【Ticker Symbol YOU】英伟达AI算力革命：从Blackwell到Ruben

1
00:00:00,000 --> 00:00:03,640
我很高兴与投资界分享这次独家专访

2
00:00:03,640 --> 00:00:06,320
大多数人认为英伟达是一家硬件公司

3
00:00:06,320 --> 00:00:08,900
专门生产训练大型AI模型的芯片

4
00:00:08,900 --> 00:00:10,599
但你即将了解一个内幕

5
00:00:10,599 --> 00:00:13,019
看看故事的另一面

6
00:00:13,019 --> 00:00:17,509
我有幸邀请到英伟达AI基础设施产品负责人乔·多尔

7
00:00:17,509 --> 00:00:21,548
过去四年里，乔一直在部署支撑

8
00:00:21,548 --> 00:00:24,260
地球上最强大的AI模型的硬件和软件

9
00:00:24,260 --> 00:00:28,420
他分享了关于AI未来发展的几个惊人见解

10
00:00:28,420 --> 00:00:31,719
但这只是我即将在直播中涵盖的众多技术之一

11
00:00:31,719 --> 00:00:33,899
几周后在GTC大会上

12
00:00:33,899 --> 00:00:36,920
GTC是英伟达的大型AI峰会

13
00:00:36,920 --> 00:00:40,869
展示机器人和自动驾驶领域的重大突破

14
00:00:40,869 --> 00:00:43,030
AI代理及其驱动芯片

15
00:00:43,030 --> 00:00:44,409
以及更多内容

16
00:00:44,409 --> 00:00:48,628
通过我的链接注册GTC免费在线会场

17
00:00:48,628 --> 00:00:52,048
即可赢取英伟达RTX 590显卡

18
00:00:52,048 --> 00:00:53,649
只需参加任意一场会

19
00:00:53,649 --> 00:00:56,990
截屏作为凭证并在会议后发送给我

20
00:00:56,990 --> 00:00:58,500
使用下方链接

21
00:00:58,500 --> 00:01:01,320
GTC应进入每位投资者的视野

22
00:01:01,320 --> 00:01:04,849
英伟达的AI推理生态系统同样值得关注

23
00:01:04,849 --> 00:01:05,888
你的时间很宝贵

24
00:01:05,888 --> 00:01:07,888
让我们直接进入正题

25
00:01:07,888 --> 00:01:09,528
能在这里与大家见面我非常开心

26
00:01:09,528 --> 00:01:10,429
感谢您的时间

27
00:01:10,429 --> 00:01:11,159
顺便说一句

28
00:01:11,159 --> 00:01:14,400
 Jensen在主旨演讲中提到了许多精彩内容

29
00:01:14,400 --> 00:01:17,159
他详细谈到的其中一个重点是

30
00:01:17,159 --> 00:01:23,129
英伟达实际上为维拉·鲁宾世代共同设计了六款芯片

31
00:01:23,319 --> 00:01:24,439
这需要深入解析

32
00:01:24,439 --> 00:01:26,379
我希望能与您一起探讨所有内容

33
00:01:26,379 --> 00:01:31,549
从GPU本身开始，逐步深入到机架级系统层面

34
00:01:31,549 --> 00:01:32,328
如果可以的话

35
00:01:32,328 --> 00:01:35,370
那我们就从鲁宾芯片开始吧

36
00:01:35,370 --> 00:01:38,560
黑尔威尔和鲁宾有何区别

37
00:01:38,560 --> 00:01:39,099
哦

38
00:01:39,099 --> 00:01:42,719
鲁宾在多个方面与黑尔威尔不同

39
00:01:42,719 --> 00:01:47,530
我们共同设计了六款芯片

40
00:01:47,530 --> 00:01:50,250
我们分析了数据中心需求

41
00:01:50,250 --> 00:01:52,209
然后反向推导

42
00:01:52,209 --> 00:01:56,349
在六款芯片中需要哪些配置才能实现最佳性能

43
00:01:56,349 --> 00:01:57,930
最佳能效

44
00:01:57,930 --> 00:01:59,060
最低成本

45
00:01:59,060 --> 00:01:59,579
没错

46
00:01:59,579 --> 00:02:03,260
这就是鲁宾的核心特点

47
00:02:03,260 --> 00:02:04,760
这种极致协同设计

48
00:02:04,760 --> 00:02:07,700
所有芯片共同制造

49
00:02:07,700 --> 00:02:08,949
共同设计

50
00:02:08,949 --> 00:02:11,788
协同工作以实现最佳性能

51
00:02:11,788 --> 00:02:17,039
当你提到查看数据中心需求时，这些是否由今天的AI模型驱动

52
00:02:17,039 --> 00:02:18,699
或者是什么在绝对驱动

53
00:02:18,699 --> 00:02:23,639
模型确实是推动计算需求的主要因素，尤其是Moe模型

54
00:02:23,639 --> 00:02:29,079
特别是混合专家模型，它们需要生成大量许多token

55
00:02:29,079 --> 00:02:29,639
因素

56
00:02:29,639 --> 00:02:32,938
更多token源于其推理过程

57
00:02:32,938 --> 00:02:35,780
同时模型规模也在持续扩大

58
00:02:35,780 --> 00:02:38,430
模型规模越大，智能水平越高

59
00:02:38,430 --> 00:02:39,830
来自推理能力

60
00:02:39,830 --> 00:02:44,330
这正在产生巨大的计算需求

61
00:02:44,330 --> 00:02:47,919
而Rubin正是为此设计的，明白

62
00:02:47,919 --> 00:02:51,400
请谈谈Blackwell与Rubin的区别

63
00:02:51,400 --> 00:02:55,050
具体到GPU的功耗和性能

64
00:02:55,050 --> 00:02:59,210
在推理工作负载的功耗和性能方面

65
00:02:59,210 --> 00:03:04,569
Rubin相比Blackwell性能提升高达十倍

66
00:03:04,569 --> 00:03:05,520
哇

67
00:03:05,520 --> 00:03:07,340
每瓦特性能提升十倍

68
00:03:07,340 --> 00:03:10,919
这意味着在固定延迟下

69
00:03:10,919 --> 00:03:11,659
可以看到

70
00:03:11,659 --> 00:03:15,210
通过我们在詹森演讲中展示的帕累托图表

71
00:03:15,210 --> 00:03:16,590
在特定延迟水平下

72
00:03:16,590 --> 00:03:17,090
非常

73
00:03:17,090 --> 00:03:17,590
你知道的

74
00:03:17,590 --> 00:03:19,069
高延迟但

75
00:03:19,069 --> 00:03:19,629
Uh

76
00:03:19,629 --> 00:03:21,649
这对模型用户很有利

77
00:03:21,649 --> 00:03:22,169
所以是的

78
00:03:22,169 --> 00:03:24,568
十倍性能提升覆盖整个机架规模

79
00:03:24,568 --> 00:03:26,609
是否在机架级架构中

80
00:03:26,609 --> 00:03:30,430
这里展示的是Blackwall Ultra代计算单元

81
00:03:30,430 --> 00:03:35,650
我可以展示组件及其分解情况

82
00:03:35,650 --> 00:03:37,689
我们有两个超级芯片

83
00:03:37,689 --> 00:03:38,770
