camp2023-57070-eng-Energy_Consumption_of_Data_Centers_opus.srt

1
00:00:00,000 --> 00:00:29,980
 [ Music ]

2
00:00:30,300 --> 00:00:31,460
 >> Hello everyone.

3
00:00:31,460 --> 00:00:35,580
 Welcome to day five on Bits and Boymers stage.

4
00:00:35,580 --> 00:00:40,860
 It's so nice to see all of you even though you look a bit more tired

5
00:00:40,860 --> 00:00:43,300
 than two days ago when I last saw you.

6
00:00:43,300 --> 00:00:51,540
 But I hope we are all going to look as sunny as this day when this talk is over.

7
00:00:51,540 --> 00:00:55,620
 Speaking of a sunny day, please take care too.

8
00:00:55,620 --> 00:00:57,300
 Do you know?

9
00:00:57,300 --> 00:00:59,240
 Drink water, yes.

10
00:00:59,240 --> 00:01:02,820
 And use sunscreen and be in the shadow.

11
00:01:02,820 --> 00:01:04,160
 Take breaks, you know.

12
00:01:04,160 --> 00:01:05,940
 Don't just run from talk to talk.

13
00:01:05,940 --> 00:01:07,180
 Take breaks.

14
00:01:07,180 --> 00:01:08,080
 Sit in the shade.

15
00:01:08,080 --> 00:01:13,140
 And also look that people around you are doing the same.

16
00:01:13,140 --> 00:01:22,420
 Yes. And if you have any thoughts or comments, you can also always post them

17
00:01:22,420 --> 00:01:26,360
 on Mastodon using the hashtag #bitsandboymer.

18
00:01:26,860 --> 00:01:36,300
 Boymer with a E. So, and now I'm really happy that I can announce our speaker, Thomas.

19
00:01:36,300 --> 00:01:40,380
 Thomas is mainly working with clouds and cloud security.

20
00:01:40,380 --> 00:01:44,820
 And when he was looking for energy consumption of data centers,

21
00:01:44,820 --> 00:01:46,560
 he didn't find enough information.

22
00:01:46,560 --> 00:01:48,520
 So he went looking for them.

23
00:01:48,520 --> 00:01:50,960
 And he's going to present them to us today.

24
00:01:50,960 --> 00:01:54,080
 So give a warm applause and welcome to Thomas.

25
00:01:54,080 --> 00:01:55,360
 [ Applause ]

26
00:01:55,360 --> 00:01:56,100
 >> Thank you very much.

27
00:01:56,100 --> 00:02:00,060
 So this is what I'm normally doing.

28
00:02:00,060 --> 00:02:03,260
 Kubernetes, cloud security, and this stuff.

29
00:02:03,260 --> 00:02:05,460
 But let's start.

30
00:02:05,460 --> 00:02:13,580
 So energy consumption of data centers is something which is not really well published.

31
00:02:13,580 --> 00:02:19,040
 So the data center owner are not necessarily publishing a lot of numbers.

32
00:02:19,040 --> 00:02:24,700
 And you see there are a lot of articles and the best ones I have presented here.

33
00:02:24,700 --> 00:02:31,320
 And this study here is from 2020.

34
00:02:31,320 --> 00:02:37,280
 So they show an energy forecast which is exponentially growing.

35
00:02:37,280 --> 00:02:42,480
 And always if you see something exponentially growing, you should be very,

36
00:02:42,480 --> 00:02:46,860
 very careful and aware that something could terribly go wrong.

37
00:02:48,260 --> 00:02:54,440
 If you compare the data center usage with the energy consumption of entire countries,

38
00:02:54,440 --> 00:03:02,040
 you see all the data centers are more or less in the range of South Africa,

39
00:03:02,040 --> 00:03:05,040
 even comparable to the UK or Indonesia.

40
00:03:05,040 --> 00:03:13,160
 So if the data centers would be a country, you see it's a bigger country.

41
00:03:15,840 --> 00:03:23,580
 If you look into the German numbers, you see we are spending 16 terawatt hours on data centers.

42
00:03:23,580 --> 00:03:31,700
 Mostly the heat which is produced by the CPUs is not used anymore.

43
00:03:31,700 --> 00:03:36,420
 There is no decarbonization possible because it's just heat.

44
00:03:36,420 --> 00:03:42,260
 And it is at the moment 3% of the electric power consumption.

45
00:03:43,700 --> 00:03:49,500
 And the dangerous part is it has an annual growth of 6%.

46
00:03:49,500 --> 00:03:55,620
 This means it will double probably around every 12 years.

47
00:03:55,620 --> 00:04:05,140
 So if the data centers from now in the camp, if we talk again, it's a camp in 12 years,

48
00:04:05,140 --> 00:04:10,980
 which would be 2035, then the data centers have doubled the energy consumption

49
00:04:10,980 --> 00:04:12,780
 if this trend continues.

50
00:04:13,780 --> 00:04:17,020
 For all of you who are not familiar with exponential growth,

51
00:04:17,020 --> 00:04:20,500
 here is something which is growing exponentially.

52
00:04:20,500 --> 00:04:26,900
 The first thing, first picture is just radioactive decay.

53
00:04:26,900 --> 00:04:32,900
 So this is a chain reaction or, with other words, it's an atomic bomb.

54
00:04:32,900 --> 00:04:41,180
 And the bomb explosion stops if all the fuel is spent or something other is limited.

55
00:04:41,180 --> 00:04:46,020
 And another example for exponential growth is the spread of COVID.

56
00:04:46,020 --> 00:04:53,740
 In your body, the virus load also increases in the first steps exponentially.

57
00:04:53,740 --> 00:04:56,020
 And the limit is your body.

58
00:04:56,020 --> 00:05:02,260
 If your body cannot provide more resources for the virus,

59
00:05:02,260 --> 00:05:05,140
 then the exponential growth stops.

60
00:05:05,140 --> 00:05:06,540
 The same with the population.

61
00:05:07,100 --> 00:05:13,220
 We had this in COVID and this is the danger of something increasing exponentially,

62
00:05:13,220 --> 00:05:17,500
 that you simply have a limit which you don't want to hit.

63
00:05:17,500 --> 00:05:23,460
 And here you see, this is the green curve,

64
00:05:23,460 --> 00:05:30,500
 is the growth of the data centers predicted for the next years.

65
00:05:30,500 --> 00:05:36,500
 [COUGHS]

66
00:05:36,500 --> 00:05:49,540
 And this makes it dangerous and so something limiting will happen.

67
00:05:49,540 --> 00:05:51,300
 We don't know exactly what.

68
00:05:51,300 --> 00:05:56,900
 It could simply be that they go out of business because the price is too high.

69
00:05:57,860 --> 00:05:59,260
 It could also be...

70
00:05:59,260 --> 00:06:02,340
 Oh, sorry.

71
00:06:02,340 --> 00:06:11,300
 It could also be that it is competing with other resources

72
00:06:11,300 --> 00:06:16,700
 and that simply we cannot spend the energy we use in data centers on other things.

