再決定要不要選 AMD。\n因為用比較低的價格拿到更多 VRAM 當然很香，\n但前提是你要用的東西真的都有支援。\n不然最後省下硬體的錢，\n卻花很多時間在處理環境跟相容性，\n使用體驗反而會變差。\n那 Apple 相關產品這邊就比較特別。\n如果今天主要是跑語言模型，\n確實是可以考慮 Mac 相關的產品。\n因為現在 Mac 用的 M 系列 這些晶片，像是 M1、M2，\n其實都是 Apple 自己設計的，\n這整套架構就叫 Apple Silicon。\n它跟一般桌機一個很大的差別，\n是 CPU 跟 GPU 可以共用同一塊記憶體。\nApple 把這個叫做 Unified\nMemory，也就是統一記憶體。\n一般 NVIDIA 電腦通常是系統 RAM\n跟顯示卡 VRAM 分開算。\n假設你今天買一張 24GB 的 4090，\n模型主要可以直接放進 GPU 的空間\n就是這 24GB。\n但 Mac 不太一樣，\n它可以直接做到 64GB、\n128GB，甚至更高的共享記憶體，\nCPU 跟 GPU 都可以使用。\n所以如果今天碰到的是非常大的語言模型，\nMac 反而有一個很特殊的優勢。\n因為有些模型不是電腦完全算不動，\n而是一般消費級顯卡的 VRAM 根本放不下。\n這時候高記憶體的 Mac，\n就有機會直接把更大的模型放進去跑。\n所以如果今天我的工作主要是 Local LLM，\n我其實真的會把 Mac 放進選項。\n而且現在也不只是語言模型。\n像 MiniMax H3 這種很大的影音模型，\n現在也已經有人真的在 M 系列 Mac 上面跑通。\n只是這裡就要講到 Apple 跟\nNVIDIA 另外一個很大的差別。\n前面講 NVIDIA 的時候有提到 CUDA，\n你可以把 CUDA 理解成目前很多 AI 軟體\n最常使用的一套 GPU 運算環境。\nApple 沒有 CUDA，\n它有自己的一套系統。\n像 Metal 是 Apple 讓程式\n使用 GPU 的底層技術，\nMLX 則是 Apple 做給機器學習使用的框架。\n其實一般使用者不用記這些名字，\n真正要知道的只有一件事情：\n同一個 AI 模型，\n在 NVIDIA 上可以跑，\n不代表直接搬到 Mac 就一定可以跑。\n很多模型需要有人另外把它適\n配到 Apple 的環境，\n或者做成專門給 Mac 使用的版本。\n這也是為什麼我覺得本地 AI 有一個很重要的觀念：\n跑得動，跟適不適合拿來工作，是兩回事。\n像 MiniMax H3 現在 Mac 確實可以跑，\n這件事情本身很厲害，\n也代表 Apple 在本地 AI 這一塊已經\n不是只能跑語言模型。\n但是如果我今天的工作是大量做圖片、\nComfyUI、影片生成，\n而且新的模型一出來我就想馬上拿來測，\n那目前我自己還是會比較偏 NVIDIA。\n因為很多新模型跟新的 workflow，\n通常還是先在 CUDA 上面被完整測試。\nNVIDIA 很可能照著官方或者社群\n教學就可以開始跑，\nMac 有時候則要等別人移植、\n調整流程，或者自己處理一些相容性的問題。\n所以 Apple 不是不能跑 AI，\n比較準確的說法是：\n如果主要跑大型語言模型，\nMac 的大容量共享記憶體其實非常有吸引力；\n但如果圖片、影片、ComfyUI、新模型全部都想玩，\n目前 NVIDIA 的整體生態還是比較省事。\n因為我買一台工作機，\n在意的不是「網路上有沒有人成功跑過一次」。\n我更在意的是，\n我每天真的要用這些東西的時候，\n到底麻不麻煩。\n那講到這裡，其實還有一個界線要先說清楚。\n前面我們一直在講 24GB、\n32GB，甚至 Mac 的大記憶體，\n這些都還是一般人真的有可能買回家的機器。\n但如果今天你的目標是：\n我要在家裡直接跑完整版的超大型模型。\n那就完全是另外一個等級了。\n像現在 DeepSeek V4 Flash 這種模型，\n雖然它每次回答的時候只會動用其中一部分，\n但整套模型本身還是非常大。\n你可以把它想成一間很大的公司，\n一次工作可能只叫幾個部門出來，\n但其他部門還是得存在。\n所以如果真的要完整部署這種模型，\n通常就不是買一張 5090 可以解決，\n而是會開始進到多張資料中心 GPU、\n伺服器這種等級。\n像 GB200、\nB200 這些當然很強，但價格、耗電、\n散熱跟整套機器都已經不是一般使用\n者會考慮的東西。\n所以我覺得一般人在玩 Local AI，\n大部分指的都是所謂的量化版本的模型，\n所以不需要追求：\n「最大的模型我也一定要滿血跑。」\n我們比較實際的做法，\n本來就是找適合自己的硬體 GPU，\n然後挑選適合的量化模型。\n只要它已經可以完成我的工作，\n那麼就是一個值得投資的考量了\n那回到一般人的工作機，\n顯卡選完之後，\n我自己還會補兩個東西。\n第一個是系統 RAM。\n前面 VRAM 是顯示卡自己的記憶體，\nRAM 則是整台電腦平常在用的記憶體。\n如果我都已經組到 24GB VRAM 這種等級，\n我自己通常會直接抓 64GB RAM。\n不是因為 32GB 不能用，\n而是實際工作時你不可能只開一個模型。\n瀏覽器、IDE、\nAgent、ComfyUI 都可能一起跑，\n而且模型太大塞不進 VRAM 的時候，\n有些東西還可以暫時丟到系統 RAM。\n這種做法通常會比較慢，\n但至少可以讓一些原本塞不下的模型跑起來。\n所以 64GB 對我來說比較像是留一點空間，\n不要每次換模型都開始撞牆。\n還有像是前面提到的 Mac 也是同概念，\n只是會直接整個記憶體考量而已。再來就是 SSD。\n這個真的玩本地 AI 之後很容易爆。\n因為模型十幾 GB、\n幾十 GB 很正常，\n圖片跟影片還會再加 VAE、\nLoRA、Encoder，\n各種版本全部累積起來會大量的吃滿空間，\n用一般儲存會影響到載入這些東西的速度。\n而且你一定不會只抓一個模型。\n所以如果真的要長期把這台當 AI 工作機，\nSSD 我也會盡量抓高一點。\n所以綜合前面這些條件，\n如果今天是我自己要組一台比較全面的\n本地 AI 工作機，\n我目前還是會優先選 NVIDIA，\n而且 VRAM 我會從 24GB 左右開始看。\n原因很簡單，一個是 CUDA 生態現在還是比較完整，\n很多新的模型、\nComfyUI workflow 跟影音\n工具通常都先支援 NVIDIA；\n另外就是 24GB 這個級距，\n已經可以涵蓋不少語言、\n圖片跟部分影片模型，\n不會每碰到一個新模型就先卡在 VRAM 不夠。\n有預算的話，32GB、\n48GB 當然會更舒服，\n但這就看每個人的預算。\n再來是系統記憶體。\n如果顯卡都已經選到這個等級，\n我自己會建議至少 64GB RAM。\n因為實際跑模型的時候，\n不只是 GPU 在工作，系統本身、Agent、\nComfyUI、瀏覽器，\n甚至模型 offload 都可能一起吃到 RAM。\n最後就是 SSD，\n這個我也不會省太多。\n就如前面說的，\n載入初始化的模型，\n很容易就可以吃掉大量空間，\n而且 SSD 的速度也會直接影響模型第\n一次載入的時間。\n所以如果要把它簡化成三件事情：\nGPU 跟 VRAM，\n主要決定模型能不能順利跑，\n以及跑得多快。\nRAM，決定系統有沒有足夠空間去載入、\noffload 跟同時處理其他工作。\nSSD，決定你能放多少模型，\n以及模型載入時要等多久。\n如果只是要我直接給一套比較不容易踩雷的方向，\n那我目前會是：NVIDIA、24GB VRAM 起跳、\n64GB RAM，\n再準備夠大的 SSD。\n這不是唯一答案，\n但如果今天目標是語言、圖片、影片、\nAgent 都想碰，\n而且希望新模型出來不要每次先處理\n一堆相容性問題，\n我覺得這是目前比較實際的一個起點。", "language": "zh", "duration": 876.0, "segments": [ { "id": 0, "start": 0.0, "end": 3.06, "text": "我最近開始比較大量串接 API 跑 Agent，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 1, "start": 3.28, "end": 6.06, "text": "後面又開始碰圖片跟影片生成之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 2, "start": 6.08, "end": 8.18, "text": "有一個問題變得越來越明顯。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 3, "start": 8.36, "end": 10.56, "text": "就是這些東西真的用多了之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 4, "start": 10.64, "end": 12.3, "text": "費用消耗其實非常快。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 5, "start": 12.56, "end": 14.74, "text": "所以我那時候很自然就開始想：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 6, "start": 15.0, "end": 17.74, "text": "現在開源模型都已經做到這種程度了，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 7, "start": 17.94, "end": 21.34, "text": "那我乾脆把模型放在自己的電腦裡面跑，不就好了？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 8, "start": 21.62, "end": 23.195, "text": "語言模型現在有千問、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 9, "start": 23.195, "end": 24.52, "text": "Gemma 這些選擇，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 10, "start": 24.76, "end": 28.16, "text": "圖片跟影片也有越來越多模型可以自己部署。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 11, "start": 28.38, "end": 31.14, "text": "而且 Agent 又會一直反覆呼叫模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 12, "start": 31.42, "end": 32.92, "text": "在這樣的消耗底下，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 13, "start": 32.98, "end": 35.02, "text": "還不如本地的模型來去使用，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 14, "start": 35.3, "end": 37.16, "text": "結果真的開始研究之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 15, "start": 37.26, "end": 40.28, "text": "我第一個卡住的地方根本不是模型怎麼裝。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 16, "start": 40.64, "end": 42.84, "text": "而是：我要買什麼電腦？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 17, "start": 43.06, "end": 45.84, "text": "這件事情比我原本想像中麻煩很多。