1
00:00:00,000 --> 00:00:12,000
今天這期視頻我們再談一談AI小盒子。就是AMD的AIMAX 395-400、英偉達的DGX Spark。然後蘋果的M5 Max和Ultra。

2
00:00:12,000 --> 00:00:26,000
為什麼呢?因為最近Deepseek V4.1 Flash這個重磅模型發布了。這個模型可以說是相當的牛逼。經過我這兩天的實測,我一開始認為它可能說能夠勉強替代Deepseek V4 Pro。

3
00:00:26,000 --> 00:00:30,460
但是深度使用之后我发现它是完全可以替代DeepSig V4 Pro

4
00:00:30,460 --> 00:00:34,320
并且有这个论坛有人发帖子

5
00:00:34,320 --> 00:00:36,720
就是比较权威的这些机构他们测评之后

6
00:00:36,720 --> 00:00:39,120
认为这个什么DeepSig V4.1 Flush

7
00:00:39,120 --> 00:00:43,200
在很多任务场景中都能够跟这个什么GPT-6 Astra

8
00:00:43,200 --> 00:00:46,320
跟这个Cloud 5.1这样的顶级旗舰模型

9
00:00:46,320 --> 00:00:49,260
做到相当稍微落后的一点水平

10
00:00:49,260 --> 00:00:51,960
它的成本就要低上一两个档次

11
00:00:51,960 --> 00:00:55,940
那么这个模型它刚好就能够在小盒子上跑起来

12
00:00:55,940 --> 00:00:59,480
在这种情况下就不得不讲它了

13
00:00:59,480 --> 00:01:02,720
就把这个小盒子在这个提一下

14
00:01:02,720 --> 00:01:05,240
虽然我本地都是跑这个7.820GB

15
00:01:05,240 --> 00:01:07,840
我一直主张大家去买这个独立的显卡

16
00:01:07,840 --> 00:01:10,200
比如说4090 48G 7900 XTX

17
00:01:10,200 --> 00:01:12,860
或者说是什么RTX Pro 4500 5000 6000

18
00:01:12,860 --> 00:01:13,620
是吧

19
00:01:13,620 --> 00:01:16,260
因为他们能够在本地跑大模型的同时

20
00:01:16,260 --> 00:01:18,060
也能够跑这个什么MiliMax H3

21
00:01:18,060 --> 00:01:20,340
跑这个LTX 2.3这样的视频模型

22
00:01:20,340 --> 00:01:21,460
跑Confi UI生态

23
00:01:21,460 --> 00:01:23,260
生产力是非常不错的

24
00:01:23,260 --> 00:01:25,740
是吧

25
00:01:25,740 --> 00:01:27,140
但是毕竟他们是跑不了

26
00:01:27,140 --> 00:01:29,400
Deep Seek 威斯福拉奇这样的

27
00:01:29,400 --> 00:01:32,740
可以说是目前性价比最高的这个模型

28
00:01:32,740 --> 00:01:35,720
那么我们首先从这个AMD谈起

29
00:01:35,720 --> 00:01:37,220
就是AMD这家公司呢

30
00:01:37,220 --> 00:01:39,700
它的这个什么AI Max 395跟这个400

31
00:01:39,700 --> 00:01:41,520
它现在陆续的发布了

32
00:01:41,520 --> 00:01:44,140
395好像是128级的这个内存

33
00:01:44,140 --> 00:01:46,700
400好像是可以支持到192级了

34
00:01:46,700 --> 00:01:48,880
但是呢我就有点没搞清楚

35
00:01:48,880 --> 00:01:51,340
你说它为什么不搞256级版本的呢

36
00:01:51,340 --> 00:01:52,640
我还去问了一下AI

37
00:01:52,640 --> 00:01:54,440
就这家公司它骂得非常的脑残

38
00:01:54,440 --> 00:01:54,840
知道吗

39
00:01:54,840 --> 00:01:56,140
它在设计这个芯片的时候

40
00:01:56,140 --> 00:01:58,780
它的硬件规格就只能跑到192G

41
00:01:58,780 --> 00:02:00,520
就支持不到256G

42
00:02:00,520 --> 00:02:03,140
你说这个硬件设计是得有多脑残

43
00:02:03,140 --> 00:02:03,900
才能够想得出来

44
00:02:03,900 --> 00:02:07,320
而且它这个395

45
00:02:07,320 --> 00:02:09,220
在这个市场反响

46
00:02:09,220 --> 00:02:11,160
比较一般的情况下

47
00:02:11,160 --> 00:02:12,240
你这个400出来

48
00:02:12,240 --> 00:02:14,380
就算说你这个核心部分

49
00:02:14,380 --> 00:02:15,320
你不去改

50
00:02:15,320 --> 00:02:18,080
你起码把395的一些大的毛病

51
00:02:18,080 --> 00:02:18,500
给它修改

52
00:02:18,500 --> 00:02:20,460
比如说它两台小盒子之间

53
00:02:20,460 --> 00:02:21,680
没有高速的互联方案

54
00:02:21,680 --> 00:02:22,700