73
00:06:16,700 --> 00:06:23,660
 Here you see the comparison of data centers with aviation shipping.

74
00:06:23,660 --> 00:06:26,540
 And here it's getting interesting.

75
00:06:26,860 --> 00:06:30,220
 It's even competing with food production.

76
00:06:30,220 --> 00:06:31,220
 So...

77
00:06:31,220 --> 00:06:37,500
 And the range is because the data is not really clear.

78
00:06:37,500 --> 00:06:43,540
 The data center owners do not really need to give out the data.

79
00:06:43,540 --> 00:06:46,540
 If you look into...

80
00:06:46,540 --> 00:06:53,420
 If you run your data center on your own or if you should go to a cloud provider,

81
00:06:53,980 --> 00:06:56,020
 the cloud providers, the hyperscalers...

82
00:06:56,020 --> 00:06:57,020
 [COUGHS]

83
00:06:57,020 --> 00:07:04,700
 The hyperscalers are the best in terms of energy saving

84
00:07:04,700 --> 00:07:06,500
 because they have the biggest data centers.

85
00:07:06,500 --> 00:07:08,020
 This is normal.

86
00:07:08,020 --> 00:07:16,380
 And so from an energy perspective, the advice would be go to the cloud

87
00:07:16,380 --> 00:07:19,820
 and don't build your own data center.

88
00:07:19,820 --> 00:07:22,660
 [COUGHS]

89
00:07:23,420 --> 00:07:24,420
 Oh, sorry.

90
00:07:24,420 --> 00:07:28,620
 This is obviously competing with privacy.

91
00:07:28,620 --> 00:07:33,580
 So you have the option not to go into a cloud

92
00:07:33,580 --> 00:07:39,020
 and have GDPR-compliant co-location or your own data center

93
00:07:39,020 --> 00:07:42,900
 or you have to save energy and go into the cloud.

94
00:07:42,900 --> 00:07:47,340
 If you compare another thing,

95
00:07:49,620 --> 00:07:53,820
 then you see it's not only consuming energy, it's also consuming water.

96
00:07:53,820 --> 00:07:58,540
 So the 15 megawatt...

97
00:07:58,540 --> 00:07:59,540
 [COUGHS]

98
00:07:59,540 --> 00:08:07,940
 The 15 megawatt data center is around the energy consumption of a hospital

99
00:08:07,940 --> 00:08:10,620
 or an Almond Orchard.

100
00:08:10,620 --> 00:08:14,700
 And not the energy consumption, sorry, so the water consumption

101
00:08:14,700 --> 00:08:18,780
 or comparable to a golf course.

102
00:08:19,780 --> 00:08:25,420
 So if you have the choice to build a data center, a hospital or a golf course,

103
00:08:25,420 --> 00:08:31,020
 then it depends on your preferences, what you prefer.

104
00:08:31,020 --> 00:08:40,100
 Data center is just for maybe computing, gaming.

105
00:08:40,100 --> 00:08:48,020
 Or for the hospital, yes, this is just for the normal population.

106
00:08:48,020 --> 00:08:51,460
 And if this competes, you have immediately a conflict.

107
00:08:51,460 --> 00:08:57,460
 And normally, the best lobbying wins

108
00:08:57,460 --> 00:09:04,740
 or the best use of venture capital.

109
00:09:04,740 --> 00:09:08,260
 And I don't know how much venture capital goes to the hospitals.

110
00:09:12,220 --> 00:09:20,900
 If you see here, there is some number which is always used for advertising.

111
00:09:20,900 --> 00:09:23,780
 This is power usage effectiveness.

112
00:09:23,780 --> 00:09:30,100
 It has been invented by Google because Google looks very good with this.

113
00:09:30,100 --> 00:09:37,060
 And you see here, it's a number and they improve all the time.

114
00:09:40,780 --> 00:09:44,060
 And this number is simply the total energy consumption

115
00:09:44,060 --> 00:09:48,220
 by the real load which is used for computation.

116
00:09:48,220 --> 00:09:51,420
 This means if the number would be one,

117
00:09:51,420 --> 00:09:54,420
 then it would be a 100% effective data center

118
00:09:54,420 --> 00:10:01,220
 without need of cooling and energy for cooling and infrastructure.

119
00:10:01,220 --> 00:10:04,460
 And if always, if you see the number, it's like in the...

120
00:10:04,460 --> 00:10:08,580
 Another example is if it's in the Deutsche Bahn,

121
00:10:08,580 --> 00:10:14,060
 the Deutsche Bahn train shows you only the speed if it's fast.

122
00:10:14,060 --> 00:10:19,220
 It does not show you if your ICE is spending 30 kilometers per hour.

123
00:10:19,220 --> 00:10:26,220
 It shows you if it is going with 200 kilometers.

124
00:10:26,220 --> 00:10:29,300
 And so if you see the numbers, normally they are good.

125
00:10:29,300 --> 00:10:31,780
 You don't ever see the bad numbers.

126
00:10:31,780 --> 00:10:38,220
 And here is another curve which shows how Google makes it.

127
00:10:39,180 --> 00:10:42,980
 The energy, the water consumption of Google data centers is much higher

128
00:10:42,980 --> 00:10:49,300
 than the water consumption of other cloud providers.

129
00:10:49,300 --> 00:10:58,380
 So and then they compare obviously with others,

130
00:10:58,380 --> 00:11:00,420
 with countries, with other cloud providers.

131
00:11:00,420 --> 00:11:03,780
 So you should take these numbers with care

132
00:11:04,780 --> 00:11:09,180
 because on this scale, a renewable energy consumption,

133
00:11:09,180 --> 00:11:14,420
 Amazon is better than Germany.

134
00:11:14,420 --> 00:11:17,100
 And I simply don't believe these numbers.

135
00:11:17,100 --> 00:11:21,620
 So resource conflicts,

136
00:11:21,620 --> 00:11:26,780
 they are popping up everywhere where data centers are built.

137
00:11:26,780 --> 00:11:33,140
 In Virginia, people are gathering against an Amazon facility

138
00:11:34,060 --> 00:11:36,140
 with a consumption of 600 megawatt.

139
00:11:36,140 --> 00:11:43,260
 600 megawatt is more or less half a power plant.

140
00:11:43,260 --> 00:11:50,940
 In Ireland, the data centers make...

141
00:11:50,940 --> 00:11:57,060
 They take 18% of the energy consumption of Ireland

142
00:11:57,060 --> 00:12:01,900
 and the national strategy of CO2 reduction

143
00:12:02,540 --> 00:12:07,620
 now is going to cull 65,000 cows

144
00:12:07,620 --> 00:12:15,580
 because they want to limit their carbon dioxide production.

145
00:12:15,580 --> 00:12:18,660
 So you have here data centers against cows.

146
00:12:18,660 --> 00:12:23,500
 And if you remember Ireland, in my memory,

147
00:12:23,500 --> 00:12:30,540
 Ireland was also very famous for herding cows and sheep.

