Coherent ｜ NPO and CPO for AI Data Centers

1
00:00:00,000 --> 00:00:01,080
下一个发言者是

2
00:00:01,080 --> 00:00:01,580
呃

3
00:00:01,580 --> 00:00:02,200
但代表连贯性

4
00:00:02,200 --> 00:00:04,330
但代表连贯性

5
00:00:04,330 --> 00:00:04,889
呃

6
00:00:04,889 --> 00:00:06,129
我们继续吧

7
00:00:07,129 --> 00:00:11,019
谢谢，灯光在我去第一张幻灯片前

8
00:00:11,019 --> 00:00:15,369
我想强调我的重点是营销动态

9
00:00:15,369 --> 00:00:17,769
为了帮助你们理解我的思路

10
00:00:17,769 --> 00:00:21,969
我想和你们谈谈Theodore Levitt

11
00:00:22,289 --> 00:00:24,129
他是哈佛大学教授

12
00:00:24,129 --> 00:00:30,120
1960年他写了一篇获奖麦肯锡论文

13
00:00:30,120 --> 00:00:33,960
后来成为商学院经典教材

14
00:00:33,960 --> 00:00:35,340
必读内容

15
00:00:35,340 --> 00:00:39,170
他称之为营销近视症

16
00:00:39,170 --> 00:00:40,869
短视行为

17
00:00:40,869 --> 00:00:46,490
他的警告是不要将市场定义得太狭窄

18
00:00:46,490 --> 00:00:49,869
否则技术变革时你会措手不及

19
00:00:49,869 --> 00:00:51,988
这不可避免会发生

20
00:00:51,988 --> 00:00:57,408
因此你能明白我为何选择下一张幻灯片副标题

21
00:00:57,408 --> 00:00:57,628
请

22
00:00:59,069 --> 00:00:59,929
精彩的开场

23
00:00:59,929 --> 00:01:01,304
大家

24
00:01:02,409 --> 00:01:05,209
我认为NPO非常简单

25
00:01:05,209 --> 00:01:07,129
我这里展示三个图表

26
00:01:07,129 --> 00:01:11,189
之前发言者曾向我介绍过

27
00:01:11,189 --> 00:01:13,730
NPO是左侧图表

28
00:01:13,730 --> 00:01:15,430
你有光学器件在PCB上

29
00:01:16,590 --> 00:01:22,489
我认为当前最先进技术CPo是光学器件在基板上

30
00:01:22,489 --> 00:01:27,230
使用ABF基板属于多芯片模块领域

31
00:01:27,230 --> 00:01:31,400
下一代将是CPo在中介层

32
00:01:31,719 --> 00:01:37,250
实用经验法则就是简单说不必如此

33
00:01:37,250 --> 00:01:39,989
但如果在系统PCB上

34
00:01:39,989 --> 00:01:41,409
如果是插座式安装

35
00:01:41,409 --> 00:01:43,170
且超出CPo范围

36
00:01:43,170 --> 00:01:45,025
我称之为NPO

37
00:01:46,620 --> 00:01:50,379
但Theodore Levitt提醒我要拓宽视野

38
00:01:50,379 --> 00:01:52,659
我们转到下一张幻灯片

39
00:01:55,769 --> 00:02:00,209
如果仅从客户需求而非技术角度思考

40
00:02:00,209 --> 00:02:02,689
你会看到

41
00:02:03,989 --> 00:02:12,319
历史上网络行业一直面临模块化与集成化的根本矛盾

42
00:02:12,479 --> 00:02:18,370
模块化能降低风险耦合

43
00:02:18,370 --> 00:02:22,129
更大的生态系统可维护性

44
00:02:23,289 --> 00:02:28,780
但需付出性能损耗因跨越模块边界

45
00:02:28,780 --> 00:02:31,819
无法快速响应

46
00:02:32,900 --> 00:02:33,699
另一方面

47
00:02:33,699 --> 00:02:38,769