两个超级芯片

84
00:03:38,770 --> 00:03:38,969
好的

85
00:03:38,969 --> 00:03:42,769
每个超级芯片包含两个Blackwall Alter GPU

86
00:03:42,769 --> 00:03:47,008
并在其中一个超级芯片上配备一个CPU

87
00:03:47,008 --> 00:03:48,188
然后两个超级芯片组合在一起

88
00:03:48,188 --> 00:03:51,110
形成四GPU加CPU配置

89
00:03:51,110 --> 00:03:56,899
同时还有Connect X8超级节点作为组件

90
00:03:56,899 --> 00:04:00,338
这将在后续讨论Vera Rubin时成为关键区别

91
00:04:00,338 --> 00:04:01,829
这些组件如何迁移

92
00:04:01,829 --> 00:04:03,528
嗯

93
00:04:03,528 --> 00:04:05,929
可以看到这是混合冷却系统

94
00:04:05,929 --> 00:04:06,348
好的

95
00:04:06,348 --> 00:04:07,528
这些是冷板

96
00:04:07,528 --> 00:04:10,569
对超级芯片及其组件进行液冷

97
00:04:10,569 --> 00:04:14,699
而在机架前半部分

98
00:04:14,699 --> 00:04:16,699
应说明这是风冷设计

99
00:04:16,699 --> 00:04:18,620
这些是我实际观察到的部分

100
00:04:18,620 --> 00:04:20,480
所有风扇顶部

101
00:04:20,480 --> 00:04:23,040
这里有八个风扇得到了它，所以八个风扇

102
00:04:23,040 --> 00:04:24,860
然后我们有一个蓝色区域

103
00:04:24,860 --> 00:04:25,800
DPU

104
00:04:25,800 --> 00:04:26,300
呃

105
00:04:26,300 --> 00:04:27,860
这也是这个托盘的一部分

106
00:04:27,860 --> 00:04:30,199
这是南北方向的交通

107
00:04:30,199 --> 00:04:30,879
呃

108
00:04:30,879 --> 00:04:31,860
连接存储设备

109
00:04:31,860 --> 00:04:36,310
将数据导入计算机架

110
00:04:36,310 --> 00:04:38,209
因此它会为

111
00:04:38,209 --> 00:04:40,110
GPU得到了它，所以是的

112
00:04:40,110 --> 00:04:42,120
DPU负责数据的输入输出

113
00:04:42,120 --> 00:04:43,360
然后所有处理过程

114
00:04:43,360 --> 00:04:44,759
所有神奇的操作都在

115
00:04:44,759 --> 00:04:46,790
超级芯片本身得到了它

116
00:04:46,790 --> 00:04:48,470
因此有两种网络流量

117
00:04:48,470 --> 00:04:51,329
南北方向是在同一机架内

118
00:04:51,329 --> 00:04:53,930
东西方向是连接多个机架

119
00:04:53,930 --> 00:04:55,290
这就是我们应该这样理解的吗

120
00:04:55,290 --> 00:04:56,730
这才是正确的理解方式

121
00:04:56,730 --> 00:04:57,110
是的

122
00:04:57,110 --> 00:04:57,490
是的

123
00:04:57,490 --> 00:05:01,360
嗯，我以为NVIDIA只是GPU设计商

124
00:05:01,360 --> 00:05:03,120
但Grace其实是CPU对吧

125
00:05:03,120 --> 00:05:05,519
那么CPU的作用是什么

126
00:05:05,519 --> 00:05:07,600
CPU处理大量管理任务

127
00:05:07,600 --> 00:05:08,180
所以

128
00:05:08,180 --> 00:05:08,920
例如

129
00:05:08,920 --> 00:05:10,209
当你进行

130
00:05:10,209 --> 00:05:10,889
呃

131
00:05:10,889 --> 00:05:12,410
你正在使用推理

132
00:05:12,410 --> 00:05:14,910
并希望你的模型能够

133
00:05:14,910 --> 00:05:15,250
呃

134
00:05:15,250 --> 00:05:16,298
为你生成一些代码

135
00:05:16,298 --> 00:05:17,519
并且想要制作

136
00:05:17,519 --> 00:05:18,678
可能混合一点应用

137
00:05:18,678 --> 00:05:23,589
一个需要运行的Python应用，CPU可以执行该应用

138
00:05:23,589 --> 00:05:27,290
GPU无法运行由模型生成的应用

139
00:05:28,370 --> 00:05:31,029
但它也处理其他类型的任务

140
00:05:31,029 --> 00:05:35,819
比如数据库分析等更适合CPU的功能

141
00:05:35,819 --> 00:05:38,319
能够加速这类任务

142
00:05:38,319 --> 00:05:39,079
哦，所以真的

143
00:05:39,079 --> 00:05:40,420
整个理念类似于

144
00:05:40,420 --> 00:05:41,579
让GPU负责

145
00:05:41,579 --> 00:05:42,480
它们最擅长的工作

146
00:05:42,480 --> 00:05:44,279
然后CPU处理其他事务

147
00:05:44,279 --> 00:05:47,519
显然CPU在某些方面远优于GPU

148
00:05:47,519 --> 00:05:51,680
因此可以将任务分配到适合的芯片上

149
00:05:51,680 --> 00:05:51,920
没错

150
00:05:51,920 --> 00:05:52,420
正确

151
00:05:52,420 --> 00:05:54,850
你还提到了一种叫做DPU的东西

152
00:05:54,850 --> 00:05:55,810
你能给我们详细讲解一下吗

153
00:05:55,810 --> 00:05:56,910
DSO DPU

154
00:05:56,910 --> 00:06:01,639
Bluefield DPU数据处理单元将处理南北向流量

155
00:06:01,639 --> 00:06:02,720
南北向流量

156
00:06:02,720 --> 00:06:03,019
是的

157
00:06:03,019 --> 00:06:04,839
当你连接到存储设备时

158
00:06:04,839 --> 00:06:06,639
它位于不同的机架上

159
00:06:06,649 --> 00:06:10,178
将会进行压缩和加密

160
00:06:10,658 --> 00:06:14,259
所有这些都将由Bluefield中的GPU管理

161
00:06:14,259 --> 00:06:14,999
Bluefield三号

162
00:06:14,999 --> 00:06:16,978
这个目标就是确保

163
00:06:16,978 --> 00:06:20,069
CPU和GPU不会执行这些任务

164
00:06:20,069 --> 00:06:20,810
卸载任务

165
00:06:20,810 --> 00:06:23,750
将所有功能从CPU和GPU卸载

166
00:06:23,750 --> 00:06:26,120
在硬件中加速这些功能

167
00:06:26,120 --> 00:06:26,639
嗯

168
00:06:26,639 --> 00:06:32,100
这样就能获得最快的数 据访问速度，为GPU提供数据，这很有道理

169
