WEBVTT
Kind: captions
Language: zh-Hant

00:00:00.000 --> 00:00:01.580
相信大家對於九月這批

00:00:01.580 --> 00:00:03.161
新的 Mac mini 跟

00:00:03.161 --> 00:00:04.583
Mac Studio,

00:00:04.583 --> 00:00:06.480
到底要買哪一台,買什麼規格

00:00:06.700 --> 00:00:07.820
都有很大的疑問

00:00:08.040 --> 00:00:08.965
我也一樣

00:00:08.965 --> 00:00:11.742
我打開購買頁排了一個下午

00:00:11.742 --> 00:00:13.130
越排越不確定

00:00:13.350 --> 00:00:14.894
因為官網把記憶體

00:00:14.894 --> 00:00:16.051
GPU 核心數

00:00:16.051 --> 00:00:17.209
跟記憶體頻寬

00:00:17.209 --> 00:00:18.560
分成三個地方寫

00:00:18.780 --> 00:00:20.636
你看得到每一個數字

00:00:20.636 --> 00:00:23.110
但看不到它們是連在一起的

00:00:23.330 --> 00:00:24.889
我先給你其中一個結論

00:00:24.889 --> 00:00:26.760
你等一下就知道我在說什麼

00:00:26.980 --> 00:00:28.604
如果你跑代理的時候

00:00:28.604 --> 00:00:30.768
每一輪都要讀進很長的內容

00:00:30.768 --> 00:00:32.031
卻只回你一小段

00:00:32.031 --> 00:00:34.196
時間就可能主要花在處理輸

00:00:34.196 --> 00:00:36.180
入的運算上,不是花在吐字

00:00:36.400 --> 00:00:37.913
這一段主要看算力

00:00:37.913 --> 00:00:39.993
所以模型裝得下、吐字很快

00:00:39.993 --> 00:00:41.316
都不代表你按下

00:00:41.316 --> 00:00:42.640
送出之後不用等

00:00:42.860 --> 00:00:44.540
這支影片就是要幫助你

00:00:44.540 --> 00:00:46.220
選一台適合自己的規格

00:00:46.440 --> 00:00:47.874
先把買之前必須

00:00:47.874 --> 00:00:49.717
知道的那幾個規格,各

00:00:49.717 --> 00:00:51.560
自卡住你什麼,講清楚

00:00:51.780 --> 00:00:52.782
然後你拿自己

00:00:52.782 --> 00:00:53.950
的使用情境去對

00:00:54.170 --> 00:00:55.415
你就抓得到最適

00:00:55.415 --> 00:00:56.660
合你的那個規格

00:00:57.110 --> 00:00:58.323
購買 mac 之前

00:00:58.323 --> 00:01:00.230
有兩件事情一定要先理解

00:01:00.450 --> 00:01:02.460
第一個叫什麼。預填充

00:01:02.680 --> 00:01:04.405
它會把你輸入的整段

00:01:04.405 --> 00:01:06.130
提示詞一次全部算完

00:01:06.350 --> 00:01:07.773
這一段算完之後

00:01:07.773 --> 00:01:09.400
它會把每一層算出

00:01:09.400 --> 00:01:11.230
來的中間結果存下來

00:01:11.450 --> 00:01:13.370
就是所謂的 KV cache

00:01:13.590 --> 00:01:14.845
第二個是解碼

00:01:14.845 --> 00:01:17.145
它開始一個字一個字回你

00:01:17.145 --> 00:01:18.400
每次吐一個字

00:01:18.620 --> 00:01:19.530
它都要把整份

00:01:19.530 --> 00:01:21.350
KV cache 從頭掃一次

00:01:21.570 --> 00:01:22.887
這兩件事卡在

00:01:22.887 --> 00:01:24.424
完全不同的地方

00:01:24.424 --> 00:01:26.620
預填充是一次算一大片

00:01:26.840 --> 00:01:27.932
主要靠算力

00:01:27.932 --> 00:01:30.117
解碼是一個字一個字來

00:01:30.117 --> 00:01:32.520
每次都要把資料整份搬過

00:01:32.740 --> 00:01:34.900
靠的是記憶體頻寬

00:01:35.120 --> 00:01:38.011
那麼..預填充為什麼會慢呢?