集成因接近性带来更好性能

48
00:02:39,088 --> 00:02:40,408
部件更少

49
00:02:40,408 --> 00:02:42,449
可认为更可靠

50
00:02:43,568 --> 00:02:44,620
嗯

51
00:02:45,139 --> 00:02:46,560
你可以快速行动

52
00:02:46,560 --> 00:02:47,759
缺点在于

53
00:02:47,759 --> 00:02:51,430
只有拥有全部系统才能快速行动

54
00:02:51,750 --> 00:02:56,509
在行业中你会遇到处于光谱右侧的客户

55
00:02:56,509 --> 00:02:58,370
处于光谱左侧的客户

56
00:02:58,370 --> 00:03:01,150
中间则是非营利组织

57
00:03:01,150 --> 00:03:02,210
如果我们做得好

58
00:03:02,210 --> 00:03:10,620
工程师应明确理解两者平衡点的位置

59
00:03:10,620 --> 00:03:14,000
然后继续设计产品

60
00:03:14,169 --> 00:03:15,569
为了说明这一点

61
00:03:15,569 --> 00:03:16,150
嗯

62
00:03:16,150 --> 00:03:19,370
可以将其视为练习题

63
00:03:19,370 --> 00:03:21,610
让我们翻到下一页

64
00:03:22,539 --> 00:03:26,908
这里展示的可能是一个CPO系统

65
00:03:26,908 --> 00:03:28,968
甚至可能是一个NPO系统

66
00:03:28,968 --> 00:03:31,209
这里有很多相似之处

67
00:03:31,209 --> 00:03:31,628
没错

68
00:03:31,628 --> 00:03:33,919
这是我们目前看到的情况

69
00:03:33,919 --> 00:03:34,479
呃

70
00:03:34,479 --> 00:03:36,479
系统将会呈现什么样子

71
00:03:36,479 --> 00:03:37,259
嗯

72
00:03:37,259 --> 00:03:38,560
外部激光源

73
00:03:38,560 --> 00:03:39,680
PM光纤

74
00:03:39,680 --> 00:03:42,120
光纤线缆等组件

75
00:03:42,939 --> 00:03:46,969
但利维特会提醒你不能这样做

76
00:03:46,969 --> 00:03:48,030
你必须

77
00:03:48,030 --> 00:03:49,550
如果是CPO

78
00:03:49,550 --> 00:03:51,030
会有不同的需求

79
00:03:51,030 --> 00:03:52,110
如果是NPO

80
00:03:52,110 --> 00:03:53,919
会有不同的需求

81
00:03:53,919 --> 00:03:55,300
这意味着什么

82
00:03:55,300 --> 00:03:57,219
我认为我们可以提出

83
00:03:57,219 --> 00:04:06,050
或假设那些更倾向于集成化的人可能会说

84
00:04:06,050 --> 00:04:12,210
我们刚听到几位演讲者提到将激光移至O

85
00:04:12,629 --> 00:04:15,849
具有模块化思维的人

86
00:04:15,849 --> 00:04:18,250
可能会说

87
00:04:18,250 --> 00:04:19,089
哦

88
00:04:19,089 --> 00:04:22,350
使光纤耦合标准化

89
00:04:22,350 --> 00:04:28,329
以便多人解决光纤与基板连接的问题

90
00:04:28,329 --> 00:04:35,430
这展示了客户需求如何驱动规格选择

91
00:04:35,680 --> 00:04:37,199
如果你同意我的观点

92
00:04:37,199 --> 00:04:38,120
嗯

93
00:04:38,120 --> 00:04:41,879
接下来一页我有个作业

94
00:04:43,098 --> 00:04:45,858
这很难因为这里没有答案

95
00:04:45,858 --> 00:04:48,519
但当你扩大规模

96
00:04:48,519 --> 00:04:52,500
即将进行的光学扩展

97
00:04:52,500 --> 00:04:56,500
我们将面临多种选择