00:06:32,100 --> 00:06:32,399
明白了

170
00:06:32,399 --> 00:06:35,160
所以目前这三块是六颗芯片中的三颗

171
00:06:35,160 --> 00:06:35,899
CPU

172
00:06:35,899 --> 00:06:39,189
GPU和DPU以及Connex

173
00:06:39,189 --> 00:06:39,410
对的

174
00:06:39,410 --> 00:06:41,269
多说一点关于Connect X8

175
00:06:41,269 --> 00:06:43,750
这是东西向连接

176
00:06:43,750 --> 00:06:46,798
这是连接东西向的超级接口

177
00:06:46,798 --> 00:06:52,129
它还具备内联加密等功能处理东西向流量

178
00:06:52,129 --> 00:06:55,769
将连接机架间的GPU

179
00:06:55,769 --> 00:06:56,670
明白了

180
00:06:56,670 --> 00:06:58,490
所以我们有GPU

181
00:06:58,490 --> 00:06:59,509
CPU

182
00:06:59,509 --> 00:07:03,149
DPU和Connex芯片在这块板上

183
00:07:03,149 --> 00:07:04,430
另外两颗芯片在哪里

184
00:07:04,430 --> 00:07:08,259
EmuLink交换机是另一颗芯片

185
00:07:08,259 --> 00:07:09,019
嗯

186
00:07:09,019 --> 00:07:12,560
这个交换机托盘上有两颗

187
00:07:12,560 --> 00:07:13,220
嗯

188
00:07:13,220 --> 00:07:16,548
这是第五代NBLINK

189
00:07:16,548 --> 00:07:19,809
这些正在与NBLINK网络通信

190
00:07:19,809 --> 00:07:21,928
每秒1800吉比特

191
00:07:21,928 --> 00:07:23,108
1800一点

192
00:07:23,108 --> 00:07:24,699
每秒1.8太比特

193
00:07:24,699 --> 00:07:27,000
速度非常快

194
00:07:27,000 --> 00:07:34,189
这将成为Blackwell的核心神经系统

195
00:07:34,189 --> 00:07:35,470
GB300 NBL

196
00:07:35,470 --> 00:07:36,629
72明白了

197
00:07:36,629 --> 00:07:37,009
所以

198
00:07:37,009 --> 00:07:39,110
所以这是两个完全不同的托盘

199
00:07:39,110 --> 00:07:39,329
对的

200
00:07:39,329 --> 00:07:41,100
这就是计算托盘

201
00:07:41,100 --> 00:07:43,540
这就是处理数据时发生魔法的地方

202
00:07:43,540 --> 00:07:45,418
然后这是开关托盘

203
00:07:45,418 --> 00:07:50,639
我认为你之前提到的正是将所有GPU连接起来

204
00:07:50,639 --> 00:07:52,788
它将所有GPU连接在一起

205
00:07:52,788 --> 00:07:55,028
伊拉克内部有多个这样的交易

206
00:07:55,028 --> 00:07:55,408
是的

207
00:07:55,408 --> 00:07:55,968
呃

208
00:07:55,968 --> 00:07:58,088
所有GPU共有72台

209
00:07:58,088 --> 00:08:00,569
他们在伊拉克有72台

210
00:08:00,569 --> 00:08:02,490
实现全互联连接

211
00:08:02,490 --> 00:08:07,339
每个GPU必须能以全带宽与其他所有GPU通信

212
00:08:07,339 --> 00:08:10,439
这就是交换机实现的功能，所以

213
00:08:10,439 --> 00:08:11,439
每秒8太字节

214
00:08:11,439 --> 00:08:13,480
任何GPU都能与任何其他GPU通信

215
00:08:13,480 --> 00:08:15,660
这就是为什么叫计算织构的原因吗

216
00:08:15,660 --> 00:08:19,209
当我画这个网络图时应该没问题

217
00:08:19,209 --> 00:08:19,550
明白了

218
00:08:19,550 --> 00:08:20,589
所以是的

219
00:08:20,589 --> 00:08:21,970
他们称之为计算织构

220
00:08:21,970 --> 00:08:25,790
不仅因为连接所有GPU

221
00:08:25,790 --> 00:08:30,189
我们的NVLink交换芯片还包含一些计算功能

222
00:08:30,189 --> 00:08:33,710
我们称之为全归约或集体操作

223
00:08:33,710 --> 00:08:39,019
在训练时需要跨网络共享某些操作

224
00:08:39,019 --> 00:08:40,860
无需发送到所有GPU

225
00:08:40,860 --> 00:08:43,220
这些操作会在交换机内完成

226
00:08:43,220 --> 00:08:44,080
哦太棒了

227
00:08:44,080 --> 00:08:44,379
好的

228
00:08:44,379 --> 00:08:46,240
这个交换机不仅仅是连接设备

229
00:08:46,240 --> 00:08:49,659
它实际上也在执行部分计算

230
00:08:49,659 --> 00:08:51,318
这太厉害了

231
00:08:51,318 --> 00:08:51,958
好的

232
00:08:51,958 --> 00:08:54,278
我觉得我们已经介绍了五款芯片

233
00:08:54,278 --> 00:08:54,479
没错

234
00:08:54,479 --> 00:08:55,178
这样对吗

235
00:08:55,178 --> 00:08:56,039
完全正确

236
00:08:56,039 --> 00:08:57,720
第六款芯片是什么

237
00:08:57,720 --> 00:08:59,480
第六款是Spectrum Max

238
00:09:00,759 --> 00:09:02,240
我们可以看看这些机架吗

239
00:09:02,240 --> 00:09:03,068
对的

240
00:09:03,068 --> 00:09:04,769
我们去看看吧

241
00:09:05,879 --> 00:09:08,039
顶部有十个托盘

242
00:09:08,039 --> 00:09:09,399
这些是计算托盘

243
00:09:09,399 --> 00:09:11,110
九个网络托盘

244
00:09:11,110 --> 00:09:13,149
九个NVLink交换托盘

245
00:09:13,149 --> 00:09:13,610
应该说

246
00:09:13,610 --> 00:09:19,149
它们的任务是连接上方十台和下方八台GPU

247
00:09:19,549 --> 00:09:21,570
将计算托盘整合在一起对吧

248
00:09:21,570 --> 00:09:23,330
上面有什么

249
00:09:23,330 --> 00:09:26,059
这就是顶部机架

250
00:09:26,059 --> 00:09:29,340
这里有一台千兆位交换机用于遥测

251
00:09:29,340 --> 00:09:31,059
这只是某件事

252
00:09:31,059 --> 00:09:33,419
这是系统管理功能

253
00:09:33,419 --> 00:09:34,179
速度很低

254
00:09:34,179 --> 00:09:34,779
以太网

255
00:09:34,779 --> 00:09:36,139
这只是一个

256