00:01:38.011 --> 00:01:39.850
因為你輸入越長

00:01:40.070 --> 00:01:42.620
它要算的量不是等比例增加

00:01:42.620 --> 00:01:44.320
是跟長度的平方走

00:01:44.540 --> 00:01:46.316
你貼一段程式碼進去

00:01:46.316 --> 00:01:47.894
那一段要先被讀完

00:01:47.894 --> 00:01:49.670
你才會看到第一個字

00:01:49.890 --> 00:01:51.465
這裡有一個大家最

00:01:51.465 --> 00:01:52.842
容易搞錯的地方

00:01:52.842 --> 00:01:54.220
多輪對話的時候

00:01:54.440 --> 00:01:55.799
前面所有的歷史

00:01:55.799 --> 00:01:58.130
會不會每一輪都重新算一次

00:01:58.350 --> 00:01:59.417
答案是不一定

00:01:59.417 --> 00:02:01.017
如果前面那一段一個

00:02:01.017 --> 00:02:02.973
字都沒變,KV cache

00:02:02.973 --> 00:02:04.040
可以直接沿用

00:02:04.260 --> 00:02:05.425
mac只算你這

00:02:05.425 --> 00:02:06.590
一輪新加的東西

00:02:06.810 --> 00:02:08.917
那什麼時候會失效

00:02:08.917 --> 00:02:10.233
有三種情況

00:02:10.233 --> 00:02:12.340
一中就是全部重算

00:02:12.560 --> 00:02:13.772
第一種是什麼

00:02:13.772 --> 00:02:15.792
你開了一個新的子代理

00:02:15.792 --> 00:02:17.610
那是一個全新的對話

00:02:17.830 --> 00:02:19.760
前面的快取一個字都用不到

00:02:19.980 --> 00:02:20.991
第二種呢

00:02:20.991 --> 00:02:24.023
工具回傳的結果被插進中間

00:02:24.023 --> 00:02:26.550
插入點後面的全部失效

00:02:26.770 --> 00:02:27.783
第三種呢

00:02:27.783 --> 00:02:30.061
你把程式關掉再打開

00:02:30.061 --> 00:02:31.580
本地跑的時候

00:02:31.800 --> 00:02:33.850
通常不會把快取留到下一次

00:02:34.070 --> 00:02:35.510
這個差別有多大

00:02:35.510 --> 00:02:37.360
有人拿同一台mac

00:02:37.360 --> 00:02:38.800
同一個模型量過

00:02:39.020 --> 00:02:41.568
快取全中的時候,等兩點九秒

00:02:41.568 --> 00:02:43.478
快取沒中要全部重算

00:02:43.478 --> 00:02:44.540
等八十二秒

00:02:44.760 --> 00:02:47.313
同一台mac,同一個模型

00:02:47.313 --> 00:02:48.473
差二十八倍

00:02:48.473 --> 00:02:49.866
差別只在前面

00:02:49.866 --> 00:02:51.490
那段有沒有動過

00:02:51.710 --> 00:02:53.841
所以你如果覺得同一台mac

00:02:53.841 --> 00:02:55.480
有時候很快有時候很慢

00:02:55.700 --> 00:02:56.910
那不是你的錯覺

00:02:57.130 --> 00:02:58.875
還有一個例子更直接

00:02:58.875 --> 00:03:00.813
有人測同一段長提示詞

00:03:00.813 --> 00:03:02.170
量到兩點九七秒

00:03:02.390 --> 00:03:03.387
覺得很快

00:03:03.387 --> 00:03:06.128
然後他自己發現那是快取

00:03:06.128 --> 00:03:08.870
命中,關掉重測,變成十七秒

00:03:09.320 --> 00:03:10.220
先說一件事

00:03:10.220 --> 00:03:12.200
Mac 跟一般電腦差在哪

00:03:12.200 --> 00:03:14.180
差在它的記憶體是統一的

00:03:14.400 --> 00:03:16.650
CPU 跟 GPU 共用同一池

00:03:16.870 --> 00:03:18.795
所以規格表上那個記憶體

00:03:18.795 --> 00:03:20.720
數字,就是模型可以用的量

00:03:20.940 --> 00:03:22.225
這是 Mac 跑大

00:03:22.225 --> 00:03:23.510
模型真正的優勢

00:03:23.730 --> 00:03:25.528
在 Mac Studio 上

00:03:25.528 --> 00:03:26.836
你只要把記憶體選

00:03:26.836 --> 00:03:27.980
超過三十六 GB

00:03:28.200 --> 00:03:29.932
系統就會自動把你的

00:03:29.932 --> 00:03:32.050
GPU 換成貴一階的那顆

00:03:32.270 --> 00:03:34.370
頻寬從四百六十跳到六

00:03:34.370 --> 00:03:36.680
百一十四,價格多六百美元

00:03:36.900 --> 00:03:39.470
這不是加價選配,是你沒得選

00:03:39.690 --> 00:03:41.782
規格跟價格我整理成一張表

00:03:41.782 --> 00:03:43.700
放在畫面上,你暫停看就好

00:03:43.920 --> 00:03:44.410
我不念

00:03:44.630 --> 00:03:47.234
這一段如果幫你省掉在購買頁上

00:03:47.234 --> 00:03:49.280
來回猜的時間,順手按個讚

00:03:49.500 --> 00:03:50.945
讓這支片被更多正