148
00:12:30,740 --> 00:12:36,980
 And comparable, the data centers are using as much energy

149
00:12:36,980 --> 00:12:40,060
 as all the private city homes.

150
00:12:40,060 --> 00:12:47,660
 In the UK, they have in East London,

151
00:12:47,660 --> 00:12:53,460
 created a big environmental opposition movement

152
00:12:53,460 --> 00:12:58,540
 because they cannot do housing anymore.

153
00:12:58,540 --> 00:13:00,540
 If you don't have energy for housing

154
00:13:00,540 --> 00:13:04,140
 because you want to sell the energy to the data centers,

155
00:13:04,140 --> 00:13:05,940
 then this is the next problem.

156
00:13:05,940 --> 00:13:10,300
 So always where these huge data centers come on a local scale,

157
00:13:10,300 --> 00:13:16,660
 you have this kind of resource conflict, energy and/or water.

158
00:13:16,660 --> 00:13:19,900
 Same in the Netherlands.

159
00:13:23,220 --> 00:13:29,940
 And they even talk about that they need more dikes

160
00:13:29,940 --> 00:13:35,020
 so that they build higher walls against the North Sea

161
00:13:35,020 --> 00:13:40,500
 because if they pump up this water for cooling

162
00:13:40,500 --> 00:13:48,540
 and vapor it, then the groundwater is sinking.

163
00:13:48,540 --> 00:13:54,460
 And this means they need to put more energy in moving.

164
00:13:54,460 --> 00:14:00,540
 And then the next step is the saltwater comes from the ocean

165
00:14:00,540 --> 00:14:02,580
 and then they have other conflicts too.

166
00:14:02,580 --> 00:14:10,540
 This is another example in the Dallas in the US.

167
00:14:10,540 --> 00:14:15,020
 This is on the local scale, it's also an exponential growth.

168
00:14:15,020 --> 00:14:19,020
 So this means typically you can compute,

169
00:14:19,020 --> 00:14:25,500
 you need between 1.5 and 2.3 liter water per kilowatt hour energy.

170
00:14:25,500 --> 00:14:37,340
 This is typically 15 to 23 cubic meters for a 10 megawatt data center.

171
00:14:37,340 --> 00:14:41,220
 And is this a big one or not?

172
00:14:41,220 --> 00:14:42,860
 So I looked it up.

173
00:14:43,420 --> 00:14:47,180
 The typical sizes of data centers start from 15 kilowatts.

174
00:14:47,180 --> 00:14:50,340
 So this is a, the racks fit into a container.

175
00:14:50,340 --> 00:14:55,820
 The typical things we have here from NTT in Berlin or otherwise

176
00:14:55,820 --> 00:15:00,540
 are between 5 megawatt and 40 megawatt.

177
00:15:00,540 --> 00:15:02,860
 And the hyperscalers have more or less no limit.

178
00:15:02,860 --> 00:15:06,740
 So the biggest I've read about is 600 megawatt.

179
00:15:08,460 --> 00:15:15,820
 And the typical water use is between 4 and 7 million cubic meters water per year.

180
00:15:15,820 --> 00:15:20,460
 This is a medium large city for a data center.

181
00:15:20,460 --> 00:15:28,340
 And the first stops have arrived.

182
00:15:28,340 --> 00:15:33,700
 So in Ireland, Microsoft and Amazon have to hold their plants.

183
00:15:33,700 --> 00:15:34,980
 In the Netherlands too.

184
00:15:35,780 --> 00:15:40,420
 In Germany, in Brandenburg, we had the problem that Alphabet,

185
00:15:40,420 --> 00:15:42,300
 so the company behind Google,

186
00:15:42,300 --> 00:15:47,980
 is not allowed to build a data center where they originally have planted.

187
00:15:47,980 --> 00:15:50,500
 It has now moved to Mittenwalde.

188
00:15:50,500 --> 00:15:59,060
 The data center in Neuhagen would have been optimal by energy usage

189
00:15:59,060 --> 00:16:06,860
 because there is a power switch station.

190
00:16:06,860 --> 00:16:12,140
 But where they build it now, it's more or less in the middle of nowhere.

191
00:16:12,140 --> 00:16:18,100
 And they have no use for the heat and something like this.

192
00:16:18,100 --> 00:16:24,620
 In Neuhagen, they have planned a PoE value of 1.03.

193
00:16:24,620 --> 00:16:27,060
 And in Mittenwalde, they don't tell the number.

194
00:16:28,060 --> 00:16:34,860
 And one remark, all the studies I've seen so far

195
00:16:34,860 --> 00:16:39,380
 are before the artificial intelligence boom took off.

196
00:16:39,380 --> 00:16:46,540
 So what we know so far is that training chat GPT uses 1.

197
00:16:46,540 --> 00:16:47,740
 Oh, God.

198
00:16:47,740 --> 00:16:54,740
 1,287 gigawatt hours.

199
00:16:54,740 --> 00:17:01,420
 So roughly the equivalent of 120 US homes for a single year.

200
00:17:01,420 --> 00:17:06,020
 This means this comes on top of the situation here.

201
00:17:06,020 --> 00:17:12,460
 So the artificial intelligence is in the training phase using a lot of energy.

202
00:17:12,460 --> 00:17:19,300
 And here you also can compute this PoE value,

203
00:17:19,300 --> 00:17:20,620
 and this is a very bad one.

204
00:17:21,380 --> 00:17:27,260
 Here you see what you all have to take into consideration.

205
00:17:27,260 --> 00:17:29,500
 So CPU RAM and the GPUs.

206
00:17:29,500 --> 00:17:34,100
 And the GPUs are normally very similar to the Nvidia graphic cards

207
00:17:34,100 --> 00:17:36,300
 you run in your gaming computer at home.

208
00:17:36,300 --> 00:17:39,140
 And they are extremely energy hungry.

209
00:17:39,140 --> 00:17:45,260
 So here are some more numbers.

210
00:17:45,260 --> 00:17:49,180
 And the interesting part is here.

211
00:17:50,100 --> 00:17:56,500
 On an average, you can estimate by the computing cost the energy consumption.

212
00:17:56,500 --> 00:18:00,140
 So here are typical hardwares.

213
00:18:00,140 --> 00:18:02,740
 TPUs are tensor processor units.

214
00:18:02,740 --> 00:18:11,300
 So these are special graphic GPU-like devices for artificial intelligence.

215
00:18:12,060 --> 00:18:15,180
 And here you see the price.

216
00:18:15,180 --> 00:18:22,900
 And here is the equivalent for carbon dioxide in Ips.

217
00:18:22,900 --> 00:18:25,100
 So this is an American thing.

218
00:18:25,100 --> 00:18:29,860
 It's sometimes hard to calculate it in normal units.

219
00:18:29,860 --> 00:18:35,340
 More interesting is the comparison with real-life situations.

220
00:18:35,340 --> 00:18:39,300
 For example, here you see the consumption for air travel

221
00:18:39,300 --> 00:18:41,020
 between New York and San Francisco.