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 18, "start": 46.16, "end": 48.824, "text": "你開始查就會看到有人推四零九零、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 19, "start": 48.84, "end": 50.868, "text": "有人說直接買五零九零，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 20, "start": 50.88, "end": 54.12, "text": "有人覺得 AMD 同樣價格可以拿到更多記憶體，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 21, "start": 54.36, "end": 56.42, "text": "也有人說 Mac 的 Unified Memory", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 22, "start": 56.42, "end": 57.88, "text": "才適合跑大模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 23, "start": 58.18, "end": 59.873, "text": "如果你本來就不是硬體玩家，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 24, "start": 59.873, "end": 62.54, "text": "看完這些東西其實只會更亂。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 25, "start": 62.66, "end": 65.8, "text": "但我後來真的開始看不同模型怎麼跑之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 26, "start": 65.9, "end": 67.24, "text": "反而發現一件事情：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 27, "start": 67.68, "end": 70.28, "text": "如果今天是為了本地 AI 買電腦，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 28, "start": 70.46, "end": 72.9, "text": "新手其實不用先研究顯卡型號。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 29, "start": 73.16, "end": 75.44, "text": "因為第一個真正會卡住你的東西，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 30, "start": 75.54, "end": 78.8, "text": "通常不是這張卡快不快是你的模型到", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 31, "start": 78.8, "end": 80.06, "text": "底裝不裝得進去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 32, "start": 80.48, "end": 81.92, "text": "而這裡第一個要看的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 33, "start": 81.92, "end": 83.203, "text": "就是 VRAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 34, "start": 83.328, "end": 86.52, "text": "VRAM 簡單來說就是顯示卡自己的記憶體。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 35, "start": 86.74, "end": 89.98, "text": "這跟我們平常買遊戲電腦的思考方式有一點不一樣。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 36, "start": 90.4, "end": 91.92, "text": "像玩遊戲的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 37, "start": 91.94, "end": 94.88, "text": "我們很常先看這張卡 FPS 多高、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 38, "start": 95.06, "end": 96.3, "text": "新一代快了多少。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 39, "start": 96.48, "end": 98.68, "text": "但是 AI 第一關會比較直接。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 40, "start": 99.0, "end": 101.5, "text": "假設今天一個模型跟整套 workflow", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 41, "start": 101.5, "end": 104.155, "text": "需要的記憶體已經超過你的 VRAM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 42, "start": 104.18, "end": 105.82, "text": "那這張卡算力再強，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 43, "start": 105.94, "end": 107.86, "text": "你還是得開始做各種妥協。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 44, "start": 108.08, "end": 109.94, "text": "可能把模型量化得更小、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 45, "start": 110.0, "end": 112.04, "text": "把一部分東西丟到系統 RAM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 46, "start": 112.22, "end": 114.46, "text": "或者乾脆只能換小一點的模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 47, "start": 114.76, "end": 116.54, "text": "等於你用太小的 VRAM 時，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 48, "start": 116.66, "end": 119.42, "text": "你的限制就會越多。所以我現在覺得，", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "你的限制就會越多所以我現在覺得，", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 49, "start": 119.42, "end": 122.48, "text": "買本地 AI 電腦第一個問題根本不是：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 50, "start": 122.7, "end": 124.98, "text": "「4090 還是 5090？」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 51, "start": 125.14, "end": 128.175, "text": "而是：「我到底需要多少 VRAM？」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 52, "start": 128.2, "end": 129.56, "text": "這個問題先確定，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 53, "start": 129.74, "end": 131.24, "text": "後面其實簡單很多。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 54, "start": 131.605, "end": 134.14, "text": "但這裡也很容易出現另外一種誤會。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 55, "start": 134.34, "end": 136.68, "text": "不是說今天一定要 24GB，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 56, "start": 136.94, "end": 139.62, "text": "低於 24GB 就完全不能跑 AI。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 57, "start": 139.96, "end": 144.36, "text": "如果你只是想跑比較小的語言模型，8GB、12GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 58, "start": 144.46, "end": 147.04, "text": "16GB 其實都有很多東西可以玩。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 59, "start": 147.04, "end": 149.66, "text": "甚至現在各種量化做得很好，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 60, "start": 149.68, "end": 153.06, "text": "一些原本很大的模型也可以壓到一般電腦上。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 61, "start": 153.26, "end": 155.84, "text": "問題是你今天想做的事情有多廣。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 62, "start": 156.26, "end": 158.7, "text": "如果我只想在本地跑聊天模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 63, "start": 158.9, "end": 160.52, "text": "需求其實沒有這麼誇張。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 64, "start": 160.84, "end": 162.04, "text": "但如果跟我一樣，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 65, "start": 162.28, "end": 164.54, "text": "我希望這台機器除了語言模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 66, "start": 164.64, "end": 165.86, "text": "還要跑 Coding Agent、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 67, "start": 166.14, "end": 167.56, "text": "圖片、ComfyUI，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 68, "start": 167.78, "end": 169.56, "text": "甚至連影片模型都想碰，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 69, "start": 169.78, "end": 171.88, "text": "那需求就會開始完全不一樣。