00:03:50.945 --> 00:03:52.390
在選機器的人看到

00:03:52.840 --> 00:03:54.345
那把前面那兩件事套回

00:03:54.345 --> 00:03:55.700
Mac,會是什麼情況

00:03:55.920 --> 00:03:57.050
預填充吃算力

00:03:57.050 --> 00:03:58.179
而算力一直是

00:03:58.179 --> 00:04:00.250
Apple 這幾代的弱點

00:04:00.470 --> 00:04:02.687
M1 到 M4 這幾代

00:04:02.687 --> 00:04:05.459
只要你輸入的東西一多

00:04:05.459 --> 00:04:08.230
光是讀完就要等。要多久

00:04:08.450 --> 00:04:09.180
有數字

00:04:09.400 --> 00:04:11.396
有人拿 M3 Ultra 加

00:04:11.396 --> 00:04:13.574
五百一十二 GB 記憶體用了

00:04:13.574 --> 00:04:15.570
六個月,寫下他的實際體感

00:04:15.790 --> 00:04:17.715
簡單對話幾秒鐘就回

00:04:17.715 --> 00:04:19.640
這一段你不會有感覺

00:04:19.860 --> 00:04:22.331
請它改一篇七千字的文章

00:04:22.331 --> 00:04:23.230
等三十秒

00:04:23.450 --> 00:04:25.202
掛一個系統提示一萬

00:04:25.202 --> 00:04:28.121
六千字的 coding agent

00:04:28.121 --> 00:04:28.900
,等九十秒

00:04:29.120 --> 00:04:31.518
九十秒。在那九十秒裡

00:04:31.518 --> 00:04:33.650
畫面上什麼都沒有

00:04:33.870 --> 00:04:35.566
那台是 M3 Ultra

00:04:35.566 --> 00:04:36.923
記憶體頻寬已經是

00:04:36.923 --> 00:04:38.110
當時最高的一批

00:04:38.330 --> 00:04:40.216
頻寬那麼高為什麼

00:04:40.216 --> 00:04:41.867
沒有救到這一段

00:04:41.867 --> 00:04:44.460
因為這一段卡的不是頻寬

00:04:44.680 --> 00:04:45.621
那 M5 呢

00:04:45.621 --> 00:04:48.443
Apple 去年十一月發過

00:04:48.443 --> 00:04:51.030
一份研究,拿 M5 對比 M4

00:04:51.250 --> 00:04:53.622
等第一個字出現的時間

00:04:53.622 --> 00:04:55.520
快了三點三到四倍

00:04:55.520 --> 00:04:57.180
這是很大的改進

00:04:57.400 --> 00:04:58.777
而且方向對

00:04:58.777 --> 00:05:01.530
他們加強的正好是算力

00:05:01.750 --> 00:05:03.500
那這個改進是哪裡來的

00:05:03.720 --> 00:05:05.510
M5 這一代在每一個

00:05:05.510 --> 00:05:07.500
GPU 核心裡面,加了一

00:05:07.500 --> 00:05:09.290
個專門算矩陣的單元

00:05:09.510 --> 00:05:11.200
叫做神經加速器

00:05:11.420 --> 00:05:13.712
Apple 自己的說法是

00:05:13.712 --> 00:05:15.171
產生第一個字這

00:05:15.171 --> 00:05:16.630
件事被算力綁住

00:05:16.850 --> 00:05:18.647
而它完全吃得到神

00:05:18.647 --> 00:05:20.220
經加速器的好處

00:05:20.440 --> 00:05:21.414
換句話說

00:05:21.414 --> 00:05:23.850
你多花的錢買到的算力

00:05:24.070 --> 00:05:25.835
全部用在你按下送

00:05:25.835 --> 00:05:27.600
出之後那段等待上

00:05:27.820 --> 00:05:28.823
有多明顯

00:05:28.823 --> 00:05:30.830
有人拿同一台 M5

00:05:30.830 --> 00:05:33.590
同一個模型,只切這個開關

00:05:33.810 --> 00:05:35.759
讀取的速度從每秒

00:05:35.759 --> 00:05:37.708
兩百四十七跳到六

00:05:37.708 --> 00:05:39.900
百三十六,兩點五七倍

00:05:40.120 --> 00:05:42.645
吐字的速度從二十九點六

00:05:42.645 --> 00:05:45.170
到三十一,只有一點零五倍

00:05:45.390 --> 00:05:46.576
你注意一下這

00:05:46.576 --> 00:05:47.960
兩個數字的差距

00:05:48.180 --> 00:05:50.694
同一台機器同一個模型

00:05:50.694 --> 00:05:53.710
只換一個開關,一邊跳兩倍半

00:05:53.930 --> 00:05:55.613
另一邊幾乎不動

00:05:55.613 --> 00:05:57.297
這就是那兩件事

00:05:57.297 --> 00:05:58.980
真的分開的證據

00:05:59.200 --> 00:06:01.282
那買之前要注意

00:06:01.282 --> 00:06:02.770
什麼。有兩件

00:06:02.990 --> 00:06:06.619
第一,這個東西不是只有某一台有

00:06:06.619 --> 00:06:08.952
M5,M5 Pro,M5

00:06:08.952 --> 00:06:11.544
Max,M5 Ultra

00:06:11.544 --> 00:06:12.840
連 M6 都有

00:06:13.060 --> 00:06:15.255
所以它不是選型的差別

00:06:15.255 --> 00:06:17.230