不像英伟达

55
00:02:22,700 --> 00:02:24,040
英伟达有这个高速的互联方案

56
00:02:24,040 --> 00:02:26,220
可以两台128G的盒子合到一起

57
00:02:26,220 --> 00:02:28,240
形成256G的统一

58
00:02:28,240 --> 00:02:29,640
显存尺 瑞存尺

59
00:02:29,640 --> 00:02:30,840
AMD不能

60
00:02:30,840 --> 00:02:33,320
那么如果说你买395的话

61
00:02:33,320 --> 00:02:35,420
你就只能跑什么轻吻3.8Flash

62
00:02:35,420 --> 00:02:37,100
或者轻吻3.820Gb

63
00:02:37,100 --> 00:02:39,560
这两个模型在小盒子上跑都是非常的慢

64
00:02:39,560 --> 00:02:41,880
就响应速度很慢

65
00:02:41,880 --> 00:02:43,620
这个3.8Flash

66
00:02:43,620 --> 00:02:45,180
它不要说本地的小盒子跑

67
00:02:45,180 --> 00:02:47,600
它就是你用什么RTX

68
00:02:47,600 --> 00:02:48,560
跑到5000

69
00:02:48,560 --> 00:02:50,160
6000来跑它也慢

70
00:02:50,160 --> 00:02:52,300
就这个模型本身有问题

71
00:02:52,300 --> 00:02:56,120
然后这个轻微3.827币

72
00:02:56,120 --> 00:02:57,520
由于是稠密模型

73
00:02:57,520 --> 00:03:01,060
它同时激活这个270亿赛数

74
00:03:01,060 --> 00:03:03,700
比跑里这个什么DeepSick威斯福拉西的时候

75
00:03:03,700 --> 00:03:05,820
激活13币

76
00:03:05,820 --> 00:03:08,300
这个算力负担更大

77
00:03:08,300 --> 00:03:11,280
对于算力跟带宽的要求就更高

78
00:03:11,280 --> 00:03:13,520
那么IMD的小盒子跑起来

79
00:03:13,520 --> 00:03:16,740
就算是你把这个什么SG浪给它部署到位了

80
00:03:16,740 --> 00:03:19,540
它这个体验也是不怎么样的

81
00:03:19,540 --> 00:03:21,380
就你加入什么DeepFush2呀

82
00:03:21,380 --> 00:03:23,500
投机尖码是吧

83
00:03:23,500 --> 00:03:25,160
它的响应其实也都一般

84
00:03:25,160 --> 00:03:31,040
如果说你升级到这个什么AI Max 400

85
00:03:31,040 --> 00:03:32,740
能够买这个新的版本

86
00:03:32,740 --> 00:03:34,340
买这个190二级的路线版本

87
00:03:34,340 --> 00:03:35,400
那里也很尴尬

88
00:03:35,400 --> 00:03:38,200
因为这个Deep Seek V4 Flash

89
00:03:38,200 --> 00:03:39,240
它这个家族

90
00:03:39,240 --> 00:03:42,580
它是这个什么NVFP4跟这个FP8的混合模型

91
00:03:42,580 --> 00:03:45,300
它这个模型本身的这个尺寸

92
00:03:45,300 --> 00:03:47,220
大概是140几左右

93
00:03:47,220 --> 00:03:48,120
140多几

94
00:03:48,120 --> 00:03:51,000
你放上去之后再扣除系统开销

95
00:03:51,000 --> 00:03:52,660
扣除这个框架开销

96
00:03:52,660 --> 00:03:55,680
那里剩下来的就只能跑

97
00:03:55,680 --> 00:03:57,820
单个绘画长生下文

98
00:03:57,820 --> 00:03:59,340
你想多开难度很高

99
00:03:59,340 --> 00:04:02,680
就你很难理解

100
00:04:02,680 --> 00:04:04,220
这家公司为什么会这么老残

101
00:04:04,220 --> 00:04:07,940
就他这么多年来被这个什么

102
00:04:07,940 --> 00:04:09,760
英伟达吊打是有原因的

103
00:04:09,760 --> 00:04:10,980
就是他设计产品的时候

104
00:04:10,980 --> 00:04:12,560
非常的粗暴

105
00:04:12,560 --> 00:04:13,320
非常的粗放

106
00:04:13,320 --> 00:04:15,020
完全不考虑市场需求

107
00:04:15,020 --> 00:04:16,660
可能他们的设计师

108
00:04:16,660 --> 00:04:18,080
也用英伟达的显卡

109
00:04:18,080 --> 00:04:19,220
他可能不用自家的显卡

110
00:04:19,220 --> 00:04:23,140
你比如说他跟什么英塔去搞这个什么

111
00:04:23,140 --> 00:04:25,040
芯片竞争的时候

112
00:04:25,040 --> 00:04:27,660
我们都认为他渗出市场之后

113
00:04:27,660 --> 00:04:28,640