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 70, "start": 172.2, "end": 173.32, "text": "以我自己來說，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 71, "start": 173.44, "end": 177.6, "text": "我會把 24GB 左右當成比較實際的全用途起點。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 72, "start": 177.6, "end": 180.3, "text": "不是因為 24GB 有什麼魔法。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 73, "start": 180.62, "end": 181.92, "text": "而是從這裡開始，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 74, "start": 182.16, "end": 184.64, "text": "你會比較少遇到「模型明明可以跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 75, "start": 184.66, "end": 185.98, "text": "但是為了塞進去，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 76, "start": 186.08, "end": 187.98, "text": "每一個 workflow 都要重新折", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 77, "start": 187.98, "end": 189.4, "text": "騰一次」這種狀況。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 78, "start": 189.86, "end": 193.44, "text": "像 3090、4090 都是 24GB，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 79, "start": 193.68, "end": 196.58, "text": "而現在 5090 已經來到 32GB。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 80, "start": 196.82, "end": 201.02, "text": "對 AI 來說，5090 多出來的 8GB 有時候比單純", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 81, "start": 201.02, "end": 202.22, "text": "效能提升還有感。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 82, "start": 202.22, "end": 205.82, "text": "因為它可能直接讓原本卡在邊界的模型或", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 83, "start": 205.82, "end": 207.9, "text": "更長的 Context 放得進去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 84, "start": 208.08, "end": 209.88, "text": "所以這也是為什麼我覺得，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 85, "start": 210.04, "end": 212.68, "text": "本地 AI 看顯卡不能只看快多少，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 86, "start": 212.86, "end": 215.14, "text": "重要是能不能運行需要的模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 87, "start": 215.4, "end": 218.42, "text": "這時，記憶體容量就是你能不能運行這類", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 88, "start": 218.42, "end": 219.74, "text": "模型的第一個關鍵。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 89, "start": 220.096, "end": 222.62, "text": "這件事情到了影片模型會更明顯。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 90, "start": 222.82, "end": 225.12, "text": "因為一個影片 workflow 很多時候", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 91, "start": 225.12, "end": 226.68, "text": "不只有主要生成模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 92, "start": 226.68, "end": 228.88, "text": "你可能還有 Text Encoder、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 93, "start": 229.04, "end": 230.28, "text": "VAE、LoRA，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 94, "start": 230.52, "end": 232.09, "text": "甚至其他前後處理。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 95, "start": 232.09, "end": 236.12, "text": "像 Wan 2.2 官方自己的 720p 單卡推理範例，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 96, "start": 236.28, "end": 240.353, "text": "就直接把 24GB VRAM 當成可以執行的級距。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 97, "start": 240.4, "end": 244.3, "text": "但這也不是代表所有影片模型一定需要 24GB。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 98, "start": 244.64, "end": 246.72, "text": "現在很多模型透過量化、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 99, "start": 246.8, "end": 247.92, "text": "CPU offload，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 100, "start": 248.08, "end": 249.9, "text": "甚至社群特別做的版本，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 101, "start": 250.1, "end": 252.26, "text": "16GB 一樣可以硬塞進去。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 102, "start": 252.52, "end": 255.48, "text": "像 MiniMax H3 現在甚至已經有人", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 103, "start": 255.48, "end": 256.96, "text": "把消費級 16GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 104, "start": 257.14, "end": 259.28, "text": "24GB 的部署方式做出來。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 105, "start": 259.6, "end": 263.26, "text": "所以我覺得這裡真正要理解的不是某個固定數字。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 106, "start": 263.54, "end": 267.06, "text": "而是：能跑，跟跑得舒服，是兩件事情。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 107, "start": 267.48, "end": 270.58, "text": "16GB 可能透過很多優化真的跑成功。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 108, "start": 270.82, "end": 273.24, "text": "但是如果我今天每天都要用它工作，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 109, "start": 273.4, "end": 276.44, "text": "每一個新模型都要研究半天怎麼省 VRAM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 110, "start": 276.44, "end": 279.62, "text": "那我自己就不會把這個叫做很舒服的配置。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 111, "start": 279.64, "end": 283.18, "text": "Wan 2.2 官方的 720p TI2V", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 112, "start": 283.26, "end": 286.213, "text": "範例就是至少 24GB VRAM；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 113, "start": 286.213, "end": 288.49, "text": "MiniMax H3 官方原始部署", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 114, "start": 288.49, "end": 290.18, "text": "甚至以多 GPU 為主，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 115, "start": 290.46, "end": 291.92, "text": "消費卡之所以能跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 116, "start": 291.98, "end": 295.26, "text": "很多是後來靠量化與 offload 把門檻壓下來。