你買哪一台都吃得到

00:06:17.450 --> 00:06:19.393
第二,有這個硬體是

00:06:19.393 --> 00:06:21.822
不是就自動用得到。不是

00:06:21.822 --> 00:06:23.280
要 macOS

00:06:23.280 --> 00:06:24.980
二十六點二以上

00:06:25.200 --> 00:06:27.530
而且軟體要支援才會啟用

00:06:27.750 --> 00:06:28.937
但要注意這是

00:06:28.937 --> 00:06:31.113
Apple 拿自己比自己

00:06:31.113 --> 00:06:32.300
它沒有回答跟

00:06:32.300 --> 00:06:34.080
NVIDIA 差多少

00:06:34.300 --> 00:06:35.995
那個差距目前只

00:06:35.995 --> 00:06:37.689
有上一代的數字

00:06:37.689 --> 00:06:40.110
同一個模型同一段輸入

00:06:40.330 --> 00:06:42.918
M3 Ultra 要五點

00:06:42.918 --> 00:06:45.248
五七秒,NVIDIA 

00:06:45.248 --> 00:06:47.836
那台小機器一點四七秒

00:06:47.836 --> 00:06:49.130
差三點八倍

00:06:49.350 --> 00:06:51.371
那 Mac Studio

00:06:51.371 --> 00:06:52.786
上面那顆最大的

00:06:52.786 --> 00:06:54.403
M5 Ultra 呢

00:06:54.403 --> 00:06:55.615
它九月二十二

00:06:55.615 --> 00:06:57.030
號才第一次出貨

00:06:57.250 --> 00:06:58.092
在那之前

00:06:58.092 --> 00:07:00.197
這顆晶片沒有人拿到過

00:07:00.197 --> 00:07:01.250
能追上多少

00:07:01.470 --> 00:07:02.393
要等出貨之後

00:07:02.393 --> 00:07:03.470
才有人測得出來

00:07:03.920 --> 00:07:05.371
那另外一半呢

00:07:05.371 --> 00:07:06.580
解碼跟頻寬

00:07:06.800 --> 00:07:09.084
解碼是一個字一個字產生

00:07:09.084 --> 00:07:10.330
每產生一個字

00:07:10.550 --> 00:07:12.482
mac都要把整份模型

00:07:12.482 --> 00:07:14.800
權重從記憶體讀進晶片一次

00:07:15.020 --> 00:07:16.918
所以它被什麼綁住

00:07:16.918 --> 00:07:18.105
記憶體頻寬

00:07:18.105 --> 00:07:19.054
頻寬翻倍

00:07:19.054 --> 00:07:21.190
吐字速度就接近翻倍

00:07:21.410 --> 00:07:23.415
所以跑分表上最常看到的那

00:07:23.415 --> 00:07:25.420
個數字,每秒幾個 token

00:07:25.640 --> 00:07:27.673
在 Mac 上為什麼會很好看

00:07:27.673 --> 00:07:28.690
就是這個原因

00:07:28.910 --> 00:07:31.322
M5 這一代吐字也有進步

00:07:31.322 --> 00:07:33.076
官方的數字是快百

00:07:33.076 --> 00:07:34.830
分之十九到二十七

00:07:35.050 --> 00:07:37.643
比起預填充那個三到四倍小

00:07:37.643 --> 00:07:40.020
很多,但你看跑分表的時候

00:07:40.240 --> 00:07:41.490
感覺是很直接的

00:07:41.710 --> 00:07:42.767
問題在這裡

00:07:42.767 --> 00:07:44.669
你日常真正在做的事

00:07:44.669 --> 00:07:46.360
決定你有沒有感覺

00:07:46.580 --> 00:07:48.250
你平常是哪一種

00:07:48.250 --> 00:07:49.919
如果你只是聊天

00:07:49.919 --> 00:07:51.350
問一句答一句

00:07:51.570 --> 00:07:53.465
輸入很短輸出很長

00:07:53.465 --> 00:07:55.598
那頻寬就是你的體感

00:07:55.598 --> 00:07:57.730
數字好看你就真的爽

00:07:57.950 --> 00:07:58.115
但

00:07:58.115 --> 00:08:01.080
如果你在做 agentic coding

00:08:01.080 --> 00:08:02.398
,或者同時開好幾個

00:08:02.398 --> 00:08:03.880
代理,情況完全反過來

00:08:04.100 --> 00:08:06.494
那種工作是什麼形狀

00:08:06.494 --> 00:08:09.686
讀進去一大坨,吐出來一小撮

00:08:09.686 --> 00:08:10.750
實測下來

00:08:10.970 --> 00:08:13.435
每一輪送進去二十萬 token

00:08:13.435 --> 00:08:15.020
,只吐出不到一千個字

00:08:15.240 --> 00:08:16.333
這種形狀下

00:08:16.333 --> 00:08:18.082
你的時間幾乎全部

00:08:18.082 --> 00:08:20.050
花在預填充,不在吐字

00:08:20.270 --> 00:08:21.657
頻寬再高也救不了

00:08:21.657 --> 00:08:23.390
這樣的體驗將會非常差

00:08:23.610 --> 00:08:25.380
你大概率會想要退貨的

00:08:25.600 --> 00:08:28.061
那多開幾個代理會怎麼樣

00:08:28.061 --> 00:08:29.850
你每開一個子代理

