統一記憶體是 Apple Silicon 跑本地 AI 的底牌：CPU 和 GPU 共用同一塊記憶體，模型整包載入，不必在系統記憶體和顯示卡記憶體之間搬。96GB（Apple 官網目前的比較機型配置）對上 M1 Max 的 32GB，「本地跑得動多大的模型」直接跳了一級。

Unified memory is Apple Silicon's ace for local AI: CPU and GPU share one pool, so a model loads whole — no shuffling between system RAM and VRAM. 96GB (the config on Apple's current compare page) against the M1 Max's 32GB moves "how big a model runs locally" up a full tier.

具體差在哪：32GB 的上限約是 14B 級模型；96GB 能讓 32B 常駐、夜間載入 70B 審查、週末單獨跑 120B。搭配 819GB/s 的頻寬餵 GPU，一台安靜的桌機就是務實的本地推論機——不用組雙顯卡、不用機房。

Concretely: 32GB tops out around 14B-class models; 96GB keeps a 32B resident, loads a 70B reviewer at night, and runs a 120B solo on weekends. With 819GB/s feeding the GPU, one quiet desktop becomes a practical local-inference box — no dual-GPU build, no server room.

CPU 從 M1 Ultra 的 20 核、M2 Ultra 的 24 核，到 M3 Ultra 的 32 核（24 個效能核心）。最有感的場景是 Xcode：大專案 clean build、同時開好幾台模擬器跑 UI 測試，效能核心多就是能一次吃下更多平行工作，等待明顯變短。

CPU went from 20 cores (M1 Ultra) to 24 (M2 Ultra) to 32 (24 performance cores) on M3 Ultra. Where it shows most is Xcode: a big clean build, or several simulators running UI tests at once, simply chews through more parallel work. The waiting gets noticeably shorter.

另一個容易被忽略的是 Thunderbolt 5（120Gb/s），比 TB4 的 40Gb/s 快三倍。外接高速 SSD 當專案／素材碟、接高解析度螢幕、未來擴充的餘裕都大很多。加上前後合計六個 TB5、兩個 USB-A、HDMI、10Gb 乙太網路和 SDXC 讀卡機，桌面接線終於不用再靠一堆轉接。

Easy to overlook: Thunderbolt 5 (120Gb/s) — 3× the 40Gb/s of TB4. External high-speed SSDs as project/asset drives, high-res displays, and future headroom all get a lot more comfortable. With six TB5 ports plus two USB-A, HDMI, 10Gb Ethernet, and an SDXC reader, my desk finally sheds its pile of dongles.

M3 世代第一次帶來 硬體加速光線追蹤 和 AV1 解碼，這兩項在 M1／M2 都沒有。AV1 是 YouTube、串流平台越來越常用的編碼，硬體解碼代表看片、處理這類素材時更省電、更順。媒體引擎也給到兩個影片解碼、四個影片編碼、四個 ProRes 編解碼引擎——如果你同時也剪影片、做內容，這代的媒體處理是實打實的加分。

The M3 generation is the first to bring hardware-accelerated ray tracing and AV1 decode — neither existed on M1 or M2. AV1 is increasingly common on YouTube and streaming platforms; hardware decode means smoother, more power-efficient playback and handling of that footage. The media engine also packs two video-decode, four video-encode, and four ProRes encode/decode engines — a real bonus if you also edit video or make content on the side.

為什麼這樣分？因為頻寬和記憶體上限決定了

Why split it this way? Bandwidth and the memory ceiling decide

本地大模型的生成速度，粗略正比於「記憶體頻寬 ÷ 模型大小」。同一顆 70B 模型放在 M3（819GB/s）大約 19 tok/s、放到 M1（400GB/s）只剩約 9 tok/s——這就是「大模型必須待在 M3」的物理依據。反過來，14B 這種小模型頻寬需求低，放在 M1 反而飛快（理論可到 40 tok/s 級），拿來做清洗、分類、檢索這類高頻雜活剛剛好，還完全不佔 M3 的記憶體。

Local LLM generation speed is roughly proportional to "memory bandwidth ÷ model size." The same 70B model runs at about 19 tok/s on M3 (819GB/s) but only ~9 tok/s on M1 (400GB/s) — that's the physical reason big models must live on M3. Conversely, a small 14B model has low bandwidth needs and actually flies on M1 (theoretically 40+ tok/s), perfect for the high-frequency grunt work of cleaning, classifying, and retrieval — without touching M3's memory at all.

再加上記憶體上限：M1 的 32GB（GPU 實際可用約 22–24GB）大概只能餵到 14B 級；32B 以上一定得在 96GB 的 M3。所以分工不是我硬湊的，是硬體先幫我畫好的線。

Plus the memory ceiling: M1's 32GB (about 22–24GB usable by the GPU) tops out around 14B; anything above 32B has to run on the 96GB M3. So the division of labor isn't something I forced — the hardware drew the line first.

1+1>2 實際長什麼樣

What 1+1>2 actually looks like

白天，M3 專心當開發機：Xcode、模擬器、32B 待命；M1 在背景爬蟲、做向量檢索——兩台各用各的記憶體頻寬，誰也不等誰。這就是加法的部分。

By day, the M3 is purely a dev machine — Xcode, simulators, the 32B on standby — while the M1 crawls and runs vector retrieval in the background. Each burns its own memory bandwidth; neither waits on the other. That's the additive part.

大於的部分在流程：要產技術週報或每日 code review 時，兩台接成一條「生成＋審查」的產線——單獨任何一台都做不到這件事。

The "greater than" part is the pipeline: for a weekly tech digest or daily code review, the two machines chain into a generate-and-review line — something neither could run alone.

順便解決了備份和「跑跑看」的安全問題

It also solves backup and the "just run it" safety problem

上面拓撲圖的後兩條流，就是 M1 的另外兩份工作：備份（每晚收 git 鏡像、當 Time Machine 目的地）和風險沙盒——AI 產出、要跑過才知道對不對的 code，一律丟到 M1 的隔離帳號執行。最壞情況是重灌 M1；主力機的環境和資產毫髮無傷。一句話原則：風險跑 M1，資產放 M3。

The last two flows in the topology above are M1's other two jobs: backup (nightly git mirror + Time Machine target) and risk sandbox — AI-generated code that can only be judged by running it goes to an isolated account on M1. Worst case, M1 gets reinstalled; the main machine's environment and assets stay untouched. One-line rule: run risk on M1, keep assets on M3.

值得考慮：需要在本地跑大型 AI 模型（吃大容量統一記憶體與高頻寬）、重度 Xcode／多模擬器／大型專案編譯、專業影音剪輯、或想要一台能撐好幾年的桌機主力。

Worth considering if: you run large AI models locally (hungry for high-capacity unified memory and bandwidth), do heavy Xcode / many simulators / big-project builds, professional video work, or want one desktop main machine that lasts years.

可以先緩緩：如果你手上已經是 M2 Ultra 且 192GB 記憶體夠用，這代的頻寬和神經網路引擎幾乎沒動，跳代的邊際效益不大；只是文書、瀏覽、輕度開發的話，Max 版本或更入門的機種其實更划算。買最貴的不等於買最對的——先看自己卡在哪一格，再決定要不要為那格付錢。

Can wait if: you're on an M2 Ultra and 192GB is enough — bandwidth and Neural Engine barely moved, so the marginal gain of jumping a gen is small. For docs, browsing, and light dev, a Max model or something more entry-level is the better value. The most expensive isn't the most right — figure out which row you're actually stuck on, then decide whether to pay for that row.