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 117, "start": 295.52, "end": 296.58, "text": "更重要的是，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 118, "start": 296.68, "end": 300.24, "text": "像是量化模型通常會有犧牲掉的品質或者能力，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 119, "start": 300.52, "end": 304.265, "text": "這未必是一件好事。所以如果有人跟你說：", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "這未必是一件好事所以如果有人跟你說：", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 120, "start": 304.265, "end": 307.58, "text": "「我 16GB 就什麼都能跑。」他不一定錯。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 121, "start": 307.78, "end": 309.1, "text": "但你要先問的是：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 122, "start": 309.21, "end": 310.48, "text": "他到底怎麼跑？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 123, "start": 310.54, "end": 311.2, "text": "跑多快？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 124, "start": 311.46, "end": 313.44, "text": "為了跑起來犧牲了多少東西？", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 125, "start": 313.74, "end": 316.52, "text": "這比單純看到「成功運行」重要很多。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 126, "start": 316.949, "end": 319.02, "text": "那等到 VRAM 大概確定之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 127, "start": 319.12, "end": 320.36, "text": "我第二個會看的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 128, "start": 320.38, "end": 323.76, "text": "才是這張卡後面的軟體生態。這也是為什麼", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "才是這張卡後面的軟體生態這也是為什麼", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 129, "start": 323.76, "end": 325.74, "text": "如果今天是一個完全的新手，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 130, "start": 325.88, "end": 328.0, "text": "要我推薦一條比較少踩坑的路，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 131, "start": 328.0, "end": 330.517, "text": "我現在還是會比較偏 NVIDIA。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 132, "start": 330.517, "end": 332.8, "text": "原因不是 AMD 的卡不能跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 133, "start": 332.92, "end": 335.6, "text": "而是 NVIDIA 有一套叫做 CUDA 的東西。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 134, "start": 335.8, "end": 337.52, "text": "你可以先把它理解成：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 135, "start": 337.56, "end": 340.16, "text": "很多 AI 軟體跟顯示卡溝通時，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 136, "start": 340.24, "end": 341.64, "text": "最常用的一套環境。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 137, "start": 341.96, "end": 345.58, "text": "現在很多 AI 工具、新模型，尤其圖片、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 138, "start": 345.66, "end": 349.04, "text": "影片這一類，開發者第一個測試的通常就是", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 139, "start": 349.04, "end": 350.28, "text": "NVIDIA 加 CUDA。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 140, "start": 350.52, "end": 352.4, "text": "所以你實際使用的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 141, "start": 352.5, "end": 354.0, "text": "差別不一定只是速度。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 142, "start": 354.0, "end": 355.6, "text": "而是同一個模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 143, "start": 355.82, "end": 359.08, "text": "別人用 NVIDIA 可能照教學裝一裝就可以跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 144, "start": 359.26, "end": 360.62, "text": "你換成別的平台，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 145, "start": 360.82, "end": 362.88, "text": "有時候還要自己處理相容性、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 146, "start": 362.96, "end": 365.66, "text": "版本，甚至某些功能根本還沒支援。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 147, "start": 365.96, "end": 369.4, "text": "這個差距在單純跑語言模型時可能沒那麼明顯，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 148, "start": 369.6, "end": 372.6, "text": "但開始碰 ComfyUI、圖片、影片，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 149, "start": 372.96, "end": 374.34, "text": "尤其剛出的新模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 150, "start": 374.34, "end": 375.48, "text": "就會比較有感。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 151, "start": 375.76, "end": 378.6, "text": "所以我覺得 NVIDIA 很大一部分的優勢，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 152, "start": 378.68, "end": 380.16, "text": "其實不是顯卡本身，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 153, "start": 380.4, "end": 382.6, "text": "而是它後面那整套 CUDA 生態。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 154, "start": 382.96, "end": 385.32, "text": "至於 AMD 當然不是不能用。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 155, "start": 385.44, "end": 387.48, "text": "AMD 也有自己對應的環境，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 156, "start": 387.5, "end": 390.8, "text": "叫 ROCm，現在支援已經比以前完整很多，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 157, "start": 391.0, "end": 393.14, "text": "很多主流 AI 工具也都能跑。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 158, "start": 393.36, "end": 395.08, "text": "只是如果今天是新手，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 159, "start": 395.22, "end": 398.58, "text": "我會比較建議你先確認自己固定要用哪些模型、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 160, "start": 398.68, "end": 401.48, "text": "哪些工具，再決定要不要選 AMD。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 161, "start": 401.76, "end": 405.3, "text": "因為用比較低的價格拿到更多 VRAM 當然很香，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 162, "start": 405.54, "end": 408.2, "text": "但前提是你要用的東西真的都有支援。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 163, "start": 408.2, "end": 410.393, "text": "不然最後省下硬體的錢，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 164, "start": 410.393, "end": 413.28, "text": "卻花很多時間在處理環境跟相容性，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 165, "start": 413.52, "end": 415.08, "text": "使用體驗反而會變差。