00:09:36,139 --> 00:09:38,779
这只是用于自身管理的系统

257
00:09:38,779 --> 00:09:39,100
它并不

258
00:09:39,100 --> 00:09:42,299
它没有处理AI计算数据

259
00:09:42,299 --> 00:09:42,879
它在进行管理

260
00:09:42,879 --> 00:09:44,320
如果GPU宕机

261
00:09:44,320 --> 00:09:44,960
就像

262
00:09:44,960 --> 00:09:46,559
帮我理解遥测是什么意思

263
00:09:46,559 --> 00:09:47,960
以及这些遥测数据代表什么

264
00:09:47,960 --> 00:09:49,220
我只是在查看

265
00:09:49,220 --> 00:09:50,899
机架本身的各项功能

266
00:09:50,899 --> 00:09:52,000
我在查看

267
00:09:52,000 --> 00:09:52,940
运行时间

268
00:09:52,940 --> 00:09:54,360
我在查看健康状态和状态

269
00:09:54,360 --> 00:09:55,220
大概就是健康状态

270
00:09:55,220 --> 00:09:56,139
检查确认

271
00:09:56,139 --> 00:09:58,159
进行全面诊断

272
00:09:58,159 --> 00:10:02,629
你提到还有另一种机架会并排放置

273
00:10:02,629 --> 00:10:03,549
所以是的

274
00:10:03,549 --> 00:10:06,350
你会有一组计算机架

275
00:10:06,350 --> 00:10:08,049
三百台计算机架

276
00:10:08,049 --> 00:10:11,250
然后还会配备专用的光谱机架

277
00:10:11,250 --> 00:10:13,730
最大东西向网络交换机

278
00:10:13,730 --> 00:10:15,149
不过我们没有

279
00:10:15,149 --> 00:10:16,600
这里但呃

280
00:10:16,600 --> 00:10:18,559
不过功能应该是这样的

281
00:10:18,559 --> 00:10:19,740
我们称之为一个机柜

282
00:10:19,740 --> 00:10:20,659
你拥有

283
00:10:20,659 --> 00:10:23,860
可能八台三百机架

284
00:10:23,860 --> 00:10:27,919
然后会有几台配备光谱Max的交换机机架

285
00:10:27,919 --> 00:10:28,539
是的

286
00:10:28,539 --> 00:10:32,700
这很好地概述了当前的Blackwall系统

287
00:10:32,700 --> 00:10:34,359
我想了解如何

288
00:10:34,359 --> 00:10:37,639
从Blackwell到Reuben的变化

289
00:10:37,639 --> 00:10:38,078
明白了

290
00:10:38,078 --> 00:10:39,558
我们可以过去看看

291
00:10:39,558 --> 00:10:42,158
我们去看看这些托盘

292
00:10:42,419 --> 00:10:45,679
现在正在查看墙上的组件

293
00:10:45,679 --> 00:10:47,700
之前在计算托盘中讨论过

294
00:10:47,700 --> 00:10:48,879
Bluefield DPU

295
00:10:48,879 --> 00:10:49,879
Bluefield核心

296
00:10:49,879 --> 00:10:50,120
是的

297
00:10:50,120 --> 00:10:51,460
所以你可以看到它

298
00:10:51,460 --> 00:10:52,500
在墙上的位置

299
00:10:52,500 --> 00:10:55,759
这块板是模块系统的一部分

300
00:10:55,759 --> 00:10:59,090
可以插入或取出计算托盘以方便维护

301
00:10:59,090 --> 00:10:59,570
呃

302
00:10:59,570 --> 00:11:00,629
然后所有类似的情况

303
00:11:00,629 --> 00:11:03,149
连接X9位于中间

304
00:11:03,149 --> 00:11:03,690
呃

305
00:11:03,690 --> 00:11:06,950
这块板上有两个连接X9

306
00:11:06,950 --> 00:11:10,710
每个计算托盘总计八个

307
00:11:10,710 --> 00:11:13,750
因此每个GPU都会分配一个

308
00:11:13,750 --> 00:11:16,149
KINEX9的传输速率达6太比特每秒

309
00:11:16,149 --> 00:11:22,950
然后我们有光子集成封装光学模块

310
00:11:22,950 --> 00:11:24,610
这真是太酷了

311
00:11:24,610 --> 00:11:24,990
是啊

312
00:11:24,990 --> 00:11:31,220
这是什么？与其使用可插拔的光学模块

313
00:11:31,220 --> 00:11:34,000
它们直接集成在芯片上

314
00:11:34,000 --> 00:11:35,500
与芯片共封装

315
00:11:35,500 --> 00:11:38,328
这在能效方面有巨大提升

316
00:11:38,328 --> 00:11:43,428
可靠性也显著提高，主要体现在这两个方面

317
00:11:43,428 --> 00:11:46,629
以前我们使用光纤收发器

318
00:11:46,629 --> 00:11:48,789
光纤光学收发器

319
00:11:48,789 --> 00:11:51,259
两端连接光纤电缆

320
00:11:51,259 --> 00:11:53,899
这些收发器内置激光器

321
00:11:53,899 --> 00:11:54,379
没错

322
00:11:54,379 --> 00:11:56,139
需要供电对吧

323
00:11:56,139 --> 00:11:58,059
这就是要消除的部分

324
00:11:58,059 --> 00:12:01,320
通过将封装与芯片集成

325
00:12:01,320 --> 00:12:04,629
这对性能或功耗有何实际影响

326
00:12:04,629 --> 00:12:06,769
从性能角度来看

327
00:12:06,769 --> 00:12:08,049
性能保持不变

328
00:12:08,049 --> 00:12:08,450
是的

329
00:12:08,450 --> 00:12:09,809
但会带来

330
00:12:09,809 --> 00:12:13,850
功耗降低和可靠性提升

331
00:12:13,850 --> 00:12:17,210
因为可插拔激光器可能

332
00:12:17,210 --> 00:12:17,789
你知道的

333
00:12:17,789 --> 00:12:19,070
有时非常不可靠

334
00:12:19,070 --> 00:12:21,470
需要频繁更换

335
00:12:21,470 --> 00:12:23,269
但如果共封装在这里

336
00:12:23,269 --> 00:12:24,009
在

337
00:12:24,009 --> 00:12:24,919
在芯片上

338
00:12:24,919 --> 00:12:26,220
可靠性大幅提升

339
00:12:26,220 --> 00:12:26,759
比如

340
00:12:26,759 --> 00:12:28,720
可能提升十倍左右

341
00:12:28,720 --> 00:12:28,960
太棒了

342
00:12:28,960 --> 00:12:29,980
差异巨大

343
00:12:29,980 --> 00:12:32,429
这些组件在机架中的位置

344
00:12:32,429 --> 00:12:37,089