222
00:18:41,940 --> 00:18:49,060
 This is what you normally consume in one year, or what a US citizen consumes.

223
00:18:49,060 --> 00:18:56,700
 This is an XZEP36, 1156 Ips.

224
00:18:56,700 --> 00:19:00,460
 And a car for the full lifetime.

225
00:19:00,460 --> 00:19:05,820
 And here, there you see the training of one model with a GPU.

226
00:19:05,820 --> 00:19:10,500
 This is a neural processing pipeline.

227
00:19:11,500 --> 00:19:15,500
 It doesn't look so much, but you have to tune it,

228
00:19:15,500 --> 00:19:17,900
 and you have to do a lot of experiments.

229
00:19:17,900 --> 00:19:25,660
 And this means this increases to that number which is comparable to a car

230
00:19:25,660 --> 00:19:27,180
 in the entire lifetime.

231
00:19:27,180 --> 00:19:34,140
 So this means training in an artificial intelligence model

232
00:19:34,140 --> 00:19:39,100
 is comparable to running a car for the entire lifetime of a typical car.

233
00:19:40,100 --> 00:19:45,580
 Compared to that, Bitcoin is the worst.

234
00:19:45,580 --> 00:19:48,020
 It's a total waste of energy.

235
00:19:48,020 --> 00:19:55,980
 It is consuming more energy than some even developed countries.

236
00:19:55,980 --> 00:20:02,900
 From my personal experience, we had always these things.

237
00:20:02,900 --> 00:20:07,820
 Some developers come with a genius of an idea,

238
00:20:07,820 --> 00:20:11,340
 and they want to go live in critical infrastructure

239
00:20:11,340 --> 00:20:13,020
 so there is no cloud involved.

240
00:20:13,020 --> 00:20:19,100
 For example, a trading software which should be run on OpenShift.

241
00:20:19,100 --> 00:20:22,380
 The vendor came up with 10 nodes of HBase.

242
00:20:22,380 --> 00:20:25,660
 HBase is this Google-like database you can host for free.

243
00:20:25,660 --> 00:20:28,140
 It's an Apache project.

244
00:20:28,140 --> 00:20:33,100
 And then we tried to measure the load and calculated the real needs.

245
00:20:33,100 --> 00:20:36,780
 And finally, it was hard to detect the load.

246
00:20:37,780 --> 00:20:41,220
 And this means even my SQL database would be oversized.

247
00:20:41,220 --> 00:20:49,860
 And finally, we could have run this on an SQLite database on a Raspberry Pi.

248
00:20:49,860 --> 00:20:59,740
 And this means we have a trend in software that it has an incentive

249
00:20:59,740 --> 00:21:06,140
 that the team which spends the most resources is preferred.

250
00:21:06,140 --> 00:21:09,380
 So another experience from 10 years ago

251
00:21:09,380 --> 00:21:18,380
 was I have automated the database rollout for a customer,

252
00:21:18,380 --> 00:21:21,260
 and then we don't need an Oracle team anymore.

253
00:21:21,260 --> 00:21:25,900
 And yeah, Oracle teams are expensive,

254
00:21:25,900 --> 00:21:28,340
 but they immediately looked for Hadoop,

255
00:21:28,340 --> 00:21:34,620
 which means they have a huge cluster solution for a big data file system.

256
00:21:34,620 --> 00:21:41,740
 And so they could use more resources in their daily job.

257
00:21:41,740 --> 00:21:44,140
 And this is a problem.

258
00:21:44,140 --> 00:21:47,700
 I call it server machism.

259
00:21:47,700 --> 00:21:53,540
 The more horsepower, it's the same as imposing with expensive cars.

260
00:21:53,540 --> 00:21:55,700
 The more horsepower, the more important is the team.

261
00:21:55,700 --> 00:22:00,820
 And you should rethink your big data paradigm.

262
00:22:00,820 --> 00:22:03,180
 In most of the cases, it's completely wrong.

263
00:22:04,540 --> 00:22:10,100
 If you have to increase the size of your data to compute it,

264
00:22:10,100 --> 00:22:11,500
 then something is completely wrong.

265
00:22:11,500 --> 00:22:12,940
 I've seen customers doing this.

266
00:22:12,940 --> 00:22:20,580
 And on the other side, this idea cuts totally into the business model of data centers

267
00:22:20,580 --> 00:22:26,220
 because they want to sell you compute power, storage, network power.

268
00:22:26,220 --> 00:22:31,620
 And not only the hyperscalers have this problem,

269
00:22:31,620 --> 00:22:36,220
 but also the normal ISPs and companies who do the housing.

270
00:22:36,220 --> 00:22:42,580
 In the background, you could see one of these electric cars

271
00:22:42,580 --> 00:22:48,980
 which can go a thousand kilometers without loading.

272
00:22:48,980 --> 00:22:52,700
 It's completely useless because this is not a use case.

273
00:22:52,700 --> 00:22:55,340
 We don't need these big things anymore.

274
00:22:55,340 --> 00:22:59,020
 We need electric bicycles, but not electric cars.

275
00:22:59,020 --> 00:23:01,420
 But this is not really visible at the moment.

276
00:23:01,740 --> 00:23:05,740
 [APPLAUSE]

277
00:23:05,740 --> 00:23:09,740
 Thank you.

278
00:23:09,740 --> 00:23:13,740
 Another effect I'm responsible for is this so-called rebound effect.

279
00:23:13,740 --> 00:23:17,140
 If you make something more accessible, it's used more,

280
00:23:17,140 --> 00:23:23,020
 and then the consumption increases, and then the energy consumption also increases.

281
00:23:23,020 --> 00:23:28,460
 I looked it up and it is known as Yevon's paradox.

282
00:23:28,460 --> 00:23:35,020
 It was detected for steam engines in the beginning.

283
00:23:35,020 --> 00:23:39,780
 So if the cost increases, let's say from 100 to 80,

284
00:23:39,780 --> 00:23:47,020
 then the usage increases more than this amount of 20%.

285
00:23:47,020 --> 00:23:50,700
 So it goes from 10 to 40, so this is 40%.

286
00:23:50,700 --> 00:23:53,060
 This means if you make something cheaper, it's immediately...

287
00:23:53,060 --> 00:23:59,900
 use more. This is logical and this absolutely makes everything complicated.

288
00:23:59,900 --> 00:24:06,220
 If you make something easier, people are using it more than proportionally.

289
00:24:06,220 --> 00:24:11,540
 We have the same in the cloud and my personal observation is I'm doing

290
00:24:11,540 --> 00:24:19,900
 Kubernetes so I try to make cloud access as easy as possible and in this cloud

291
00:24:19,900 --> 00:24:26,260
 access there are hidden costs which I without exactly knowing them shrink.

292
00:24:26,260 --> 00:24:30,540
 So if you can do self provisioning you are unable to use the cloud which is my

293
00:24:30,540 --> 00:24:35,500
 daily job or do virtualization containers Kubernetes and automation.