会回馈我们玩家

114
00:04:28,640 --> 00:04:30,480
把这个什么CPU的价格做下来

115
00:04:30,480 --> 00:04:31,220
结果并不是

116
00:04:31,220 --> 00:04:32,580
他非常的邪恶

117
00:04:32,580 --> 00:04:34,700
跟英塔长期联手

118
00:04:34,700 --> 00:04:35,480
是吧

119
00:04:35,480 --> 00:04:37,660
推高这个CPU的这个入门价格

120
00:04:37,660 --> 00:04:39,940
我们一点便宜都没占到

121
00:04:39,940 --> 00:04:41,880
再比如说他搞显卡的时候

122
00:04:41,880 --> 00:04:43,020
他非常的鸡贼

123
00:04:43,020 --> 00:04:44,920
就他知道自家的这个显卡

124
00:04:44,920 --> 00:04:47,080
性能生态没有这个英伟达好

125
00:04:47,080 --> 00:04:49,100
好比说英伟达的显卡卖5000是吧

126
00:04:49,100 --> 00:04:50,560
那我扣除生态是吧

127
00:04:50,560 --> 00:04:51,540
我就卖4000

128
00:04:51,540 --> 00:04:53,920
你虽然看起来

129
00:04:53,920 --> 00:04:55,180
你买他的卡同样的贵格

130
00:04:55,180 --> 00:04:56,140
你省了1000块钱

131
00:04:56,140 --> 00:04:57,140
但是这1000块钱

132
00:04:57,140 --> 00:04:58,720
你付出的这个体验差距

133
00:04:58,720 --> 00:05:01,100
这就是你支付的价格

134
00:05:01,100 --> 00:05:03,120
你实际上一毛钱便宜都没赚到

135
00:05:03,120 --> 00:05:05,280
相反你还有可能亏的更多

136
00:05:05,280 --> 00:05:06,460
给AMD做了小白鼠

137
00:05:06,460 --> 00:05:08,120
这么多年来

138
00:05:08,120 --> 00:05:09,440
一些游戏行业的创新

139
00:05:09,440 --> 00:05:11,460
但并不是每一项游戏创新都是好的

140
00:05:11,460 --> 00:05:13,300
基本上都是英伟达搞的

141
00:05:13,300 --> 00:05:15,100
比如说这个什么光追

142
00:05:15,100 --> 00:05:17,640
再比如什么这个插针

143
00:05:17,640 --> 00:05:19,840
就是DLSS

144
00:05:19,840 --> 00:05:21,060
就是插针的AI插针

145
00:05:21,060 --> 00:05:23,920
AMD搞了个叫什么FSR

146
00:05:23,920 --> 00:05:24,200
是吧

147
00:05:24,200 --> 00:05:26,460
就是永远是老黄做什么作业

148
00:05:26,460 --> 00:05:27,580
他在后面抄什么作业

149
00:05:27,580 --> 00:05:29,180
关键每次抄的都不彻底

150
00:05:29,180 --> 00:05:29,840
抄的都不好

151
00:05:29,840 --> 00:05:33,500
然后又不愿意真正给玩家上力

152
00:05:33,500 --> 00:05:34,620
就非常的鸡贼

153
00:05:34,620 --> 00:05:36,280
比如说前一阵子

154
00:05:36,280 --> 00:05:37,860
我这个推荐他的显卡

155
00:05:37,860 --> 00:05:39,060
就是什么7900XDX

156
00:05:39,060 --> 00:05:40,800
AI Pro R9700是吧

157
00:05:40,800 --> 00:05:42,240
那你这个时候

158
00:05:42,240 --> 00:05:44,220
你不应该多推出一点类似的产品吗

159
00:05:44,220 --> 00:05:44,780
是不是

160
00:05:44,780 --> 00:05:46,400
就市场中像我们这些屌丝

161
00:05:46,400 --> 00:05:49,980
都被这个老黄的刀法割得有点疼了

162
00:05:49,980 --> 00:05:51,900
你这个时候推出一些好的产品

163
00:05:51,900 --> 00:05:55,120
最起码说你能跟什么4080S30RG

164
00:05:55,120 --> 00:05:57,780
跟什么409048G摩托卡给干一干

165
00:05:57,780 --> 00:06:00,420
你都不要去比什么RTX Pro5600了

166
00:06:00,420 --> 00:06:01,480
你推出这样的卡

167
00:06:01,480 --> 00:06:02,620
卖的便宜一点

168
00:06:02,620 --> 00:06:04,880
让利下玩家的生态不就起来了吗

169
00:06:04,880 --> 00:06:05,820
并不会

170
00:06:05,820 --> 00:06:08,060
他只要出新的产品

171
00:06:08,060 --> 00:06:08,860
他就贵得很

172
00:06:08,860 --> 00:06:10,460
价格就对标老黄

173