00:08:30.070 --> 00:08:32.617
就是一次全新的預填充

00:08:32.617 --> 00:08:34.400
開四個就是四次

00:08:34.620 --> 00:08:36.200
所以同一台mac

00:08:36.200 --> 00:08:38.175
你拿它聊天會覺得很快

00:08:38.175 --> 00:08:40.150
拿它跑代理會覺得很慢

00:08:40.370 --> 00:08:41.857
兩個感覺都是真的

00:08:41.857 --> 00:08:43.159
因為你在用的是

00:08:43.159 --> 00:08:44.460
兩個不同的規格

00:08:44.910 --> 00:08:46.524
如果你買 Mac 是為

00:08:46.524 --> 00:08:48.316
了在本地跑語言模型,有

00:08:48.316 --> 00:08:49.930
一類模型特別適合你

00:08:50.150 --> 00:08:51.360
叫做 MoE

00:08:51.580 --> 00:08:52.238
它的全名是

00:08:52.238 --> 00:08:53.422
Mixture of

00:08:53.422 --> 00:08:54.870
Experts,混合專家

00:08:55.090 --> 00:08:56.514
一般的模型裡面

00:08:56.514 --> 00:08:58.956
負責運算的那一層只有一組

00:08:58.956 --> 00:09:00.380
你每產生一個字

00:09:00.600 --> 00:09:02.200
這一組全部都要動一次

00:09:02.420 --> 00:09:05.138
MoE 把那一層拆成很多組

00:09:05.138 --> 00:09:07.630
常見的是幾十組到幾百組

00:09:07.850 --> 00:09:09.167
然後它多了一

00:09:09.167 --> 00:09:10.703
個很小的分派器

00:09:10.703 --> 00:09:12.020
每產生一個字

00:09:12.240 --> 00:09:13.920
分派器決定這次要叫醒

00:09:13.920 --> 00:09:15.600
哪幾組,通常只叫兩三組

00:09:15.820 --> 00:09:17.454
沒被叫到的那幾百組

00:09:17.454 --> 00:09:19.270
這個字完全不參與運算

00:09:19.490 --> 00:09:21.415
那沒被叫到的那些呢

00:09:21.415 --> 00:09:23.340
它們還是佔著記憶體

00:09:23.560 --> 00:09:25.810
沒被叫到不等於不用放進來

00:09:26.030 --> 00:09:27.811
那你怎麼看出一個

00:09:27.811 --> 00:09:30.259
模型是不是 MoE。看名字

00:09:30.259 --> 00:09:32.040
你會看到兩個數字

00:09:32.260 --> 00:09:35.380
像是三十 B,A 三 B 這種寫法

00:09:35.600 --> 00:09:37.257
前面那個是總共多大

00:09:37.257 --> 00:09:38.730
決定你裝不裝得下

00:09:38.950 --> 00:09:41.404
A 後面那個是每次啟用多少

00:09:41.404 --> 00:09:43.040
決定你跑起來多快

00:09:43.260 --> 00:09:44.716
那這件事為什麼剛

00:09:44.716 --> 00:09:45.990
好對 Mac 有用

00:09:46.210 --> 00:09:47.764
回想一下前面講的

00:09:47.764 --> 00:09:49.513
Mac 的強項是什麼

00:09:49.513 --> 00:09:50.290
記憶體大

00:09:50.510 --> 00:09:52.839
因為它是統一記憶體

00:09:52.839 --> 00:09:54.910
整池都能給模型用

00:09:55.130 --> 00:09:57.572
Mac 的弱項是算力比較差

00:09:57.572 --> 00:09:58.996
讀你貼進去那一

00:09:58.996 --> 00:10:00.420
大段的時候會等

00:10:00.640 --> 00:10:02.205
頻寬也不是最頂的

00:10:02.205 --> 00:10:03.770
吐字速度有天花板

00:10:03.990 --> 00:10:05.237
MoE 要的剛

00:10:05.237 --> 00:10:06.692
好是 Mac 有的

00:10:06.692 --> 00:10:07.731
它吃記憶體

00:10:07.731 --> 00:10:09.393
而記憶體是 Mac

00:10:09.393 --> 00:10:10.640
最不缺的東西

00:10:10.860 --> 00:10:12.134
MoE 省的剛

00:10:12.134 --> 00:10:13.620
好是 Mac 缺的

00:10:13.620 --> 00:10:15.106
每次只算兩三組

00:10:15.106 --> 00:10:16.380
算力需求變小

00:10:16.600 --> 00:10:18.450
要搬的權重也少很多

00:10:18.670 --> 00:10:20.087
所以同一台 Mac

00:10:20.087 --> 00:10:21.503
跑一個一千一百七

00:10:21.503 --> 00:10:22.920
十億參數的 MoE

00:10:23.140 --> 00:10:24.625
可以比跑一個七百

00:10:24.625 --> 00:10:26.110
億的一般模型還快

00:10:26.330 --> 00:10:27.959
這也是為什麼你會看到

00:10:27.959 --> 00:10:29.588
有人拿 Mac 跑那些名

00:10:29.588 --> 00:10:31.380
字聽起來很誇張的大模型

00:10:31.600 --> 00:10:32.810
而且真的跑得動

00:10:33.030 --> 00:10:35.321
你在 Mac 上挑模型的時候

00:10:35.321 --> 00:10:37.040
名字裡有沒有那個 A

00:10:37.260 --> 00:10:38.870