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 166, "start": 415.48, "end": 417.76, "text": "那 Apple 相關產品這邊就比較特別。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 167, "start": 418.0, "end": 420.16, "text": "如果今天主要是跑語言模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 168, "start": 420.34, "end": 422.54, "text": "確實是可以考慮 Mac 相關的產品。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 169, "start": 422.76, "end": 427.22, "text": "因為現在 Mac 用的 M 系列 這些晶片，像是 M1、M2，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 170, "start": 427.42, "end": 429.34, "text": "其實都是 Apple 自己設計的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 171, "start": 429.58, "end": 431.66, "text": "這整套架構就叫 Apple Silicon。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 172, "start": 432.06, "end": 434.24, "text": "它跟一般桌機一個很大的差別，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 173, "start": 434.32, "end": 437.56, "text": "是 CPU 跟 GPU 可以共用同一塊記憶體。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 174, "start": 437.56, "end": 439.62, "text": "Apple 把這個叫做 Unified", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 175, "start": 439.62, "end": 441.96, "text": "Memory，也就是統一記憶體。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 176, "start": 442.14, "end": 444.64, "text": "一般 NVIDIA 電腦通常是系統 RAM", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 177, "start": 444.64, "end": 446.76, "text": "跟顯示卡 VRAM 分開算。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 178, "start": 447.12, "end": 450.12, "text": "假設你今天買一張 24GB 的 4090，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 179, "start": 450.78, "end": 453.26, "text": "模型主要可以直接放進 GPU 的空間", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 180, "start": 453.26, "end": 454.54, "text": "就是這 24GB。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 181, "start": 454.92, "end": 456.1, "text": "但 Mac 不太一樣，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 182, "start": 456.24, "end": 458.22, "text": "它可以直接做到 64GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 183, "start": 458.36, "end": 461.565, "text": "128GB，甚至更高的共享記憶體，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 184, "start": 461.565, "end": 463.64, "text": "CPU 跟 GPU 都可以使用。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 185, "start": 463.64, "end": 467.02, "text": "所以如果今天碰到的是非常大的語言模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 186, "start": 467.12, "end": 469.5, "text": "Mac 反而有一個很特殊的優勢。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 187, "start": 469.68, "end": 472.26, "text": "因為有些模型不是電腦完全算不動，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 188, "start": 472.36, "end": 475.52, "text": "而是一般消費級顯卡的 VRAM 根本放不下。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 189, "start": 475.78, "end": 477.38, "text": "這時候高記憶體的 Mac，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 190, "start": 477.48, "end": 480.06, "text": "就有機會直接把更大的模型放進去跑。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 191, "start": 480.34, "end": 483.54, "text": "所以如果今天我的工作主要是 Local LLM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 192, "start": 483.76, "end": 486.1, "text": "我其實真的會把 Mac 放進選項。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 193, "start": 486.34, "end": 488.54, "text": "而且現在也不只是語言模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 194, "start": 488.78, "end": 491.74, "text": "像 MiniMax H3 這種很大的影音模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 195, "start": 491.74, "end": 495.38, "text": "現在也已經有人真的在 M 系列 Mac 上面跑通。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 196, "start": 495.58, "end": 497.66, "text": "只是這裡就要講到 Apple 跟", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 197, "start": 497.66, "end": 499.82, "text": "NVIDIA 另外一個很大的差別。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 198, "start": 500.14, "end": 502.6, "text": "前面講 NVIDIA 的時候有提到 CUDA，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 199, "start": 502.8, "end": 505.56, "text": "你可以把 CUDA 理解成目前很多 AI 軟體", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 200, "start": 505.56, "end": 507.96, "text": "最常使用的一套 GPU 運算環境。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 201, "start": 508.14, "end": 509.32, "text": "Apple 沒有 CUDA，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 202, "start": 509.5, "end": 510.98, "text": "它有自己的一套系統。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 203, "start": 511.2, "end": 512.997, "text": "像 Metal 是 Apple 讓程式", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 204, "start": 512.997, "end": 514.9, "text": "使用 GPU 的底層技術，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 205, "start": 515.12, "end": 518.48, "text": "MLX 則是 Apple 做給機器學習使用的框架。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 206, "start": 518.48, "end": 521.3, "text": "其實一般使用者不用記這些名字，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 207, "start": 521.36, "end": 523.42, "text": "真正要知道的只有一件事情：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 208, "start": 523.7, "end": 525.18, "text": "同一個 AI 模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 209, "start": 525.26, "end": 526.86, "text": "在 NVIDIA 上可以跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 210, "start": 526.94, "end": 529.4, "text": "不代表直接搬到 Mac 就一定可以跑。