294
00:24:35,500 --> 00:24:41,820
 Finally more people have access to the resources and this means effectively you

295
00:24:41,820 --> 00:24:47,820
 are triggering an increase in resource consumption. Even if the effective

296
00:24:47,820 --> 00:24:51,180
 container might be smaller than the virtual machine or what you have

297
00:24:51,180 --> 00:24:59,740
 before, finally the overall, the total energy consumption may increase.

298
00:24:59,740 --> 00:25:07,660
 And this is a problem we should have in mind. Further problems if we don't have

299
00:25:07,660 --> 00:25:12,300
 any transparent information on data centers. We only know roughly

300
00:25:12,300 --> 00:25:18,860
 hyperscalers are better than self-run data centers. Here's another bad old

301
00:25:18,860 --> 00:25:24,860
 example. So this is Berlin Spandau. If you know it here is the old E-Shelter

302
00:25:24,860 --> 00:25:31,620
 data center. Here is the Finanzamt, so the tax office Spandau. And here is a lot of

303
00:25:31,620 --> 00:25:40,100
 cooling devices. And on the other side of the channel there is a coal power

304
00:25:40,100 --> 00:25:45,100
 plant which also produces heat for the local heat network. And they are

305
00:25:45,100 --> 00:25:50,420
 completely disconnected. So the heat of the data center is wasted and on the

306
00:25:50,420 --> 00:26:02,380
 other side of the channel we produce heat to get our homes warm. And I've

307
00:26:02,380 --> 00:26:09,540
 criticized it even on the last Bits on Boehme conference. NTT has now a new

308
00:26:09,540 --> 00:26:14,100
 data center in Berlin Marienpark. They have two buildings and they have

309
00:26:14,100 --> 00:26:22,020
 connected up with the GASAC. So this is a, they are not only providing gas but

310
00:26:22,020 --> 00:26:29,940
 also heat for housing and for the industry. And they have a total renewable

311
00:26:29,940 --> 00:26:38,820
 data center with in the final stage 48 megawatt. The PoE is smaller than 1.3.

312
00:26:38,820 --> 00:26:46,180
 And they are based 100% of renewable energy. And this produces 29 gigawatt

313
00:26:46,180 --> 00:26:51,180
 hours a year. So this is a reasonable solution. So at least it's not wasted.

314
00:26:51,180 --> 00:26:58,300
 But it is not possible to use the old data centers and simply attach the heat

315
00:26:58,300 --> 00:27:06,260
 transform. Further topics here from Germany. We have drought everywhere. So

316
00:27:06,260 --> 00:27:11,140
 problem is if I want to cool my data center with local water, I need water.

317
00:27:11,140 --> 00:27:16,740
 And if the groundwater is sinking, I don't have this water. So Frankfurt has

318
00:27:16,740 --> 00:27:24,940
 this problem at the moment. And they have all over this reported these problems.

319
00:27:24,940 --> 00:27:34,340
 In Brandenburg, Google had also these problems. And it is absolutely unclear

320
00:27:34,340 --> 00:27:38,900
 if they can use it. If it's in the middle of nowhere of Mittenwalde,

321
00:27:38,900 --> 00:27:44,540
 probably there is no heat consumer. So definitely they will also waste the

322
00:27:44,540 --> 00:27:51,700
 energy. So this is a Google case. What you can conclude from the public

323
00:27:51,700 --> 00:27:58,580
 available information. And they don't tell you the PoE value, which means it's

324
00:27:58,580 --> 00:28:09,620
 a bad value. Otherwise they would tell you. Probably it's in the vintage data

325
00:28:09,620 --> 00:28:14,420
 center location because Google is not always hosting their own computers in

326
00:28:14,420 --> 00:28:21,660
 their own data centers. So the new energy law, it is valid only for new data

327
00:28:21,660 --> 00:28:32,900
 centers, not for the old ones. And the PoE must be below 1.2, but only from

328
00:28:32,900 --> 00:28:43,740
 2026 on. By the way, the PoE can run. I will give an example soon.

329
00:28:43,740 --> 00:28:50,660
 Data center must have a power of at least 200 kilowatt, which is not a very

330
00:28:50,660 --> 00:29:04,500
 big one. It has been diluted by lobbying. 300. Okay. So it's even worth 300.

331
00:29:04,500 --> 00:29:12,580
 Thank you, Anke. And a lot of lobbyism took over. And they must integrate it

332
00:29:12,580 --> 00:29:15,940
 into local heating systems. But in Mittenwalde there are no local heating

333
00:29:15,940 --> 00:29:21,300
 systems on that scale. There is a Golem article on that. So we have this

334
00:29:21,300 --> 00:29:32,140
 problem. And I don't want you to go without positive hope. We have the

335
00:29:32,140 --> 00:29:36,140
 ability of building very small data centers which can integrate it in your

336
00:29:36,140 --> 00:29:41,580
 local energy and heating system. For example, by the Hochschule Osnabrück,

337
00:29:41,580 --> 00:29:49,500
 they have created on their campus data centers in real industry containers.

338
00:29:49,500 --> 00:29:54,740
 So you can rack it. You attach it to energy and to the network, and then

339
00:29:54,740 --> 00:30:01,860
 it boots. What is booting a form of OpenStack? From the sovereign cloud

340
00:30:01,860 --> 00:30:08,460
 stack. They reported it on the summit. And you can integrate these small

341
00:30:08,460 --> 00:30:17,580
 containers in your local heating providers. So you have, they call it

342
00:30:17,580 --> 00:30:24,980
 near warming. It uses renewable energy. It is kind of cloud computing inside

343
00:30:24,980 --> 00:30:31,740
 the container. But this only means everything is automated. It's done by

344
00:30:31,740 --> 00:30:37,740
 two German companies. It's 100% open source and it runs totally

345
00:30:37,740 --> 00:30:44,020
 automatically. So the only problem if you want to drop these containers

346
00:30:44,020 --> 00:30:49,660
 everywhere into your local heating systems is security. So you have to

347
00:30:49,660 --> 00:30:55,140
 build the containers in a way that nobody has access or that it is turning

348
00:30:55,140 --> 00:31:00,900
 off everything if somebody gets access. But this is solvable. What did they

349
00:31:00,900 --> 00:31:07,260
 do? They have upcycled the heat to 85 degrees of Celsius, which is the

350
00:31:07,260 --> 00:31:14,140
 lower level where these heating systems accept the energy. It's scalable

351
00:31:14,140 --> 00:31:18,660
 between 50 kilowatts, so this is a small container, and 50 megawatt,

352
00:31:18,660 --> 00:31:26,620
 which is roundabout the same as NTT. It makes economic sense. They told me

353
00:31:26,620 --> 00:31:32,940
 beyond 250 kilowatts, this means five containers or something like this.

354
00:31:32,940 --> 00:31:38,420
 This means you have two gigawatt hours or the equivalent of 400 houses.