位于独立的交换托盘或交换服务器

345
00:12:37,089 --> 00:12:38,808
属于单独机架

346
00:12:38,808 --> 00:12:40,808
这就是独立机架

347
00:12:40,808 --> 00:12:43,340
与72系列分离

348
00:12:43,340 --> 00:12:46,059
这就是东西向流量交换机架

349
00:12:47,059 --> 00:12:47,580
太厉害了

350
00:12:47,580 --> 00:12:48,929
量子MAX

351
00:12:48,929 --> 00:12:49,330
嗯

352
00:12:49,330 --> 00:12:51,409
还有用于finband的组件

353
00:12:51,409 --> 00:12:54,009
这是以太网的替代方案

354
00:12:54,009 --> 00:12:57,610
还有量子infiniband的co包光模块

355
00:12:57,610 --> 00:12:59,090
所以这两款芯片功能等效

356
00:12:59,090 --> 00:13:00,690
一款用于spectrum x以太网

357
00:13:00,690 --> 00:13:02,070
另一款用于量子infiniband

358
00:13:02,070 --> 00:13:02,818
正确

359
00:13:02,818 --> 00:13:06,599
然后还有spectrum x以太网光子交换机

360
00:13:06,599 --> 00:13:10,629
因此该co包光模块芯片已集成其中

361
00:13:10,629 --> 00:13:12,730
在以太网光子交换机内

362
00:13:12,730 --> 00:13:14,070
这就是光子部分的组件

363
00:13:14,070 --> 00:13:16,250
co包光模块明白了

364
00:13:16,250 --> 00:13:18,470
但这些组件安装在侧挂单元

365
00:13:18,470 --> 00:13:19,549
这些组件安装到

366
00:13:19,549 --> 00:13:19,870
嗯

367
00:13:19,870 --> 00:13:20,570
交换机机架

368
00:13:20,570 --> 00:13:21,210
是的

369
00:13:21,210 --> 00:13:24,149
也理解了那个部分

370
00:13:24,149 --> 00:13:28,379
如果采用量子infiniband作为东西向流量协议

371
00:13:28,379 --> 00:13:31,340
则需使用infiniband作为侧挂单元

372
00:13:31,340 --> 00:13:32,460
所以这些是

373
00:13:32,460 --> 00:13:33,889
这些是等效方案

374
00:13:33,889 --> 00:13:35,110
一款finiband

375
00:13:35,110 --> 00:13:36,529
一款用于以太网

376
00:13:36,529 --> 00:13:37,570
正确明白了

377
00:13:37,570 --> 00:13:38,049
是的

378
00:13:38,049 --> 00:13:38,909
没错

379
00:13:38,909 --> 00:13:40,389
所以我们刚才讨论的内容

380
00:13:40,389 --> 00:13:43,960
可以说是当前数据中心的顶尖技术

381
00:13:43,960 --> 00:13:44,299
没错

382
00:13:44,299 --> 00:13:49,320
Blackwell Ultra目前是数据中心的顶尖产品

383
00:13:49,320 --> 00:13:51,820
然后Jensen推出了Vera Rubin

384
00:13:51,820 --> 00:13:53,399
我们提到的六款芯片

385
00:13:53,399 --> 00:13:55,340
我们讨论了Blackwell系列

386
00:13:55,340 --> 00:13:58,779
这是一个与之前完全不同的计算模块

387
00:13:58,779 --> 00:14:00,840
能否详细说明差异

388
00:14:00,840 --> 00:14:01,440
哦对

389
00:14:01,440 --> 00:14:01,980
差异很多

390
00:14:01,980 --> 00:14:03,120
所以嗯

391
00:14:03,120 --> 00:14:05,179
我们整体设计

392
00:14:05,179 --> 00:14:06,639
采用模块化设计

393
00:14:06,639 --> 00:14:07,139
明白了

394
00:14:07,139 --> 00:14:09,700
这意味着这里有插槽

395
00:14:09,700 --> 00:14:12,720
这些组件可轻松滑入滑出

396
00:14:12,720 --> 00:14:14,309
并锁定到位

397
00:14:14,309 --> 00:14:17,350
无需大量线缆连接

398
00:14:17,350 --> 00:14:21,840
处理模块间的所有连接

399
00:14:21,840 --> 00:14:25,000
同时线缆布局也已优化

400
00:14:25,000 --> 00:14:25,429
是的

401
00:14:25,429 --> 00:14:27,789
所以中间有一个管道装置

402
00:14:27,789 --> 00:14:31,590
它负责管理大量液体的分配

403
00:14:31,590 --> 00:14:34,070
总体来看在GB上

404
00:14:34,070 --> 00:14:34,470
三百

405
00:14:34,470 --> 00:14:36,070
有四十三根软管

406
00:14:36,070 --> 00:14:38,350
这里有一个风扇机柜

407
00:14:38,350 --> 00:14:39,960
因为它是混合冷却系统

408
00:14:39,960 --> 00:14:40,559
呃

409
00:14:40,559 --> 00:14:43,549
GB三百的下半部分是风扇冷却

410
00:14:43,549 --> 00:14:44,149
这个

411
00:14:44,149 --> 00:14:45,389
我们已经去除了它

412
00:14:45,389 --> 00:14:45,830
呃

413
00:14:45,830 --> 00:14:48,070
现在我们完全采用液体冷却

414
00:14:48,070 --> 00:14:50,850
八个风扇变为零个风扇

415
00:14:50,850 --> 00:14:52,190
零根软管

416
00:14:52,190 --> 00:14:55,429
同时还移除了大量电缆

417
00:14:55,429 --> 00:14:57,049
所以实现无缆化

418
00:14:57,049 --> 00:15:01,090
我正在努力拼凑

419
00:15:01,090 --> 00:15:01,769
我正在观察的内容

420
00:15:01,769 --> 00:15:03,970
这些位置应该是两个超级芯片所在

421
00:15:03,970 --> 00:15:05,250
这些就是超级芯片

422
00:15:05,250 --> 00:15:06,269
可以滑入滑出

423
00:15:06,269 --> 00:15:07,318
自动锁定到位

424
00:15:07,318 --> 00:15:09,658
所以这里有两块Ruben芯片

425
00:15:09,658 --> 00:15:11,879
上面安装了一个Vera

426
00:15:11,879 --> 00:15:12,918
所以啊

427
00:15:12,918 --> 00:15:16,429
另一个重要点是现在模块化设计

428
00:15:16,429 --> 00:15:18,549
所有机柜可滑入滑出

429
00:15:22,100 --> 00:15:24,299
组装这个结构并进行

430
00:15:24,299 --> 00:15:27,000
效率提升二十倍

431
00:15:27,000 --> 00:15:30,629