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 211, "start": 529.6, "end": 531.785, "text": "很多模型需要有人另外把它適", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 212, "start": 531.785, "end": 533.12, "text": "配到 Apple 的環境，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 213, "start": 533.26, "end": 535.62, "text": "或者做成專門給 Mac 使用的版本。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 214, "start": 535.84, "end": 539.44, "text": "這也是為什麼我覺得本地 AI 有一個很重要的觀念：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 215, "start": 539.68, "end": 543.0, "text": "跑得動，跟適不適合拿來工作，是兩回事。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 216, "start": 543.3, "end": 546.72, "text": "像 MiniMax H3 現在 Mac 確實可以跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 217, "start": 546.72, "end": 548.7, "text": "這件事情本身很厲害，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 218, "start": 548.9, "end": 551.52, "text": "也代表 Apple 在本地 AI 這一塊已經", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 219, "start": 551.52, "end": 553.14, "text": "不是只能跑語言模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 220, "start": 553.4, "end": 556.1, "text": "但是如果我今天的工作是大量做圖片、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 221, "start": 556.22, "end": 557.96, "text": "ComfyUI、影片生成，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 222, "start": 558.18, "end": 560.84, "text": "而且新的模型一出來我就想馬上拿來測，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 223, "start": 561.0, "end": 563.6, "text": "那目前我自己還是會比較偏 NVIDIA。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 224, "start": 563.78, "end": 566.02, "text": "因為很多新模型跟新的 workflow，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 225, "start": 566.14, "end": 568.82, "text": "通常還是先在 CUDA 上面被完整測試。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 226, "start": 569.12, "end": 571.56, "text": "NVIDIA 很可能照著官方或者社群", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 227, "start": 571.56, "end": 572.72, "text": "教學就可以開始跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 228, "start": 572.72, "end": 575.3, "text": "Mac 有時候則要等別人移植、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 229, "start": 575.34, "end": 578.54, "text": "調整流程，或者自己處理一些相容性的問題。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 230, "start": 579.0, "end": 581.16, "text": "所以 Apple 不是不能跑 AI，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 231, "start": 581.46, "end": 583.08, "text": "比較準確的說法是：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 232, "start": 583.22, "end": 585.22, "text": "如果主要跑大型語言模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 233, "start": 585.48, "end": 588.64, "text": "Mac 的大容量共享記憶體其實非常有吸引力；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 234, "start": 588.84, "end": 593.22, "text": "但如果圖片、影片、ComfyUI、新模型全部都想玩，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 235, "start": 593.44, "end": 596.62, "text": "目前 NVIDIA 的整體生態還是比較省事。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 236, "start": 596.74, "end": 598.4, "text": "因為我買一台工作機，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 237, "start": 598.62, "end": 601.46, "text": "在意的不是「網路上有沒有人成功跑過一次」。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 238, "start": 601.46, "end": 603.08, "text": "我更在意的是，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 239, "start": 603.16, "end": 605.38, "text": "我每天真的要用這些東西的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 240, "start": 605.48, "end": 606.68, "text": "到底麻不麻煩。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 241, "start": 606.955, "end": 610.0, "text": "那講到這裡，其實還有一個界線要先說清楚。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 242, "start": 610.32, "end": 612.66, "text": "前面我們一直在講 24GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 243, "start": 612.8, "end": 615.38, "text": "32GB，甚至 Mac 的大記憶體，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 244, "start": 615.6, "end": 618.8, "text": "這些都還是一般人真的有可能買回家的機器。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 245, "start": 619.02, "end": 621.0, "text": "但如果今天你的目標是：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 246, "start": 621.26, "end": 624.18, "text": "我要在家裡直接跑完整版的超大型模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 247, "start": 624.5, "end": 626.5, "text": "那就完全是另外一個等級了。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 248, "start": 626.5, "end": 629.84, "text": "像現在 DeepSeek V4 Flash 這種模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 249, "start": 629.96, "end": 633.26, "text": "雖然它每次回答的時候只會動用其中一部分，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 250, "start": 633.42, "end": 635.72, "text": "但整套模型本身還是非常大。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 251, "start": 635.9, "end": 638.24, "text": "你可以把它想成一間很大的公司，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 252, "start": 638.5, "end": 640.86, "text": "一次工作可能只叫幾個部門出來，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 253, "start": 641.08, "end": 642.86, "text": "但其他部門還是得存在。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 254, "start": 643.2, "end": 645.8, "text": "所以如果真的要完整部署這種模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 255, "start": 645.92, "end": 648.54, "text": "通常就不是買一張 5090 可以解決，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 256, "start": 648.7, "end": 651.5, "text": "而是會開始進到多張資料中心 GPU、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 257, "start": 651.507, "end": 653.0, "text": "伺服器這種等級。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 258, "start": 653.34, "end": 654.48, "text": "像 GB200、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 259, "start": 654.49, "end": 658.12, "text": "B200 這些當然很強，但價格、耗電、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 260, "start": 658.15, "end": 661.26, "text": "散熱跟整套機器都已經不是一般使用", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 261, "start": 661.26, "end": 662.42, "text": "者會考慮的東西。