355
00:31:38,420 --> 00:31:44,180
 But the customers want at least 500 megawatt data centers. So if you

356
00:31:44,180 --> 00:31:49,580
 calculate this, this means 40 gigawatt hours or you have to have 8,000

357
00:31:49,580 --> 00:31:54,740
 houses or some kind of industrial usage for the heat. The good news is this

358
00:31:54,740 --> 00:32:00,100
 is not just science fiction because the real estate companies are

359
00:32:00,100 --> 00:32:07,660
 interested to build it into new housing solutions. So this seems to have

360
00:32:07,660 --> 00:32:13,780
 a value. Another more science fiction-like project is in the tropical

361
00:32:13,780 --> 00:32:21,500
 country of Finland. They simply have a sand battery for heat, so they heat

362
00:32:21,500 --> 00:32:29,700
 up sand in the summer or with whatever measure. Here, at the moment,

363
00:32:29,700 --> 00:32:35,740
 it's not really visible, they directly transform energy into heat into

364
00:32:35,740 --> 00:32:44,100
 the sand. This means you could also include some CPU computing in the

365
00:32:44,100 --> 00:32:50,580
 middle of this process that you get a better usage of the energy. So this

366
00:32:50,580 --> 00:32:55,620
 is a waste of energy in the term of, yes, it could also compute something

367
00:32:55,620 --> 00:33:05,500
 in between. It's called the Hot Heart Project. What we see here, it's

368
00:33:05,500 --> 00:33:11,540
 quite complicated. We have seen this with heating discussion in Germany

369
00:33:11,540 --> 00:33:19,220
 where people even triggered a campaign which was more or less a psyops

370
00:33:19,220 --> 00:33:25,940
 campaign against the green transformation of heating houses. So this

371
00:33:25,940 --> 00:33:31,900
 is more complex and this will take definitely longer. So in my conclusion,

372
00:33:31,900 --> 00:33:36,460
 what we can think about, and it is a process that will take more or less

373
00:33:36,460 --> 00:33:43,020
 10 years, 20 years or 30 years, we should integrate data centres into the

374
00:33:43,020 --> 00:33:52,180
 infrastructure with a PoE of one and use the heat. And the Gazak told

375
00:33:52,180 --> 00:33:56,700
 definitely, it's probably not the perfect solution we have, but it's

376
00:33:56,700 --> 00:33:59,980
 better to start now than wait for the perfect solution because the

377
00:33:59,980 --> 00:34:05,700
 perfect solution could look different on every location. There's a lot of

378
00:34:05,700 --> 00:34:12,540
 pressure behind because otherwise we don't get our decarbonisation goals

379
00:34:12,540 --> 00:34:21,180
 and this is really something we should not miss. So this was more or less

380
00:34:21,180 --> 00:34:25,700
 an overview. I think it was not sufficient in that term that you have

381
00:34:25,700 --> 00:34:31,140
 now a plan to do, but I think I wanted to raise at least that we have a

382
00:34:31,140 --> 00:34:37,300
 problem here and that there are people starting to have solutions, but

383
00:34:37,300 --> 00:34:41,060
 definitely not final. So if you want to contact me, here is my contact

384
00:34:41,060 --> 00:34:55,620
 data. And now we are open for Q&A.

385
00:34:55,620 --> 00:35:00,940
 So thank you so much, Thomas, for giving us all this knowledge. Over

386
00:35:00,940 --> 00:35:06,620
 there is an angel. If you have a question, please go over to the angel.

387
00:35:06,620 --> 00:35:11,820
 You can just queue with the angel and the microphone. Okay. Do we find a

388
00:35:11,820 --> 00:35:20,500
 way to do this? Wow. Good that we stopped early. So as we only have like

389
00:35:20,500 --> 00:35:25,660
 30 minutes for questions, maybe not more people queue.

390
00:35:25,660 --> 00:35:30,500
 >> Okay. >> Yeah. I just read this week an article

391
00:35:30,500 --> 00:35:36,060
 about the Google data centre plant in Brandenburg and they said they will

392
00:35:36,060 --> 00:35:42,460
 not really need much water because like most big new data centres, they

393
00:35:42,460 --> 00:35:48,940
 have closed water cycles combined with air cooling and then they just

394
00:35:48,940 --> 00:35:52,980
 don't need that big water consumption. So what is your perspective on

395
00:35:52,980 --> 00:35:57,980
 this? >> Normally this is a POE value, so normally

396
00:35:57,980 --> 00:36:04,460
 you can get a good POE only with water. And you can exchange --

397
00:36:04,460 --> 00:36:09,220
 >> It's still a water cooling system, but it's a closed system. So you

398
00:36:09,220 --> 00:36:13,620
 need to fill it once, but not all of the time.

399
00:36:13,620 --> 00:36:17,460
 >> But then it's not as efficient as they would simply vaporize the

400
00:36:17,460 --> 00:36:33,140
 water. It's simply -- if it is closed, it's better than air cooling.

401
00:36:33,140 --> 00:36:41,380
 >> We have this -- >> But you can reach a POE of 1.2, that's

402
00:36:41,380 --> 00:36:43,900
 possible? >> It's possible. And Google has done it,

403
00:36:43,900 --> 00:36:47,300
 but at the moment only with a lot of water consumption. And if they

404
00:36:47,300 --> 00:36:53,980
 can be better, I think they are happy if they are below this 1.2, but

405
00:36:53,980 --> 00:36:58,500
 they will not go down to 1.03 or something like this. So this is an

406
00:36:58,500 --> 00:37:02,620
 improvement, but it's not the thing they could have if they would

407
00:37:02,620 --> 00:37:07,740
 waste energy. >> I appreciate that there are smaller options,

408
00:37:07,740 --> 00:37:11,700
 because I also don't believe in the giant solutions. One last

409
00:37:11,700 --> 00:37:16,500
 sentence. We can reuse heat also for other things, not just for

410
00:37:16,500 --> 00:37:18,420
 warming houses. >> Exactly.

411
00:37:18,420 --> 00:37:23,340
 >> We can dry food and agriculture, we can have greenhouses to grow

412
00:37:23,340 --> 00:37:27,260
 stuff in the middle of nowhere like Mittenwalde. So there are other

413
00:37:27,260 --> 00:37:33,220
 options, too, but it's a disaster that the law has been -- has been

414
00:37:33,220 --> 00:37:38,060
 destroyed to its core. >> The problem is then every solution is

415
00:37:38,060 --> 00:37:41,900
 individual. So you cannot simply have -- yeah, we use heat for one

416
00:37:41,900 --> 00:37:45,740
 data center and scale it and run it everywhere. You have the problem

417
00:37:45,740 --> 00:37:49,020
 that it has to be designed locally, which can also be an advantage

418
00:37:49,020 --> 00:37:53,100
 for the local industry, but there is no single --

419
00:37:53,100 --> 00:37:58,060
 >> And if you force them by law, they have to find a solution, and

420
00:37:58,060 --> 00:38:01,940
 now they don't have to anymore. >> Yeah. The law is --

421
00:38:01,940 --> 00:38:08,860
 >> So maybe short questions and short answers so everyone can ask

422
00:38:08,860 --> 00:38:13,140
 their question. >> Okay. My question is in the same direction.