原本组装GB三百机柜需要两小时

432
00:15:30,629 --> 00:15:32,190
现在只需五分钟

433
00:15:32,190 --> 00:15:32,629
是的

434
00:15:32,629 --> 00:15:34,110
在这条特定路线

435
00:15:34,110 --> 00:15:35,830
这就是组装过程

436
00:15:35,830 --> 00:15:35,950
比如

437
00:15:35,950 --> 00:15:37,190
如果出现维护问题

438
00:15:37,190 --> 00:15:38,549
同时也便于维护

439
00:15:38,549 --> 00:15:38,809
没错

440
00:15:38,809 --> 00:15:41,750
操作速度大幅提升

441
00:15:41,750 --> 00:15:43,730
可维护性显著增强

442
00:15:43,730 --> 00:15:45,440
提升倍数相当可观

443
00:15:45,440 --> 00:15:45,840
不

444
00:15:45,840 --> 00:15:46,600
这完全合理

445
00:15:46,600 --> 00:15:48,340
如果不用这些线缆和软管

446
00:15:48,340 --> 00:15:50,129
只需快速拆装

447
00:15:50,129 --> 00:15:51,129
修复问题

448
00:15:51,129 --> 00:15:52,090
解决任何故障

449
00:15:52,090 --> 00:15:53,110
快速装回即可

450
00:15:53,110 --> 00:15:54,830
这种模块化设计就像这样

451
00:15:54,830 --> 00:15:56,830
我们稍后会讨论下面的其他组件

452
00:15:56,830 --> 00:15:59,788
所以两个超级芯片鲁本·维拉

453
00:15:59,788 --> 00:16:00,328
呃

454
00:16:00,328 --> 00:16:03,208
我们还有CX9连接X9

455
00:16:03,208 --> 00:16:08,620
这是该超级尼克的下一代产品，位于这些电路板和模块上

456
00:16:08,620 --> 00:16:13,480
以前它们连接在GB300超级芯片底部

457
00:16:13,480 --> 00:16:16,389
但现在它们有自己的模块，卡可以插拔

458
00:16:16,389 --> 00:16:19,669
现在可以单独维护不同组件

459
00:16:19,669 --> 00:16:20,578
是的

460
00:16:20,578 --> 00:16:25,298
新一代DPU的BlueField也是这里的模块

461
00:16:25,298 --> 00:16:26,418
可以插拔使用

462
00:16:26,418 --> 00:16:30,399
明白了，这不仅仅是关于性能

463
00:16:30,399 --> 00:16:32,980
还关乎更高的可用性对吧

464
00:16:32,980 --> 00:16:35,080
所以这是另一个乘数

465
00:16:35,080 --> 00:16:40,279
整个AI工厂的输出能力取决于可用时间我们称之为有效产出

466
00:16:40,279 --> 00:16:41,470
你想要的

467
00:16:41,470 --> 00:16:41,970
那个

468
00:16:41,970 --> 00:16:44,350
实际生成令牌的时间占比

469
00:16:44,350 --> 00:16:45,409
要最大化这个

470
00:16:45,409 --> 00:16:45,990
是的

471
00:16:45,990 --> 00:16:46,950
这很有道理

472
00:16:46,950 --> 00:16:48,009
好的

473
00:16:48,009 --> 00:16:50,549
这就是等效计算托盘

474
00:16:50,549 --> 00:16:53,389
还有一个等效交换机托盘

475
00:16:53,389 --> 00:16:53,769
对的

476
00:16:53,769 --> 00:16:54,629
没错

477
00:16:54,629 --> 00:16:56,309
看起来更加简洁流畅

478
00:16:56,309 --> 00:16:57,980
请带我看看这里的改动

479
00:16:57,980 --> 00:16:59,840
从改动来看

480
00:16:59,840 --> 00:17:00,419
嗯

481
00:17:00,419 --> 00:17:02,700
顶部有这些交换机

482
00:17:02,700 --> 00:17:04,150
百分之百液冷设计

483
00:17:04,150 --> 00:17:04,630
嗯

484
00:17:04,630 --> 00:17:05,710
这里有四个交换芯片

485
00:17:05,710 --> 00:17:06,789
这是Envy Link

486
00:17:06,789 --> 00:17:09,378
第六代重型架构

487
00:17:09,378 --> 00:17:11,898
速度是黑墙版本的两倍

488
00:17:11,898 --> 00:17:12,638
速度翻倍

489
00:17:12,638 --> 00:17:15,489
现在达到36太字节每秒

490
00:17:15,489 --> 00:17:18,869
这将显著提升我们的性能

491
00:17:18,869 --> 00:17:23,230
我之前提到的每瓦或每兆瓦、每吉瓦的十倍性能

492
00:17:23,230 --> 00:17:24,470
根据你的需求

493
00:17:25,549 --> 00:17:31,089
NBL链路速度提升是其中一部分贡献

494
00:17:31,089 --> 00:17:35,019
加上其他GPU特性我们稍后可以讨论

495
00:17:35,019 --> 00:17:35,660
还有其他

496
00:17:35,660 --> 00:17:40,210
所以黑墙机架的总GPU数量是否相同

497
00:17:40,210 --> 00:17:41,490
与鲁本机架相比

498
00:17:41,490 --> 00:17:43,789
是的，采用NBL72

499
00:17:43,789 --> 00:17:46,259
72代表GPU数量

500
00:17:46,259 --> 00:17:47,059
所以gb

501
00:17:47,059 --> 00:17:48,660
三百七十二nbl

502
00:17:48,660 --> 00:17:50,259
嗯，现在我们有薇拉·鲁宾

503
00:17:50,259 --> 00:17:52,440
NBL七十二同显卡数量

504
00:17:52,440 --> 00:17:53,960
这也使其能够

505
00:17:53,960 --> 00:17:57,400
因此对于客户来说，从一个平台迁移到另一个平台非常兼容

506
00:17:57,400 --> 00:17:57,920
嗯

507
00:17:57,920 --> 00:18:01,529
这也是保持相同显卡数量的目标之一

508
00:18:01,529 --> 00:18:03,730
相同的Mgx Rap架构

509
00:18:04,930 --> 00:18:07,730
这能让客户更加便捷

510
00:18:07,730 --> 00:18:09,490
生态系统是

511
00:18:09,490 --> 00:18:10,210
你知道的

512
00:18:10,210 --> 00:18:12,670
我们已经与这些机架合作了两代

513
00:18:12,670 --> 00:18:13,049
现在

514
00:18:13,049 --> 00:18:14,329
现在我们迎来了第三代

515
00:18:14,329 --> 00:18:18,579
它们将能够快速运行并以高效率部署

516
00:18:18,579 --> 00:18:19,259
在我们这边

517
00:18:19,259 --> 00:18:19,759
客户不

518
00:18:19,759 --> 00:18:20,819
这完全合理