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 262, "start": 662.7, "end": 665.18, "text": "所以我覺得一般人在玩 Local AI，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 263, "start": 665.36, "end": 668.24, "text": "大部分指的都是所謂的量化版本的模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 264, "start": 668.46, "end": 669.7, "text": "所以不需要追求：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 265, "start": 669.96, "end": 672.2, "text": "「最大的模型我也一定要滿血跑。」", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 266, "start": 672.42, "end": 674.2, "text": "我們比較實際的做法，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 267, "start": 674.22, "end": 676.72, "text": "本來就是找適合自己的硬體 GPU，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 268, "start": 677.02, "end": 678.9, "text": "然後挑選適合的量化模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 269, "start": 678.9, "end": 681.32, "text": "只要它已經可以完成我的工作，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 270, "start": 681.36, "end": 683.9, "text": "那麼就是一個值得投資的考量了", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 271, "start": 684.181, "end": 685.9, "text": "那回到一般人的工作機，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 272, "start": 686.0, "end": 687.12, "text": "顯卡選完之後，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 273, "start": 687.2, "end": 688.88, "text": "我自己還會補兩個東西。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 274, "start": 689.14, "end": 690.66, "text": "第一個是系統 RAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 275, "start": 690.96, "end": 693.64, "text": "前面 VRAM 是顯示卡自己的記憶體，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 276, "start": 693.86, "end": 696.72, "text": "RAM 則是整台電腦平常在用的記憶體。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 277, "start": 696.9, "end": 700.5, "text": "如果我都已經組到 24GB VRAM 這種等級，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 278, "start": 700.62, "end": 703.34, "text": "我自己通常會直接抓 64GB RAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 279, "start": 703.48, "end": 705.94, "text": "不是因為 32GB 不能用，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 280, "start": 705.94, "end": 709.1, "text": "而是實際工作時你不可能只開一個模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 281, "start": 709.16, "end": 710.98, "text": "瀏覽器、IDE、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 282, "start": 711.12, "end": 713.68, "text": "Agent、ComfyUI 都可能一起跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 283, "start": 713.88, "end": 716.28, "text": "而且模型太大塞不進 VRAM 的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 284, "start": 716.52, "end": 718.84, "text": "有些東西還可以暫時丟到系統 RAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 285, "start": 718.94, "end": 721.0, "text": "這種做法通常會比較慢，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 286, "start": 721.2, "end": 724.32, "text": "但至少可以讓一些原本塞不下的模型跑起來。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 287, "start": 724.66, "end": 727.96, "text": "所以 64GB 對我來說比較像是留一點空間，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 288, "start": 728.16, "end": 730.37, "text": "不要每次換模型都開始撞牆。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 289, "start": 730.5, "end": 733.32, "text": "還有像是前面提到的 Mac 也是同概念，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 290, "start": 733.32, "end": 737.78, "text": "只是會直接整個記憶體考量而已。再來就是 SSD。", "source": "video_source_subtitle", "authority": "video_source_subtitle", "text_original": "只是會直接整個記憶體考量而已再來就是 SSD。", "correction_applied": true, "correction_source": "llm_correct_a_presegmentation" }, { "id": 291, "start": 738.02, "end": 740.6, "text": "這個真的玩本地 AI 之後很容易爆。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 292, "start": 740.86, "end": 742.38, "text": "因為模型十幾 GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 293, "start": 742.38, "end": 743.92, "text": "幾十 GB 很正常，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 294, "start": 744.08, "end": 746.37, "text": "圖片跟影片還會再加 VAE、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 295, "start": 746.37, "end": 747.85, "text": "LoRA、Encoder，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 296, "start": 747.88, "end": 750.98, "text": "各種版本全部累積起來會大量的吃滿空間，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 297, "start": 751.26, "end": 754.3, "text": "用一般儲存會影響到載入這些東西的速度。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 298, "start": 754.52, "end": 756.74, "text": "而且你一定不會只抓一個模型。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 299, "start": 757.02, "end": 760.08, "text": "所以如果真的要長期把這台當 AI 工作機，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 300, "start": 760.08, "end": 762.58, "text": "SSD 我也會盡量抓高一點。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 301, "start": 763.008, "end": 764.88, "text": "所以綜合前面這些條件，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 302, "start": 765.04, "end": 767.88, "text": "如果今天是我自己要組一台比較全面的", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 303, "start": 767.88, "end": 769.14, "text": "本地 AI 工作機，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 304, "start": 769.3, "end": 771.547, "text": "我目前還是會優先選 NVIDIA，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 305, "start": 771.547, "end": 774.98, "text": "而且 VRAM 我會從 24GB 左右開始看。