423
00:38:13,140 --> 00:38:18,060
 Isn't there a law which forbids to use fresh water with all the

424
00:38:18,060 --> 00:38:22,660
 dirt grouts and water -- >> Not at the moment. So you need a

425
00:38:22,660 --> 00:38:30,020
 license like Tesla in Brandenburg or like brown coal in -- but then

426
00:38:30,020 --> 00:38:32,660
 if you have this license, you can do whatever you want with the

427
00:38:32,660 --> 00:38:39,380
 local water. >> Hey, thanks for the talk. A question about the

428
00:38:39,380 --> 00:38:42,740
 software side of things. So as you mentioned, there's a lot of

429
00:38:42,740 --> 00:38:45,900
 opportunities to -- especially when dimensioning a system to

430
00:38:45,900 --> 00:38:49,100
 reduce the size and therefore reduce the consumption and make

431
00:38:49,100 --> 00:38:51,700
 better use of the resources that we already have in those data

432
00:38:51,700 --> 00:38:54,940
 centers. Do you think that working on the foundational

433
00:38:54,940 --> 00:38:57,580
 technologies, so for example building more efficient compilers

434
00:38:57,580 --> 00:39:00,540
 with smaller software closures that end up in all of our Docker

435
00:39:00,540 --> 00:39:05,060
 containers would make an actual impact in the long term?

436
00:39:05,060 --> 00:39:10,500
 >> All this helps. Even better CPUs would help, and you have to

437
00:39:10,500 --> 00:39:12,860
 look into it. For example, there are a lot of comparisons that

438
00:39:12,860 --> 00:39:15,860
 Python is inefficient, which is not true, because Python is

439
00:39:15,860 --> 00:39:20,580
 mostly a wrapper around good old C or Fortran libraries. So you

440
00:39:20,580 --> 00:39:24,860
 have to look into your application and measure the energy. And

441
00:39:24,860 --> 00:39:29,180
 effectively, you would need somebody in your team who is aware

442
00:39:29,180 --> 00:39:34,900
 of this and can scale the energy load down. This is something you

443
00:39:34,900 --> 00:39:37,660
 have to learn. It's not your daily business probably as a

444
00:39:37,660 --> 00:39:41,420
 developer, but somebody should have this knowledge in your team

445
00:39:41,420 --> 00:39:45,620
 or you should have somebody. This would be my proposal.

446
00:39:45,620 --> 00:39:49,180
 >> Take it as a small suggestion for people interested. Take up

447
00:39:49,180 --> 00:39:52,180
 open source projects, especially the ones that everyone uses, and

448
00:39:52,180 --> 00:39:54,260
 try to make small improvements there. You can have a huge

449
00:39:54,260 --> 00:39:56,260
 impact in this terms.

450
00:39:56,260 --> 00:39:57,260
 >> Yeah.

451
00:39:57,260 --> 00:40:03,500
 >> Hey, thanks for the talk. Really interesting, super cool.

452
00:40:03,500 --> 00:40:06,460
 Very naive question. So one of the things that we can actually

453
00:40:06,460 --> 00:40:09,660
 do with the increasing computing demand that we will have is

454
00:40:09,660 --> 00:40:14,180
 recapturing the energy in terms of heat. That's great. But very

455
00:40:14,180 --> 00:40:20,020
 naive question. Do we know what the fraction of low energy use of

456
00:40:20,020 --> 00:40:23,140
 our society is? So, you know, heat compared to electricity?

457
00:40:23,140 --> 00:40:29,980
 >> Oh, that's a good question. I have no numbers for that. But I

458
00:40:29,980 --> 00:40:35,620
 think it depends on the climate and the local situation. In

459
00:40:35,620 --> 00:40:40,180
 the isolation of the housing, so Germany is very bad on that.

460
00:40:40,180 --> 00:40:44,620
 Scandinavia is, because it's in a tropical area, obviously is

461
00:40:44,620 --> 00:40:50,780
 much better on that. So we have to improve on all levels. So

462
00:40:50,780 --> 00:40:56,060
 isolation, heat usage, better effective data center, something

463
00:40:56,060 --> 00:40:58,060
 like this. Otherwise it would not work.

464
00:40:58,060 --> 00:41:01,540
 >> As soon as we have to increase the exergy again by

465
00:41:01,540 --> 00:41:05,260
 converting heat back to electricity, it becomes really

466
00:41:05,260 --> 00:41:08,420
 inefficient. So, I mean, it's better than wasting it, of

467
00:41:08,420 --> 00:41:11,300
 course. It's free in that sense. But this is something that I

468
00:41:11,300 --> 00:41:12,820
 think we should look at.

469
00:41:12,820 --> 00:41:17,900
 >> Yeah. So it's not very efficient to transform this kind

470
00:41:17,900 --> 00:41:22,860
 of energy in electricity again. This is thermodynamics, yes.

471
00:41:22,860 --> 00:41:28,740
 Entropy is, heat is entropy on -- thank you.

472
00:41:28,740 --> 00:41:31,100
 >> Hi. Thanks for the amazing and much needed talk. I was

473
00:41:31,100 --> 00:41:34,540
 wondering, you identified the main drivers for electricity

474
00:41:34,540 --> 00:41:37,940
 consumption in data centers being compute power. I was

475
00:41:37,940 --> 00:41:42,780
 wondering if storage also generates usage.

476
00:41:42,780 --> 00:41:46,060
 >> Yes, it does. But it's proportional. So you have more

477
00:41:46,060 --> 00:41:50,700
 or less storage network is some proportional. And I did not

478
00:41:50,700 --> 00:41:52,980
 want to overload the talk with all these details. There are

479
00:41:52,980 --> 00:41:58,180
 details somewhere in the Internet. But, yeah, it also

480
00:41:58,180 --> 00:41:59,340
 contributes.

481
00:41:59,340 --> 00:42:04,740
 >> Hi. Thanks for the great talk. Are there any conflicts

482
00:42:04,740 --> 00:42:11,460
 expected between Mittenwalde and Grünheide, Tesla? Because

483
00:42:11,460 --> 00:42:13,940
 Tesla is also --

484
00:42:13,940 --> 00:42:16,620
 >> Tesla is also consumer. I think the real conflict is

485
00:42:16,620 --> 00:42:26,420
 between coal and Tesla. The coal consumption in the brown coal

486
00:42:26,420 --> 00:42:30,940
 region is much higher. The water consumption is much higher.

487
00:42:30,940 --> 00:42:34,420
 This has also an impact if they stop using water. In Berlin,

488
00:42:34,420 --> 00:42:40,740
 the river Spree will go down 75 percent, which is more or less

489
00:42:40,740 --> 00:42:43,580
 a catastrophe for Kreuzberg and southern Berlin.

490
00:42:43,580 --> 00:42:54,620
 >> Is it right to say the cooling is the most important and

491
00:42:54,620 --> 00:43:01,900
 most inefficient point on data center energy consumption?