00:06:10,460 --> 00:06:11,480
他给你让的利

174
00:06:11,480 --> 00:06:13,460
永远是你在生态上面吃的亏

175
00:06:13,460 --> 00:06:15,000
就是你吃多少亏

176
00:06:15,000 --> 00:06:15,720
他给你让多少利

177
00:06:15,720 --> 00:06:18,340
就它的小盒子你就不要碰了

178
00:06:18,340 --> 00:06:20,140
但如果说你现在就已经买了

179
00:06:20,140 --> 00:06:21,980
我就建议你认真去研究一下

180
00:06:21,980 --> 00:06:22,860
这个SG浪

181
00:06:22,860 --> 00:06:23,960
研究一下这个D-Flash

182
00:06:23,960 --> 00:06:24,920
D-Spark这些技术

183
00:06:24,920 --> 00:06:26,360
它能够缓解你的痛苦

184
00:06:26,360 --> 00:06:28,660
把千万3.820GB给剖好就行了

185
00:06:28,660 --> 00:06:30,160
其他的就不要多想了

186
00:06:30,160 --> 00:06:32,160
你也不要买它的盒子

187
00:06:32,160 --> 00:06:33,480
来跑这个Dipsick V-Flash

188
00:06:33,480 --> 00:06:34,700
因为那样子很傻

189
00:06:34,700 --> 00:06:39,160
下面就说这个英伟达的小盒子

190
00:06:39,160 --> 00:06:42,180
DGX Spark两台能够跑这个V-Flash

191
00:06:42,180 --> 00:06:44,620
这个是经过社区的多轮实测

192
00:06:44,620 --> 00:06:45,660
然后我们这个论坛

193
00:06:45,660 --> 00:06:48,140
也有不少帖子是实测跑起来的

194
00:06:48,140 --> 00:06:49,880
体验是非常不错的

195
00:06:49,880 --> 00:06:51,040
有真实的帖子

196
00:06:51,040 --> 00:06:53,600
你们可以到论坛去

197
00:06:53,600 --> 00:06:55,120
跟这些大神交流

198
00:06:55,120 --> 00:06:57,600
跟实际跑起来的人交流

199
00:06:57,600 --> 00:06:59,160
看看他们是怎么配置的

200
00:06:59,160 --> 00:07:01,380
这一块你就不用犹豫了

201
00:07:01,380 --> 00:07:03,640
如果说你刚好有这么七八万块钱

202
00:07:03,640 --> 00:07:04,120
预算在

203
00:07:04,120 --> 00:07:06,660
你想在本地跑这个什么

204
00:07:06,660 --> 00:07:08,780
Deepseek vs.Flug

205
00:07:08,780 --> 00:07:11,360
哪怕说你也想跑这个轻微3.8都可以

206
00:07:11,360 --> 00:07:13,980
甚至他跑轻微3.8Flug

207
00:07:13,980 --> 00:07:15,260
也比AMD的积极快

208
00:07:15,260 --> 00:07:19,100
如果说你还想兼容一下这个什么

209
00:07:19,100 --> 00:07:20,060
Confi UI

210
00:07:20,060 --> 00:07:22,520
它的这个Confi UI确实跑起来是慢了

211
00:07:22,520 --> 00:07:23,520
跑Milimax H3

212
00:07:23,520 --> 00:07:26,340
跑这个什么LTXR.3

213
00:07:26,340 --> 00:07:27,480
都慢的跟污涵一样

214
00:07:27,480 --> 00:07:29,300
但是你毕竟是买了两个盒子嘛

215
00:07:29,300 --> 00:07:30,140
你同时跑嘛

216
00:07:30,140 --> 00:07:31,840
它毕竟还是能跑的嘛

217
00:07:31,840 --> 00:07:35,380
就买老黄的东西就是没有什么压力

218
00:07:35,380 --> 00:07:36,440
是吧

219
00:07:36,440 --> 00:07:37,900
它两台盒子高速互联

220
00:07:37,900 --> 00:07:38,840
这个互联呢

221
00:07:38,840 --> 00:07:40,340
贷款达到了200几每秒

222
00:07:40,340 --> 00:07:41,240
是非常夸张的

223
00:07:41,240 --> 00:07:42,520
你可以对比一下嘛

224
00:07:42,520 --> 00:07:43,220
这个雷电5

225
00:07:43,220 --> 00:07:44,120
就苹果现在

226
00:07:44,120 --> 00:07:45,720
如果说零两代核子互联的话

227
00:07:45,720 --> 00:07:47,440
它是零零五零零五是八世纪

228
00:07:47,440 --> 00:07:50,280
这个八世纪里面并不是所有东西都可以拿出来的

229
00:07:50,280 --> 00:07:52,220
它只有六十几左右可以用来通信

230
00:07:52,220 --> 00:07:54,860
你就可以想一想一想

231