519
00:18:20,819 --> 00:18:21,240
好的

520
00:18:21,240 --> 00:18:23,539
我们现在可以看看薇拉·鲁宾吗

521
00:18:23,539 --> 00:18:25,279
是的

522
00:18:25,279 --> 00:18:26,920
这就是薇拉·鲁宾

523
00:18:26,920 --> 00:18:27,500
嗯

524
00:18:27,500 --> 00:18:28,480
这就是薇拉·鲁宾

525
00:18:28,480 --> 00:18:29,710
BL七十二机架

526
00:18:29,710 --> 00:18:31,109
你可以看到

527
00:18:31,109 --> 00:18:31,509
你知道的

528
00:18:31,509 --> 00:18:36,509
它在形态和外观上与GB三百非常相似

529
00:18:36,509 --> 00:18:38,189
最大的

530
00:18:38,189 --> 00:18:40,969
最大的区别在于计算模块

531
00:18:40,969 --> 00:18:42,809
你会看到没有散热口

532
00:18:42,809 --> 00:18:44,789
GB三百曾有散热口

533
00:18:44,789 --> 00:18:48,400
因为计算模块下半部分仍有风扇

534
00:18:48,400 --> 00:18:49,039
好的

535
00:18:49,039 --> 00:18:49,500
是的

536
00:18:49,500 --> 00:18:50,900
然后我们去掉了这些风扇

537
00:18:50,900 --> 00:18:53,400
计算模块全部采用100%液冷

538
00:18:53,400 --> 00:18:54,720
这也是为什么前面板设计如此

539
00:18:54,720 --> 00:18:56,679
你不再看到散热口了

540
00:18:56,679 --> 00:19:01,160
但整体仍保留九个交换模块

541
00:19:01,160 --> 00:19:05,130
顶部仍是十个计算单元，底部八个

542
00:19:05,130 --> 00:19:07,430
顶部仍保持相同的遥测系统

543
00:19:07,430 --> 00:19:11,170
仍为机架顶部遥测，配备千兆交换机

544
00:19:11,170 --> 00:19:16,720
现在从黑威尔到鲁宾的机架级对比

545
00:19:16,720 --> 00:19:20,298
谈谈机架级的性能提升

546
00:19:20,298 --> 00:19:25,230
机架级性能提升是十倍十倍的十倍

547
00:19:25,230 --> 00:19:25,670
嗯

548
00:19:25,670 --> 00:19:28,650
每秒每兆瓦或每瓦的令牌数

549
00:19:28,650 --> 00:19:29,219
嗯

550
00:19:29,219 --> 00:19:31,038
这就是机架级别的性能指标

551
00:19:31,038 --> 00:19:33,179
一种性能评估指标

552
00:19:33,179 --> 00:19:35,278
这是混合专家模型的表现

553
00:19:35,278 --> 00:19:37,400
类似Kimi K2的思考方式

554
00:19:37,400 --> 00:19:40,719
这是一个拥有万亿参数的超大型模型

555
00:19:40,719 --> 00:19:45,880
这将完全适配并优化在单个机架中

556
00:19:45,880 --> 00:19:47,059
嗯，借助

557
00:19:47,059 --> 00:19:47,740
你知道的

558
00:19:47,740 --> 00:19:49,579
多亏了NvLink技术

559
00:19:49,579 --> 00:19:55,890
混合专家模型中的专家分布在72块GPU上

560
00:19:55,890 --> 00:20:00,660
这能提升每秒处理的token数量

561
00:20:00,660 --> 00:20:02,599
这里展示的是Khyber RP

562
00:20:02,599 --> 00:20:02,859
好的

563
00:20:02,859 --> 00:20:05,200
这将是Reuben Ultra系列

564
00:20:05,200 --> 00:20:06,539
继Rubin之后

565
00:20:06,539 --> 00:20:09,348
这是2026年的产品，2027年推出

566
00:20:09,348 --> 00:20:11,128
我们将推出Rubin Ultra

567
00:20:11,128 --> 00:20:11,588
好的

568
00:20:11,588 --> 00:20:14,729
这将是不同于以往的机架架构

569
00:20:14,729 --> 00:20:16,490
与前三代产品不同

570
00:20:16,809 --> 00:20:19,170
我们将部署更多计算资源

571
00:20:19,170 --> 00:20:19,630
是的

572
00:20:19,630 --> 00:20:22,910
我发现这里多了很多托盘

573
00:20:22,910 --> 00:20:27,869
每个容器包含18个计算托盘

574
00:20:27,869 --> 00:20:29,440
这里有四个容器

575
00:20:29,440 --> 00:20:32,539
每个容器配备最多72块GPU

576
00:20:32,539 --> 00:20:34,299
所以会有288块

577
00:20:34,299 --> 00:20:37,969
从144块升级到288块

578
00:20:37,969 --> 00:20:39,269
或者说是72块

579
00:20:39,269 --> 00:20:40,608
72块扩展到280块

580
00:20:40,608 --> 00:20:40,848
好的

581
00:20:40,848 --> 00:20:42,689
GPU数量实现四倍增长

582
00:20:42,689 --> 00:20:44,249
每个容器

583
00:20:44,249 --> 00:20:47,980
我提到的四个部分相当于整个机架

584
00:20:47,980 --> 00:20:51,180
这里相当于四个机架的GPU算力

585
00:20:51,180 --> 00:20:54,230
每个容器包含72块计算单元

586
00:20:54,230 --> 00:20:56,230
超高计算密度

587
00:20:56,230 --> 00:20:56,910
是的

588
00:20:56,910 --> 00:20:58,990
这就是架构不同的原因

589
00:20:58,990 --> 00:21:01,029
采用刀片式架构

590
00:21:01,029 --> 00:21:02,549
而非托盘式设计

591
00:21:03,589 --> 00:21:08,710
每个容器有18个计算刀片

592
00:21:08,710 --> 00:21:08,970
抱歉

593
00:21:08,970 --> 00:21:10,109
这些都是计算单元

594
00:21:10,109 --> 00:21:12,179
前面全是计算模块

595
00:21:12,179 --> 00:21:12,719
对的

596
00:21:12,719 --> 00:21:18,460
后面是用于NvLink连接的交换刀片

597
00:21:18,460 --> 00:21:18,960
明白了

598
00:21:18,960 --> 00:21:22,868
这就是性能跃升的关键

599
00:21:22,868 --> 00:21:26,828
这就是Rubin Ultra带来的性能突破

600
00:21:26,828 --> 00:21:28,699
凯伯河中的超频

601
00:21:28,939 --> 00:21:32,618
我们尚未在雷布恩超频上展示任何性能表现

602
00:21:32,618 --> 00:21:37,000
但每代之间性能提升将呈数倍增长

603
00:21:38,000 --> 00:21:40,440
芯片层面将实现显著性能提升

604
00:21:40,440 --> 00:21:41,220