是你第一個要看的東西

00:10:39.090 --> 00:10:41.004
有一件事不要算過頭

00:10:41.004 --> 00:10:42.706
你不能拿啟用的那

00:10:42.706 --> 00:10:44.620
一小塊直接去除頻寬

00:10:44.840 --> 00:10:46.770
然後說速度會變幾倍

00:10:46.770 --> 00:10:49.128
中間還有挑選專家的成本

00:10:49.128 --> 00:10:50.200
那個不是零

00:10:50.650 --> 00:10:51.376
說到這裡

00:10:51.376 --> 00:10:53.556
規格對應到什麼已經清楚了

00:10:53.556 --> 00:10:55.190
那就分成三種人來看

00:10:55.410 --> 00:10:57.462
第一種,你的 AI 全部都用雲端的

00:10:57.462 --> 00:11:00.248
ChatGPT,Claude,Gemini

00:11:00.248 --> 00:11:01.274
,你不打算在自己

00:11:01.274 --> 00:11:02.300
mac上跑模型

00:11:02.520 --> 00:11:04.924
那這些頻寬跟核心數的差別

00:11:04.924 --> 00:11:06.527
對你幾乎沒有意義

00:11:06.527 --> 00:11:08.330
十六 GB,二十四 GB

00:11:08.550 --> 00:11:10.400
或者三十二 GB 都夠用

00:11:10.620 --> 00:11:11.587
那要花多少

00:11:11.587 --> 00:11:13.329
八百九十九到一千兩

00:11:13.329 --> 00:11:15.070
百九十九美元這一段

00:11:15.290 --> 00:11:16.583
M6 的 Mac mini

00:11:16.583 --> 00:11:17.100
就可以了

00:11:17.320 --> 00:11:19.619
真的要偶爾在本地跑一下

00:11:19.619 --> 00:11:22.128
三十二 GB 可以放 Qwen

00:11:22.128 --> 00:11:23.800
那個三十 B,A 三 B

00:11:24.020 --> 00:11:25.509
總共三百零五億

00:11:25.509 --> 00:11:27.210
每次啟用三十三億

00:11:27.210 --> 00:11:29.550
四位元之後十八點六 GB

00:11:29.770 --> 00:11:30.645
第二種呢

00:11:30.645 --> 00:11:31.738
你要剪影片

00:11:31.738 --> 00:11:33.050
同時想在本地

00:11:33.050 --> 00:11:34.580
跑一些語言模型

00:11:34.800 --> 00:11:36.322
這一種的甜蜜點在哪

00:11:36.322 --> 00:11:36.999
我認為是

00:11:36.999 --> 00:11:39.029
Mac Studio 六十四

00:11:39.029 --> 00:11:40.890
GB,三千七百九十九美元

00:11:41.110 --> 00:11:42.803
理由是六十四 GB 剛

00:11:42.803 --> 00:11:44.119
好讓三百二十億

00:11:44.119 --> 00:11:46.000
參數這一級跑得很舒服

00:11:46.220 --> 00:11:47.830
七百億也塞得進去

00:11:48.050 --> 00:11:49.599
而且六十四 GB 在

00:11:49.599 --> 00:11:51.730
Studio 上必然是四

00:11:51.730 --> 00:11:53.860
十核那顆,頻寬六百一十四

00:11:54.080 --> 00:11:55.898
你為了容量付的錢

00:11:55.898 --> 00:11:58.170
順便買到了最高的頻寬

00:11:58.390 --> 00:12:00.174
這一級你能跑哪些

00:12:00.174 --> 00:12:01.735
選擇一下子變多

00:12:01.735 --> 00:12:03.519
三十幾 B 的 MoE

00:12:03.519 --> 00:12:05.080
跑起來都很輕鬆

00:12:05.300 --> 00:12:07.511
如果你買到一百二十八 GB

00:12:07.511 --> 00:12:09.538
那 gpt-oss-120b

00:12:09.538 --> 00:12:10.643
是最好的例子

00:12:10.863 --> 00:12:13.207
一千一百七十億總參數

00:12:13.207 --> 00:12:14.849
六十二點八 GB

00:12:14.849 --> 00:12:17.193
但每次只啟用五十一億

00:12:17.413 --> 00:12:18.288
第三種呢

00:12:18.288 --> 00:12:20.696
你對本地大模型有很強的

00:12:20.696 --> 00:12:23.103
需求,你要跑那種真正大的

00:12:23.323 --> 00:12:25.910
那六十四 GB 夠不夠。不夠

00:12:25.910 --> 00:12:28.733
你要看的是二百五十六 GB

00:12:28.953 --> 00:12:30.687
甚至五百一十二 GB

00:12:30.687 --> 00:12:31.843
的統一記憶體

00:12:32.063 --> 00:12:34.781
那個級距只有 Mac Studio

00:12:34.781 --> 00:12:36.412
的 M5 Ultra 有

00:12:36.412 --> 00:12:38.042
二百五十六 GB 是一

00:12:38.042 --> 00:12:39.673
萬零七百九十九美元

00:12:39.893 --> 00:12:41.285
為什麼要那麼多

00:12:41.285 --> 00:12:43.074
因為真正大的那幾個

00:12:43.074 --> 00:12:44.863
開源模型是這個量級

00:12:45.083 --> 00:12:46.823
DeepSeek 那個

00:12:46.823 --> 00:12:48.911
V4 Flash,兩千八百四