00:07:54,860 --> 00:07:56,500
老黄的这个互联方案也好

232
00:07:56,500 --> 00:07:57,340
还是这个小盒子

233
00:07:57,340 --> 00:07:58,900
它本身这个价值也好

234
00:07:58,900 --> 00:08:01,440
我认为它未来有很长一段时间的可玩性

235
00:08:11,239 --> 00:08:12,499
满大满算是吧

236
00:08:12,499 --> 00:08:13,959
你也就能够用上了

237
00:08:13,959 --> 00:08:15,699
那这个模型比起这个什么

238
00:08:15,699 --> 00:08:17,779
当下这个威斯夫拉奇牛逼在哪里

239
00:08:17,779 --> 00:08:20,659
第一是它的跑分应该会高出一截

240
00:08:20,659 --> 00:08:22,199
第二就是做任务的这个能力很强

241
00:08:22,199 --> 00:08:23,839
第三它是原生的多魔态

242
00:08:23,839 --> 00:08:28,299
这个可能说到时候又会有一些新的玩法出来

243
00:08:28,299 --> 00:08:30,139
况且就算是现在你用这个什么

244
00:08:30,139 --> 00:08:32,019
威斯夫拉奇的这个视觉版本

245
00:08:32,019 --> 00:08:35,579
它本身就能够非常好的干活了

246
00:08:35,579 --> 00:08:36,679
性价比就非常高了

247
00:08:36,679 --> 00:08:40,199
就这个核子是值得长期持有的

248
00:08:40,199 --> 00:08:41,559
未来有可能还会再涨价

249
00:08:41,559 --> 00:08:43,439
甚至我们可以关注一下

250
00:08:43,439 --> 00:08:45,279
搞不好老黄这个比较技贼的奸商

251
00:08:45,279 --> 00:08:47,239
他也有可能会发布下一代的产品

252
00:08:47,239 --> 00:08:51,619
但是那可能要很长的时间了

253
00:08:51,619 --> 00:08:53,399
因为他刚刚发布了这个什么

254
00:08:53,399 --> 00:08:54,419
RTX SPARC是吧

255
00:08:54,419 --> 00:08:54,959
这个笔记本

256
00:08:54,959 --> 00:08:56,519
也是引用了相同的架构

257
00:08:56,519 --> 00:08:58,639
我觉得怎么也得等明年了

258
00:08:58,639 --> 00:09:00,419
再上量这个时间

259
00:09:00,419 --> 00:09:03,039
我觉得如果说你有钱的话

260
00:09:03,039 --> 00:09:04,459
要马上生产

261
00:09:04,459 --> 00:09:05,639
那我觉得就不要等了

262
00:09:05,639 --> 00:09:07,279
就这个小盒子

263
00:09:07,279 --> 00:09:09,399
他未来的保值率一定是非常不错的

264
00:09:09,399 --> 00:09:11,039
当然我们要接受

265
00:09:11,039 --> 00:09:12,199
它是ARM的生态

266
00:09:12,199 --> 00:09:13,239
它这个什么

267
00:09:13,239 --> 00:09:15,299
没有奇差机用的Windows

268
00:09:15,299 --> 00:09:16,259
有了Windows之后

269
00:09:16,259 --> 00:09:18,039
它的生态也跟标准的X86Windows

270
00:09:18,039 --> 00:09:19,059
也是有差距的

271
00:09:19,059 --> 00:09:20,059
这是它明显不如

272
00:09:20,059 --> 00:09:22,659
AI Max 395 400的一个地方

273
00:09:22,659 --> 00:09:24,659
但是相信你买来是跑AI的

274
00:09:24,659 --> 00:09:24,899
是吧

275
00:09:24,899 --> 00:09:27,159
也不是用它来这个办公的

276
00:09:27,159 --> 00:09:28,719
这个体验差距

277
00:09:28,719 --> 00:09:29,979
我认为是可以接受的

278
00:09:29,979 --> 00:09:32,199
甚至即便是你办公

279
00:09:32,199 --> 00:09:33,839
你想办法把这个ARM Windows

280
00:09:33,839 --> 00:09:34,539
给它装上去

281
00:09:34,539 --> 00:09:37,379
它也体验也差不多哪去

282
00:09:37,379 --> 00:09:39,039
轻度的办公都能用

283
00:09:39,039 --> 00:09:40,859
那最后呢就是讲一讲苹果

284
00:09:40,859 --> 00:09:43,639
苹果呢就是说M5MAX本身就能打

285
00:09:43,639 --> 00:09:46,479
就它用来跑这个什么轻温3.8RG

286
00:09:46,479 --> 00:09:48,259
是这三个盒子中

287
00:09:48,259 --> 00:09:49,339
应该说表现最好的

288
00:09:49,339 --> 00:09:51,759
因为它是640G的带宽

289
00:09:51,759 --> 00:09:53,539