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 306, "start": 775.12, "end": 778.9, "text": "原因很簡單，一個是 CUDA 生態現在還是比較完整，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 307, "start": 778.96, "end": 780.108, "text": "很多新的模型、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 308, "start": 780.108, "end": 782.22, "text": "ComfyUI workflow 跟影音", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 309, "start": 782.22, "end": 784.36, "text": "工具通常都先支援 NVIDIA；", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 310, "start": 784.36, "end": 786.72, "text": "另外就是 24GB 這個級距，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 311, "start": 786.74, "end": 788.32, "text": "已經可以涵蓋不少語言、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 312, "start": 788.44, "end": 790.2, "text": "圖片跟部分影片模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 313, "start": 790.48, "end": 793.48, "text": "不會每碰到一個新模型就先卡在 VRAM 不夠。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 314, "start": 793.6, "end": 795.76, "text": "有預算的話，32GB、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 315, "start": 795.9, "end": 798.02, "text": "48GB 當然會更舒服，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 316, "start": 798.22, "end": 799.96, "text": "但這就看每個人的預算。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 317, "start": 800.22, "end": 801.98, "text": "再來是系統記憶體。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 318, "start": 802.12, "end": 804.24, "text": "如果顯卡都已經選到這個等級，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 319, "start": 804.4, "end": 806.98, "text": "我自己會建議至少 64GB RAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 320, "start": 807.46, "end": 809.14, "text": "因為實際跑模型的時候，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 321, "start": 809.24, "end": 812.64, "text": "不只是 GPU 在工作，系統本身、Agent、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 322, "start": 812.92, "end": 814.58, "text": "ComfyUI、瀏覽器，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 323, "start": 814.76, "end": 817.5, "text": "甚至模型 offload 都可能一起吃到 RAM。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 324, "start": 817.82, "end": 819.28, "text": "最後就是 SSD，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 325, "start": 819.28, "end": 821.3, "text": "這個我也不會省太多。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 326, "start": 821.44, "end": 822.74, "text": "就如前面說的，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 327, "start": 822.82, "end": 824.26, "text": "載入初始化的模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 328, "start": 824.42, "end": 826.52, "text": "很容易就可以吃掉大量空間，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 329, "start": 826.84, "end": 829.74, "text": "而且 SSD 的速度也會直接影響模型第", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 330, "start": 829.74, "end": 830.98, "text": "一次載入的時間。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 331, "start": 831.26, "end": 834.02, "text": "所以如果要把它簡化成三件事情：", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 332, "start": 834.2, "end": 835.405, "text": "GPU 跟 VRAM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 333, "start": 835.405, "end": 837.68, "text": "主要決定模型能不能順利跑，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 334, "start": 837.68, "end": 838.82, "text": "以及跑得多快。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 335, "start": 839.1, "end": 842.2, "text": "RAM，決定系統有沒有足夠空間去載入、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 336, "start": 842.36, "end": 844.74, "text": "offload 跟同時處理其他工作。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 337, "start": 845.04, "end": 847.7, "text": "SSD，決定你能放多少模型，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 338, "start": 847.7, "end": 850.01, "text": "以及模型載入時要等多久。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 339, "start": 850.01, "end": 853.6, "text": "如果只是要我直接給一套比較不容易踩雷的方向，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 340, "start": 853.66, "end": 858.24, "text": "那我目前會是：NVIDIA、24GB VRAM 起跳、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 341, "start": 858.36, "end": 859.62, "text": "64GB RAM，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 342, "start": 859.88, "end": 861.7, "text": "再準備夠大的 SSD。", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 343, "start": 861.9, "end": 863.28, "text": "這不是唯一答案，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 344, "start": 863.5, "end": 866.54, "text": "但如果今天目標是語言、圖片、影片、", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 345, "start": 866.8, "end": 868.06, "text": "Agent 都想碰，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 346, "start": 868.32, "end": 870.98, "text": "而且希望新模型出來不要每次先處理", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 347, "start": 870.98, "end": 872.22, "text": "一堆相容性問題，", "source": "video_source_subtitle", "authority": "video_source_subtitle" }, { "id": 348, "start": 872.5, "end": 875.16, "text": "我覺得這是目前比較實際的一個起點。", "source": "video_source_subtitle", "authority": "video_source_subtitle" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 876.0, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260918_103114__本地_AI_電腦怎麼買_NVIDIA_Mac_VRAM_一次講懂/reference_subtitle.vtt", "reference_kind": "manual", "reference_language": "zh-TW", "selected_source": "video_source_subtitle", "source_caption_kind": "manual", "source_caption_processing": "verbatim_direct", "youtube_overlap_clamps": [], "general_high_confidence_repairs": [], "decision_reason": "complete_source_subtitle_fast_path_no_whisper", "whisper_rejected": false, "source_profile": { "segment_count": 349, "duration_coverage": 0.934474 } }, "presegmentation_correction": { "applied": true, "correction_count": 4, "deterministic_fix_count": 0 } }