在芯片层面

605
00:21:41,220 --> 00:21:42,680
在超级芯片层面

606
00:21:42,680 --> 00:21:43,599
在机架层面

607
00:21:43,599 --> 00:21:47,259
并将实现四倍的极致协同设计

608
00:21:47,259 --> 00:21:47,660
没错

609
00:21:47,660 --> 00:21:48,640
极致协同设计

610
00:21:48,640 --> 00:21:52,069
所有芯片均针对更高性能进行设计

611
00:21:52,069 --> 00:21:53,400
协同工作

612
00:21:53,424 --> 00:21:55,900
从零共同设计

613
00:21:55,900 --> 00:22:00,279
我们是否期待每代都对六颗芯片进行极致协同设计

614
00:22:00,279 --> 00:22:03,549
从现在起我们应看到六款新芯片

615
00:22:03,549 --> 00:22:04,970
对于每一代产品

616
00:22:04,970 --> 00:22:09,380
现在每年都将推出新一代GPU

617
00:22:09,380 --> 00:22:11,740
但并非所有六颗芯片每年都会协同设计

618
00:22:11,740 --> 00:22:14,619
这可能不会成为常态

619
00:22:14,619 --> 00:22:18,890
但在每一代旗舰产品发布时

620
00:22:18,890 --> 00:22:21,029
比如雷布恩六款新芯片

621
00:22:21,029 --> 00:22:21,429
是的

622
00:22:21,509 --> 00:22:24,449
还会推出与雷布恩超频配套的新芯片

623
00:22:24,449 --> 00:22:26,068
并非全部六款

624
00:22:26,068 --> 00:22:26,449
例如

625
00:22:26,449 --> 00:22:28,420
我们可能会看到维拉CPU

626
00:22:28,420 --> 00:22:29,940
而雷布恩超频

627
00:22:29,940 --> 00:22:32,359
GPU完全正确

628
00:22:32,359 --> 00:22:32,720
是的

629
00:22:32,720 --> 00:22:33,859
我对这个非常期待

630
00:22:33,859 --> 00:22:35,690
迫不及待想看到实际效果

631
00:22:35,690 --> 00:22:38,369
我们何时能了解更多相关信息

632
00:22:38,369 --> 00:22:41,859
这是今年还是明年揭晓

633
00:22:41,859 --> 00:22:42,859
所以是的

634
00:22:42,859 --> 00:22:44,779
詹森将会讨论这个

635
00:22:44,779 --> 00:22:46,279
在即将到来的一年里

636
00:22:46,279 --> 00:22:47,180
嗯

637
00:22:47,180 --> 00:22:48,619
我没有具体日期

638
00:22:48,619 --> 00:22:49,680
但确实

639
00:22:49,680 --> 00:22:51,289
我对此非常兴奋

640
00:22:51,289 --> 00:22:54,609
你最期待什么 最让你兴奋的是什么

641
00:22:54,609 --> 00:22:58,569
这种每年一代的快速进化

642
00:22:58,569 --> 00:23:02,089
在极致协同设计下创新程度

643
00:23:02,089 --> 00:23:05,230
正是最令人印象深刻之处

644
00:23:05,230 --> 00:23:05,589
没错

645
00:23:05,589 --> 00:23:05,970
没错

646
00:23:05,970 --> 00:23:09,829
从一代GPU到下一代

647
00:23:09,829 --> 00:23:13,338
工艺提升的空间有限

648
00:23:13,338 --> 00:23:15,239
技术只能不断优化

649
00:23:16,318 --> 00:23:16,598
你知道的

650
00:23:16,598 --> 00:23:20,970
这不是晶体管数量改进的因素

651
00:23:20,970 --> 00:23:23,079
从一代到下一代的过渡

652
00:23:23,079 --> 00:23:24,240
例如

653
00:23:24,240 --> 00:23:25,019
呃

654
00:23:25,019 --> 00:23:27,359
在薇拉·鲁宾和布莱克威尔之间

655
00:23:27,359 --> 00:23:31,319
晶体管数量增加了约70%

656
00:23:31,319 --> 00:23:34,888
就我们共同设计的各种芯片而言

657
00:23:34,888 --> 00:23:38,209
但我们实现了十倍的性能提升

658
00:23:38,209 --> 00:23:39,729
如果只是遵循摩尔定律

659
00:23:39,729 --> 00:23:41,769
只会是70%的增长

660
00:23:41,769 --> 00:23:44,049
而不是千倍的增长

661
00:23:44,049 --> 00:23:44,390
是的

662
00:23:44,390 --> 00:23:44,890
这不是注意力

663
00:23:44,890 --> 00:23:46,369
所以这种

664
00:23:46,369 --> 00:23:46,769
呃

665
00:23:46,769 --> 00:23:48,829
所有这些协同设计的芯片

666
00:23:48,829 --> 00:23:51,539
共同协作以最大化性能

667
00:23:51,539 --> 00:23:55,519
这是这一代及未来世代最令人惊叹之处

668
00:23:55,519 --> 00:23:56,099
是的

669
00:23:56,099 --> 00:23:57,079
这真的令人兴奋

670
00:23:57,079 --> 00:23:58,569
非常感谢您的时间

671
00:23:58,569 --> 00:23:59,970
衷心感谢乔

672
00:23:59,970 --> 00:24:03,069
德拉瓦州解析英伟达的布莱克威尔生态系统

673
00:24:03,069 --> 00:24:08,210
让我们深入了解鲁宾并解释如何加速AI模型

674
00:24:08,210 --> 00:24:10,089
更智能、更高效

675
00:24:10,089 --> 00:24:11,390
不仅仅是语言模型

676
00:24:11,390 --> 00:24:14,509
从图像和视频模型到医学

677
00:24:14,509 --> 00:24:16,470
机器人技术等更多领域

678
00:24:16,470 --> 00:24:19,650
如果你想深入了解这项技术的科学原理

679
00:24:19,650 --> 00:24:21,769
加入我参加英伟达GTC大会

680
00:24:21,769 --> 00:24:27,420
通过下方链接免费注册，参与多个在线会议

681
00:24:27,420 --> 00:24:30,400
我将在会议后宣布RTX 590显卡赠品得主

682
00:24:30,400 --> 00:24:32,130
会议几日后

683
00:24:32,130 --> 00:24:34,390
务必再次参与

684
00:24:34,390 --> 00:24:37,789
感谢英伟达赞助我的行程和媒体访问

685
00:24:37,789 --> 00:24:39,400
支持GTC现场直播

686
00:24:39,400 --> 00:24:41,460
感谢您的支持

687
00:24:41,460 --> 00:24:42,440
谢谢观看

688
00:24:42,440 --> 00:24:45,329
下次再见，我是股票代码

689
00:24:45,329 --> 00:24:46,430
我是亚历克斯

690
00:24:46,430 --> 00:24:51,250
提醒您最好的投资就是投资自己

评论采集失败

采集完成，该视频无弹幕