然后

290
00:09:53,539 --> 00:09:56,879
它如果是这个什么M5Ultra的话

291
00:09:56,879 --> 00:09:58,519
它是两个MAX并联在一起

292
00:09:58,519 --> 00:10:00,539
它芯片价的互联带宽高达几个T

293
00:10:00,539 --> 00:10:03,239
就这个总体的这个什么

294
00:10:03,239 --> 00:10:04,679
显存带宽好像是1.2T

295
00:10:04,679 --> 00:10:05,999
将近1.3T

296
00:10:05,999 --> 00:10:07,899
这个是非常能打的一个数字了

297
00:10:07,899 --> 00:10:11,719
就如果说你这个预算比较低

298
00:10:11,719 --> 00:10:14,159
那么你就买这个什么M5max

299
00:10:14,159 --> 00:10:15,959
买这个128G版本的

300
00:10:15,959 --> 00:10:18,599
用在本地跑这个什么轻微3.8G

301
00:10:18,599 --> 00:10:20,499
或者说是Flash

302
00:10:20,499 --> 00:10:23,779
它同时还能够有大量的这个绿村余量

303
00:10:23,779 --> 00:10:25,559
用来办公用来编程

304
00:10:25,559 --> 00:10:27,939
苹果的操作系统体验还是非常棒

305
00:10:27,939 --> 00:10:28,879
非常的稳定的

306
00:10:28,879 --> 00:10:31,219
你也可以把A进的就跑在这一个盒子上

307
00:10:31,219 --> 00:10:33,359
它的可玩性还是非常不错的

308
00:10:33,359 --> 00:10:36,879
但是必须要说它跑不了这个Konfi UI

309
00:10:36,879 --> 00:10:38,779
苹果是这三个盒子中

310
00:10:38,779 --> 00:10:39,979
破CONFUI最差的

311
00:10:39,979 --> 00:10:42,299
因为它的这个生态借容性不好

312
00:10:42,299 --> 00:10:43,919
而且苹果的所有芯片

313
00:10:43,919 --> 00:10:45,059
包括M5在内

314
00:10:45,059 --> 00:10:46,479
它这个叫什么

315
00:10:46,479 --> 00:10:49,839
FP3SAR的原始卷积算力是最低的

316
00:10:49,839 --> 00:10:50,979
这一块是天生残疾

317
00:10:50,979 --> 00:10:52,479
苹果设计的时候

318
00:10:52,479 --> 00:10:53,719
就没有为这个场景考虑

319
00:10:53,719 --> 00:10:57,179
就算是你勉强把这个CONFUI跑起来

320
00:10:57,179 --> 00:10:58,339
它速度也是慢的很

321
00:10:58,339 --> 00:10:59,079
当然了

322
00:10:59,079 --> 00:11:00,279
它画图是没问题的

323
00:11:00,279 --> 00:11:01,499
破音频是没问题的

324
00:11:01,499 --> 00:11:03,059
视频这一块就可以洗碎了

325
00:11:03,059 --> 00:11:04,199
你能接受这个缺点

326
00:11:04,199 --> 00:11:05,399
苹果的芯片

327
00:11:05,399 --> 00:11:07,959
苹果的小盒子性价比是非常高的

328
00:11:07,959 --> 00:11:10,619
然后如果说你要买Ultra 256G的

329
00:11:10,619 --> 00:11:12,879
它是肯定可以把DeepSeek和VS-Flug

330
00:11:12,879 --> 00:11:13,719
给跑起来的

331
00:11:13,719 --> 00:11:15,919
当然你需要知道它有一些硬性的缺点

332
00:11:15,919 --> 00:11:18,599
就是它本身没有什么FP4 FP8

333
00:11:18,599 --> 00:11:19,879
是吧

334
00:11:19,879 --> 00:11:21,099
它这个退理的时候

335
00:11:21,099 --> 00:11:22,519
只能返回BF16来退理

336
00:11:22,519 --> 00:11:24,659
当然这个也不是问题了

337
00:11:24,659 --> 00:11:26,239
因为毕竟这么大的带宽在这里

338
00:11:26,239 --> 00:11:26,559
是吧

339
00:11:26,559 --> 00:11:29,239
它还有一个你必须面对的现实

340
00:11:29,239 --> 00:11:29,939
就是它的生态

341
00:11:29,939 --> 00:11:33,179
远远不如AMD跟这个什么英伟达

342
00:11:33,179 --> 00:11:35,739
就是什么呢

343
00:11:35,739 --> 00:11:37,379
最主要的一个表现SG浪

344
00:11:37,379 --> 00:11:39,019
就SG浪

345
00:11:39,019 --> 00:11:41,479
你当下如果要火Agent的话

346
00:11:41,479 --> 00:11:41,879
我跟你讲

347
00:11:41,879 --> 00:11:42,939