00:12:48.911 --> 00:12:51.173
十億,四位元要一百五十五 GB

00:12:51.393 --> 00:12:52.332
更極端的

00:12:52.332 --> 00:12:54.211
Kimi K3 四位

00:12:54.211 --> 00:12:56.090
元要一點五一 TB

00:12:56.090 --> 00:12:58.203
開源不等於你跑得動

00:12:58.423 --> 00:13:01.006
你身邊如果有人也在這幾台之

00:13:01.006 --> 00:13:02.993
間猶豫,把這一段轉給他

00:13:03.213 --> 00:13:04.712
這三個級距是他要

00:13:04.712 --> 00:13:06.023
先搞清楚的東西

00:13:06.243 --> 00:13:07.768
五百一十二 GB 那個

00:13:07.768 --> 00:13:09.293
版本現在還下不了單

00:13:09.513 --> 00:13:12.083
官方說十月下旬才會上市

00:13:12.303 --> 00:13:14.307
這一段的價格是跳躍的

00:13:14.307 --> 00:13:15.309
不是線性的

00:13:15.309 --> 00:13:17.513
從六十四跳到二百五十六

00:13:17.733 --> 00:13:20.383
記憶體四倍,價格接近三倍

00:13:20.603 --> 00:13:22.461
所以你要先確定你真的

00:13:22.461 --> 00:13:24.133
需要跑那麼大的模型

00:13:24.353 --> 00:13:25.963
這筆錢才花得下去

00:13:26.413 --> 00:13:28.902
前面講的全部是文字模型

00:13:28.902 --> 00:13:30.713
那生圖跟生影片呢

00:13:30.933 --> 00:13:33.133
這兩件事在 Mac 上是弱的

00:13:33.133 --> 00:13:34.783
而且弱的地方不一樣

00:13:35.003 --> 00:13:35.778
先講影片

00:13:35.778 --> 00:13:37.909
因為想在本地跑 AI 的人

00:13:37.909 --> 00:13:39.653
很多是衝著這個來的

00:13:39.873 --> 00:13:41.534
而算力,正是前面講的

00:13:41.534 --> 00:13:43.563
Apple 這幾代的弱點

00:13:43.783 --> 00:13:45.597
而且不只是算力

00:13:45.597 --> 00:13:46.633
還有生態

00:13:46.853 --> 00:13:49.282
Mac 這邊沒有 TensorRT

00:13:49.282 --> 00:13:50.740
那種等級的推論優化

00:13:50.740 --> 00:13:52.521
,PyTorch 的編譯加

00:13:52.521 --> 00:13:54.303
速跟 Mac 的後端不相容

00:13:54.523 --> 00:13:56.501
ComfyUI 在 Mac

00:13:56.501 --> 00:13:59.198
上是能跑的,走的是 PyTorch

00:13:59.198 --> 00:14:00.636
的 Metal 後端

00:14:00.636 --> 00:14:02.074
但有些運算沒實作

00:14:02.074 --> 00:14:03.333
會掉回 CPU 跑

00:14:03.553 --> 00:14:05.563
在 CUDA 上是裝好就動

00:14:05.783 --> 00:14:07.665
M5 的算力是有改善

00:14:07.665 --> 00:14:09.964
同一個 app 生同一段五

00:14:09.964 --> 00:14:12.264
秒影片,筆電的 M5 Max

00:14:12.264 --> 00:14:14.982
四十核兩分鐘,M3 Ultra

00:14:14.982 --> 00:14:17.073
八十核要三分二十六秒

00:14:17.293 --> 00:14:19.264
但 ComfyUI 不是

00:14:19.264 --> 00:14:21.432
直接吃到這個算力,中間隔

00:14:21.432 --> 00:14:23.206
著 PyTorch 的

00:14:23.206 --> 00:14:25.571
Metal 後端,M5 新加的

00:14:25.571 --> 00:14:27.739
矩陣單元它才剛開始支援

00:14:27.739 --> 00:14:30.104
算力本來就是 Mac 的弱項

00:14:30.104 --> 00:14:31.680
跟 NVIDIA 比

00:14:31.680 --> 00:14:32.863
這一代還是輸

00:14:33.083 --> 00:14:34.730
如果這台 Mac 你

00:14:34.730 --> 00:14:36.788
本來就要買,它同時是你

00:14:36.788 --> 00:14:38.640
的日常機,你的剪接機

00:14:38.860 --> 00:14:40.150
你寫程式的機器

00:14:40.370 --> 00:14:41.638
那半夜掛著跑

00:14:41.638 --> 00:14:43.540
基本上只有電費的錢

00:14:43.760 --> 00:14:46.389
但是生成的影片速度非常慢

00:14:46.389 --> 00:14:48.800
並且還要考量到不是每一

00:14:48.800 --> 00:14:51.210
段影片都是可以用的片段

00:14:51.430 --> 00:14:54.000
真正能用的秒數,要再打個折

00:14:54.220 --> 00:14:56.227
所以如果你的主要目的是

00:14:56.227 --> 00:14:57.686
影片生成,那會建議

00:14:57.686 --> 00:14:59.510
還是考慮 NVIDIA

00:14:59.960 --> 00:15:01.257
最後有一個做法

00:15:01.257 --> 00:15:02.740
你不一定用得到它

00:15:02.960 --> 00:15:04.587
但我個人認為這解

00:15:04.587 --> 00:15:06.010
決方案非常有趣!