98
00:04:56,839 --> 00:04:58,879
我们可以采用高速通道

99
00:04:58,879 --> 00:05:03,600
采用扩展链路并提升信噪比

100
00:05:03,600 --> 00:05:05,220
因为需要降低延迟

101
00:05:05,220 --> 00:05:06,240
决定论

102
00:05:06,240 --> 00:05:09,079
更低的错误率等等

103
00:05:09,360 --> 00:05:12,639
或者你可以采用更高信噪比的解决方案

104
00:05:12,639 --> 00:05:14,079
比如慢速宽接口

105
00:05:14,079 --> 00:05:14,779
嗯

106
00:05:14,779 --> 00:05:15,759
包括像素

107
00:05:15,759 --> 00:05:19,079
我个人认为这是绝佳的选择

108
00:05:19,819 --> 00:05:22,139
但尚未解决的问题

109
00:05:22,139 --> 00:05:29,569
是这些新选项如何映射到客户需求

110
00:05:29,569 --> 00:05:31,170
偏好高集成度的客户

111
00:05:31,170 --> 00:05:32,790
与高模块化客户

112
00:05:32,790 --> 00:05:34,810
或中间平衡方案

113
00:05:34,810 --> 00:05:35,529
没错

114
00:05:35,529 --> 00:05:37,000
我有一些想法

115
00:05:37,000 --> 00:05:39,720
我认为在扩展过程中中间切换

116
00:05:39,720 --> 00:05:45,019
机架行将采用三种形式

117
00:05:45,019 --> 00:05:45,819
可插拔式

118
00:05:45,819 --> 00:05:47,819
以及NPO和CPO

119
00:05:47,819 --> 00:05:52,139
计算托盘可能结合CPU和NPO

120
00:05:52,139 --> 00:05:55,620
但我很期待听到大家的意见，请随时发邮件

121
00:05:56,740 --> 00:05:58,339
下一张幻灯片

122
00:05:59,980 --> 00:06:01,220
为此

123
00:06:01,220 --> 00:06:03,920
我认为参与此项目的人

124
00:06:03,920 --> 00:06:04,420
嗯

125
00:06:04,420 --> 00:06:04,819
开放

126
00:06:04,819 --> 00:06:05,500
Cpx

127
00:06:05,500 --> 00:06:07,759
MSA正在做正确的事

128
00:06:07,759 --> 00:06:09,879
涵盖范围全面

129
00:06:09,879 --> 00:06:12,680
将覆盖NPO和CPO

130
00:06:12,680 --> 00:06:14,699
各种速度和通道数

131
00:06:14,699 --> 00:06:15,500
插槽式

132
00:06:15,500 --> 00:06:16,540
焊接光学模块

133
00:06:16,540 --> 00:06:17,470
铜缆

134
00:06:17,470 --> 00:06:17,910
呃

135
00:06:17,910 --> 00:06:19,949
我我我支持

136
00:06:19,949 --> 00:06:22,589
这是正确的方向

137
00:06:23,639 --> 00:06:25,199
我准备好结束

138
00:06:25,199 --> 00:06:27,040
幻灯片平放

139
00:06:27,649 --> 00:06:30,930
我所说的是一个简单观点

140
00:06:30,930 --> 00:06:35,920
NPO应设计为

141
00:06:35,920 --> 00:06:42,120
在模块化与集成之间取得平衡

142
00:06:42,120 --> 00:06:44,199
如果我们做得正确

143
00:06:44,199 --> 00:06:46,418
若达成目标

144
00:06:46,418 --> 00:06:49,379
它不会只是CPO的过渡方案

145
00:06:49,379 --> 00:06:52,259
我认为将是一种持久架构

146
00:06:52,259 --> 00:06:54,500
将具备开放可维护性

147