你部署Lambati CPP

348
00:11:42,939 --> 00:11:45,079
部署这个什么VLM

349
00:11:45,079 --> 00:11:46,419
他们单会话可能说还可以

350
00:11:46,419 --> 00:11:47,979
一旦多会话长生下文

351
00:11:47,979 --> 00:11:49,779
你必须要上SG浪

352
00:11:49,779 --> 00:11:52,059
因为RATX缓存数

353
00:11:52,059 --> 00:11:53,019
我强调过多少次

354
00:11:53,019 --> 00:11:54,799
你们一定要去论坛

355
00:11:54,799 --> 00:11:55,839
去认真学习一下

356
00:11:55,839 --> 00:11:56,759
这些基础的概念

357
00:11:56,759 --> 00:11:58,839
RATX缓存数是非常重要的

358
00:11:58,839 --> 00:12:01,119
它能够对这个什么新的请求

359
00:12:01,119 --> 00:12:03,059
只做增量部分的这个prefill

360
00:12:03,059 --> 00:12:05,339
不用每次prefile一大段

361
00:12:05,339 --> 00:12:08,319
这个可以大大降低显卡的负载压力

362
00:12:08,319 --> 00:12:09,099
在小盒子上

363
00:12:09,099 --> 00:12:13,459
它降低的压力带来的意义就尤为重要

364
00:12:13,459 --> 00:12:16,539
但是我们也不用感到悲观

365
00:12:16,539 --> 00:12:17,939
因为我也去研究了一下

366
00:12:17,939 --> 00:12:19,859
社区现在已经有版本

367
00:12:19,859 --> 00:12:20,959
它不是主干版本

368
00:12:20,959 --> 00:12:22,479
不是SG浪官方的主干版本

369
00:12:22,479 --> 00:12:24,539
可以支持这个Radix换存数

370
00:12:24,539 --> 00:12:26,879
但是有副作用

371
00:12:26,879 --> 00:12:27,619
也有一些bug

372
00:12:27,619 --> 00:12:31,059
我相信现在即便它不能支持

373
00:12:31,059 --> 00:12:34,179
你用这个什么原始的这个prefill方式

374
00:12:34,179 --> 00:12:35,159
它也能玩

375
00:12:35,159 --> 00:12:36,039
DeepSick with Flaji

376
00:12:36,039 --> 00:12:37,239
肯定是能够跑起来的

377
00:12:37,239 --> 00:12:38,159
体验是不会差的

378
00:12:38,159 --> 00:12:40,419
你只要是买了M5 Ultra 256G

379
00:12:40,419 --> 00:12:41,379
是肯定够的

380
00:12:41,379 --> 00:12:43,759
未来的就是RedX缓存数

381
00:12:43,759 --> 00:12:45,919
一定会被官方给它加入到

382
00:12:45,919 --> 00:12:48,219
这个苹果的MLX生态中去的

383
00:12:48,219 --> 00:12:52,319
因为SG浪官方已经支持了MLX后端

384
00:12:52,319 --> 00:12:54,859
只不过暂时还没有把这个RedX缓存数

385
00:12:54,859 --> 00:12:56,159
给它加入进来

386
00:12:56,159 --> 00:12:57,559
这个是迟早的事情

387
00:12:57,559 --> 00:12:58,519
社区跑完之后

388
00:12:58,519 --> 00:12:59,759
肯定会有很多人

389
00:12:59,759 --> 00:13:02,359
向这个SG量的这个官方主板们提交请求

390
00:13:02,359 --> 00:13:04,719
一旦成熟了会被主板们合并

391
00:13:04,719 --> 00:13:08,499
到那个时候你跑这个什么Deepseek V4.1 Flush

392
00:13:08,499 --> 00:13:10,899
或者说那个时候可能都已经4.2 Flush了

393
00:13:10,899 --> 00:13:13,719
是吧那这体验是不会差的

394
00:13:13,719 --> 00:13:15,519
最后总结一下就是

395
00:13:15,519 --> 00:13:18,019
如果说你现在一定要买小盒子

396
00:13:18,019 --> 00:13:19,159
AMD的小盒子便宜

397
00:13:19,159 --> 00:13:20,319
但是我必须跟你讲

398
00:13:20,319 --> 00:13:21,759
这小盒子也是最垃圾的

399
00:13:21,759 --> 00:13:24,239
就是你不能对他有太高的期望

400
00:13:24,239 --> 00:13:26,019
他抛起来的体验是最差的

401
00:13:26,019 --> 00:13:27,879
然后英伟达的小盒子表现是最好的

402
00:13:27,879 --> 00:13:29,019
最有性价比的

403
00:13:29,019 --> 00:13:31,179
也是最有这个保值空间的

404
00:13:31,179 --> 00:13:33,239
将来甚至有可能有甚至空间

405