00:15:06.230 --> 00:15:06.898
有一個團隊

00:15:06.898 --> 00:15:08.234
把一台 NVIDIA 的

00:15:08.234 --> 00:15:09.703
DGX SPARK和一台

00:15:09.703 --> 00:15:11.440
Mac Studio 接在一起

00:15:11.660 --> 00:15:14.268
讓算力強的那台負責讀完

00:15:14.268 --> 00:15:17.350
整段,讓頻寬高的那台負責吐字

00:15:17.570 --> 00:15:20.110
他們自己測的數字是這樣

00:15:20.110 --> 00:15:22.420
同一個模型同一段輸入

00:15:22.640 --> 00:15:24.860
Mac 那台讀完花五點

00:15:24.860 --> 00:15:27.080
五七秒,吐字零點八五秒

00:15:27.080 --> 00:15:28.856
NVIDIA 那台

00:15:28.856 --> 00:15:30.410
讀完一點四七秒

00:15:30.630 --> 00:15:32.480
吐字二點八七秒

00:15:32.700 --> 00:15:34.505
你看這四個數字

00:15:34.505 --> 00:15:36.310
兩台各自贏一半

00:15:36.530 --> 00:15:38.112
兩台各贏一半

00:15:38.112 --> 00:15:41.014
接起來取兩邊快的那一段

00:15:41.014 --> 00:15:42.860
總共兩點三二秒

00:15:43.080 --> 00:15:44.450
快兩點八倍

00:15:44.670 --> 00:15:46.435
這剛好證明了前面那

00:15:46.435 --> 00:15:48.200
兩件事真的是分開的

00:15:48.420 --> 00:15:51.150
沒有一台機器兩邊都贏

00:15:51.370 --> 00:15:52.655
不過這是那個團

00:15:52.655 --> 00:15:53.940
隊自己測的數字

00:15:54.160 --> 00:15:56.205
而且現階段這個解決方案

00:15:56.205 --> 00:15:58.250
一定還有很大的改善空間

00:15:58.470 --> 00:15:59.235
即便如此

00:15:59.235 --> 00:16:01.337
我還是相信未來會有更多

00:16:01.337 --> 00:16:03.440
類似這樣的更好解決方案

00:16:03.660 --> 00:16:05.585
讓大家可以自己根據需

00:16:05.585 --> 00:16:07.510
求組合對應的硬體規格

00:16:07.960 --> 00:16:09.530
最後再收一次

00:16:09.530 --> 00:16:12.407
記憶體決定你裝不裝得下

00:16:12.407 --> 00:16:14.500
頻寬決定吐字多快

00:16:14.720 --> 00:16:16.671
算力決定你按下任

00:16:16.671 --> 00:16:18.867
務送出之後要等多久

00:16:18.867 --> 00:16:20.330
你卡在哪一件

00:16:20.550 --> 00:16:22.955
你要找的不是最強的那一台

00:16:22.955 --> 00:16:24.960
是你自己卡在哪一件事

00:16:25.180 --> 00:16:27.590
然後規格表上對應的那一行

00:16:27.810 --> 00:16:28.521
生圖的話

00:16:28.521 --> 00:16:30.477
我上面已經幫你們算過了

00:16:30.477 --> 00:16:31.900
專門買一台不划算

00:16:32.120 --> 00:16:33.370
順便用很划算

00:16:33.590 --> 00:16:35.078
跑文字模型的話

00:16:35.078 --> 00:16:36.565
裝得下是第一關

00:16:36.565 --> 00:16:37.840
頻寬是第二關

00:16:38.060 --> 00:16:38.914
相信這樣你自

00:16:38.914 --> 00:16:39.910
己就挑得出來了

00:16:40.130 --> 00:16:42.530
這批機器九月二十二號出貨

00:16:42.750 --> 00:16:44.148
到時候會有第一批

00:16:44.148 --> 00:16:45.720
真正的實測數字出來

00:16:45.940 --> 00:16:47.785
我會回來把這支影片裡所

00:16:47.785 --> 00:16:49.630
有還沒有數字的地方填滿

00:16:49.850 --> 00:16:51.657
你如果不想錯過那一集

00:16:51.657 --> 00:16:52.380
訂閱一下

00:16:52.600 --> 00:16:55.172
另外,你現在卡在哪兩個規格之間

00:16:55.172 --> 00:16:57.010
或者已經下訂了哪一台

00:16:57.230 --> 00:16:58.864
留言告訴我你的用途

00:16:58.864 --> 00:17:00.680
我看看能不能幫你判斷

00:17:00.900 --> 00:17:03.300
我是路可,讓我們一起踹踹看