00:06:54,500 --> 00:06:58,550
NPO产品将长久发展

148
00:06:58,670 --> 00:06:59,829
谢谢布莱克

149
00:06:59,829 --> 00:07:00,990
好的

150
00:07:00,990 --> 00:07:02,250
谢谢大家

151
00:07:02,250 --> 00:07:04,189
我觉得我同意你的观点

152
00:07:04,189 --> 00:07:07,990
我们经常看到临时解决方案被引入

153
00:07:07,990 --> 00:07:11,879
然后在行业内待上十年十五年

154
00:07:11,879 --> 00:07:13,959
所以需要明确什么是临时

155
00:07:15,000 --> 00:07:16,519
但有一个快速问题想问你

156
00:07:16,519 --> 00:07:18,860
我们池化一致的是

157
00:07:18,860 --> 00:07:21,259
嗯嗯

158
00:07:21,259 --> 00:07:27,420
有一家公司在开发nmpo以及成品方面

159
00:07:27,420 --> 00:07:32,660
最初是为IBM在2010年设计的

160
00:07:32,660 --> 00:07:38,459
之后他与富士通Fugaku超算项目合作

161
00:07:38,459 --> 00:07:42,079
基于像素的二维设备

162
00:07:42,079 --> 00:07:43,759
我现在忘了它们叫什么了

163
00:07:43,759 --> 00:07:46,939
但我觉得它们属于NPO类别

164
00:07:46,939 --> 00:07:48,930
按照我们今天的定义

165
00:07:48,930 --> 00:07:52,329
所以两个项目都很成功

166
00:07:52,329 --> 00:07:55,490
但都没有后续跟进

167
00:07:55,490 --> 00:07:59,319
你知道通过这个项目学到了什么

168
00:07:59,319 --> 00:08:03,889
我的意思是基于此行业应避免什么

169
00:08:03,889 --> 00:08:05,829
基于你的经验

170
00:08:05,829 --> 00:08:13,819
或Inara在像素基POS前两代的经验

171
00:08:14,720 --> 00:08:18,360
我认为我们学到的积极经验是

172
00:08:18,360 --> 00:08:22,889
我们学到的是插槽是可行的

173
00:08:22,889 --> 00:08:24,110
确实可以实现

174
00:08:24,110 --> 00:08:27,329
我认为支持插槽是个好主意

175
00:08:27,329 --> 00:08:33,789
源于我们实际实施的经验

176
00:08:33,789 --> 00:08:34,889
谨慎的一面

177
00:08:36,370 --> 00:08:41,649
我说的是不要低估会损失多少

178
00:08:41,649 --> 00:08:46,639
从前板可插拔转向板载光学

179
00:08:47,039 --> 00:08:48,120
这是巨大落差

180
00:08:49,279 --> 00:08:49,840
你会

181
00:08:49,840 --> 00:08:51,000
随着增长

182
00:08:51,000 --> 00:08:51,840
不仅于此

183
00:08:51,840 --> 00:08:53,210
你无法进行现场服务插槽

184
00:08:53,210 --> 00:08:56,309
毕竟它们仍较为脆弱

185
00:08:56,309 --> 00:08:59,169
需要专用工具拆卸光学部件

186
00:08:59,169 --> 00:09:01,330
因此突然之间

187
00:09:02,370 --> 00:09:04,570
并非像

188
00:09:04,570 --> 00:09:09,879
只需拉扯拉环就能取出光学部件

189
00:09:09,879 --> 00:09:14,399
前板可插拔与板载光学的可维护性差异大

190
00:09:14,399 --> 00:09:16,240
插拔式与板载光学的可维护性差异

191
00:09:16,240 --> 00:09:19,500
这就是我们学到的经验

192
00:09:19,500 --> 00:09:20,460
好的

193
00:09:20,460 --> 00:09:21,539
太棒了

共 2 条，按点赞数排序。

采集完成，该视频无弹幕