00:13:33,239 --> 00:13:35,919
如果说这个算力荒持续的话

406
00:13:35,919 --> 00:13:36,839
硬件荒持续的话

407
00:13:36,839 --> 00:13:38,199
对了

408
00:13:38,199 --> 00:13:39,879
英伟达的小盒子送你四个TB

409
00:13:39,879 --> 00:13:41,159
四个DP

410
00:13:41,159 --> 00:13:43,059
每个小盒子送你四个TB的存储空间

411
00:13:43,059 --> 00:13:44,619
这个当下来说也是一个价值

412
00:13:44,619 --> 00:13:46,579
就这两个小盒子

413
00:13:46,579 --> 00:13:48,499
能够为你在本地搭进你私有的一个Math

414
00:13:48,499 --> 00:13:49,599
体验

415
00:13:49,599 --> 00:13:51,259
这是附带的

416
00:13:51,259 --> 00:13:51,499
是吧

417
00:13:51,499 --> 00:13:53,079
它本身跑这个什么

418
00:13:53,079 --> 00:13:53,639
Comphe UI

419
00:13:53,639 --> 00:13:54,379
跑这个什么

420
00:13:54,379 --> 00:13:56,239
Deep Seek Vist Flash

421
00:13:56,239 --> 00:13:58,479
它体验都是非常不错的

422
00:13:58,479 --> 00:14:00,179
是吧

423
00:14:00,179 --> 00:14:00,659
画图啊

424
00:14:00,659 --> 00:14:01,119
音频啊

425
00:14:01,119 --> 00:14:02,259
当然视频可能要慢一点

426
00:14:02,259 --> 00:14:04,419
那个是小盒子的印商

427
00:14:04,419 --> 00:14:05,379
这是通用的印商

428
00:14:05,379 --> 00:14:08,959
然后如果是买苹果的话

429
00:14:08,959 --> 00:14:11,079
就是你像我一样

430
00:14:11,079 --> 00:14:12,719
平时在这个苹果的机器上面

431
00:14:12,719 --> 00:14:14,859
在上面比如说办公编程

432
00:14:14,859 --> 00:14:15,599
是吧

433
00:14:15,599 --> 00:14:16,879
同时你希望能跑个什么

434
00:14:16,879 --> 00:14:17,719
Deep Seek with Flush

435
00:14:17,719 --> 00:14:19,059
或者说跑个什么

436
00:14:19,059 --> 00:14:20,479
轻吻3.8 20GB

437
00:14:20,479 --> 00:14:22,639
那都可以的

438
00:14:22,639 --> 00:14:23,519
它都能做得到

439
00:14:23,519 --> 00:14:28,559
当然它8万多的价格

440
00:14:28,559 --> 00:14:30,459
也是这三个方案中最贵的

441
00:14:30,459 --> 00:14:30,959
单击啊

442
00:14:30,959 --> 00:14:31,859
单击就8万

443
00:14:31,859 --> 00:14:33,599
这个就是见仁见智的

444
00:14:33,599 --> 00:14:34,199
看你怎么想

445
00:14:34,199 --> 00:14:39,299
如果说你有这个相关疑问

446
00:14:39,299 --> 00:14:40,479
相关的话题想要讨论

447
00:14:40,479 --> 00:14:41,679
就欢迎到论坛中

448
00:14:41,679 --> 00:14:43,439
去看看大神们是怎么测的

449
00:14:43,439 --> 00:14:45,219
他们已经有的正式设备

450
00:14:45,219 --> 00:14:46,119
测试出来的数据

451
00:14:46,119 --> 00:14:48,639
这个比我在这里逼比较更有价值

452
00:14:48,639 --> 00:14:49,379
是吧

453
00:14:49,379 --> 00:14:51,779
我也这个让AI整理了我们论坛的帖子

454
00:14:51,779 --> 00:14:53,559
把他们这个官方的

455
00:14:53,559 --> 00:14:55,459
硬件规格数据都整理到了

456
00:14:55,459 --> 00:14:57,099
这个硬件的这个专业中去

457
00:14:57,099 --> 00:14:59,939
也把我们论坛中相关测试的帖子

458
00:14:59,939 --> 00:15:00,879
给附带进去了

459
00:15:00,879 --> 00:15:02,379
链接都加到这些专业中去了

460
00:15:02,379 --> 00:15:05,359
你们这个点点鼠标就能看得到

461
00:15:05,359 --> 00:15:05,699
是吧

462
00:15:05,699 --> 00:15:08,039
或者在手机上滑一滑屏幕就能看得到

463
00:15:08,039 --> 00:15:10,579
好吧

464
00:15:10,579 --> 00:15:11,399
本期视频就到这里

465
00:15:11,399 --> 00:15:12,099
我们下期再见