492
00:43:01,900 --> 00:43:05,420
 >> It's one of the points because it does not compute

493
00:43:05,420 --> 00:43:10,580
 anything useful. So if you waste it, it's gone. If you can use

494
00:43:10,580 --> 00:43:15,260
 it, you have a saving of some percent. But if you can use the

495
00:43:15,260 --> 00:43:21,180
 entire CPU energy, then you have a major progress on saving

496
00:43:21,180 --> 00:43:22,180
 energy.

497
00:43:22,180 --> 00:43:26,980
 >> If we solve the cooling problem, we are very much

498
00:43:26,980 --> 00:43:30,380
 better. I mean, we can see in the comparison of Google and

499
00:43:30,380 --> 00:43:34,980
 Microsoft, there are different approaches to cool. And this is

500
00:43:34,980 --> 00:43:40,940
 the point on that we can work technically. And I'm working on

501
00:43:40,940 --> 00:43:48,060
 absorption cooling using this very low caloric heat. And

502
00:43:48,060 --> 00:43:51,100
 the temperatures are about, I don't know, 50, 60 degree

503
00:43:51,100 --> 00:43:55,980
 maximum. And this is something you cannot use in many ways.

504
00:43:55,980 --> 00:43:59,340
 Normally, you can use waste heat. And that's why I want to

505
00:43:59,340 --> 00:44:03,140
 ask you how you would say this kind of absorption and

506
00:44:03,140 --> 00:44:08,300
 absorption technologies are on the board, you know, to solve

507
00:44:08,300 --> 00:44:10,940
 this problem on a big scale.

508
00:44:10,940 --> 00:44:17,660
 >> This is the idea of this people. They have a heat

509
00:44:17,660 --> 00:44:20,340
 exchange. Normally, you get 35 degrees.

510
00:44:20,340 --> 00:44:21,340
 >> It's a heat pump.

511
00:44:21,340 --> 00:44:24,780
 >> It's a heat pump, yes. And this is a lower threshold for

512
00:44:24,780 --> 00:44:27,260
 being useful for warming houses.

513
00:44:27,260 --> 00:44:33,860
 >> Yeah, but what we do is we use the heat for cooling in a

514
00:44:33,860 --> 00:44:34,860
 thermal cooling technology.

515
00:44:34,860 --> 00:44:35,860
 >> If you can do that.

516
00:44:35,860 --> 00:44:39,180
 >> Yeah, this is thermal cooling. It's not the compression

517
00:44:39,180 --> 00:44:42,580
 cooling. It's another kind of process. You don't need

518
00:44:42,580 --> 00:44:49,060
 electricity. You use the heat to cool. The COP is 0.5. From

519
00:44:49,060 --> 00:44:53,460
 2 kilowatts heat, you can get 1 kilowatt cooling. And this is

520
00:44:53,460 --> 00:44:59,620
 what I think is a very good idea. But the -- yeah, I mean,

521
00:44:59,620 --> 00:45:04,740
 the mainstream in this question does not see this solution in

522
00:45:04,740 --> 00:45:06,180
 a way I would expect it.

523
00:45:06,180 --> 00:45:12,460
 >> I would look to dive deeper into that. But I have not seen

524
00:45:12,460 --> 00:45:15,580
 that so far. It's definitely interesting.

525
00:45:15,580 --> 00:45:21,900
 >> All right. So are you -- do you want to ask a question?

526
00:45:21,900 --> 00:45:22,900
 >> Yes.

527
00:45:22,900 --> 00:45:26,100
 >> Yeah. Still. Two minutes.

528
00:45:26,100 --> 00:45:29,340
 >> Actually two questions. One is if you use the heat for

529
00:45:29,340 --> 00:45:31,660
 heating houses, I always wonder what happens in the eight,

530
00:45:31,660 --> 00:45:34,780
 nine months maybe in Europe where you don't heat the

531
00:45:34,780 --> 00:45:37,700
 houses. What do you do? These numbers seem theoretically.

532
00:45:37,700 --> 00:45:42,060
 >> Yeah. So ideally you save the heat somewhere in the sand

533
00:45:42,060 --> 00:45:45,860
 and then consume it in the winter. So this is a Scandinavian

534
00:45:45,860 --> 00:45:51,260
 solution. And this means that you need a certain scale, that

535
00:45:51,260 --> 00:45:53,500
 you need a certain environment, and you have to rethink the

536
00:45:53,500 --> 00:45:56,900
 solution carefully. So if Finland can do it, probably we also

537
00:45:56,900 --> 00:46:02,620
 can do it. But at the end, you have to consider every aspect

538
00:46:02,620 --> 00:46:07,540
 of the environment, the local situation, and so on.

539
00:46:07,540 --> 00:46:12,940
 >> And question number two is PUE is very much focused on

540
00:46:12,940 --> 00:46:16,180
 hardware, right? I guess because they can measure the energy.

541
00:46:16,180 --> 00:46:20,380
 But as you said, you can save a lot of power by programming

542
00:46:20,380 --> 00:46:24,740
 more efficiently. Like you had the example 10 H-base servers

543
00:46:24,740 --> 00:46:30,260
 versus one. Do you know of any measurement or metric that I

544
00:46:30,260 --> 00:46:35,020
 can advertise my software for being efficient with?

545
00:46:35,020 --> 00:46:37,900
 >> So for Kubernetes, Microsoft has a framework, but I have not

546
00:46:37,900 --> 00:46:41,700
 looked deeper into it. So you can measure it. But you have to

547
00:46:41,700 --> 00:46:46,060
 also take into account that the business model of the data

548
00:46:46,060 --> 00:46:49,540
 centers is to sell you more computing power. So they are not

549
00:46:49,540 --> 00:46:53,660
 interested in you saving energy. They are interested in you

550
00:46:53,660 --> 00:46:59,220
 wasting energy. So this is some -- this goes into the opposite

551
00:46:59,220 --> 00:47:06,300
 direction. And I have no idea about an incentive, how we get

552
00:47:06,300 --> 00:47:09,740
 this together, that the data center providers help you to

553
00:47:09,740 --> 00:47:10,740
 save energy.

554
00:47:10,740 --> 00:47:13,740
 >> Thanks for the amazing talk.

555
00:47:13,740 --> 00:47:15,220
 >> Is it a short question?

556
00:47:15,220 --> 00:47:17,900
 >> It's super short. It's just one thought on this. If they

557
00:47:17,900 --> 00:47:20,660
 don't have any interest in defining something to measure

558
00:47:20,660 --> 00:47:21,660
 this, we should.

559
00:47:21,660 --> 00:47:24,900
 >> Yes. Because this -- effectively, this is your money.

560
00:47:24,900 --> 00:47:28,020
 This is cost. This is directly proportional to the money you

561
00:47:28,020 --> 00:47:30,860
 spend to the clouds.

562
00:47:30,860 --> 00:47:34,940
 >> Good. Then thank you all for coming. Thank you, Thomas, for

563
00:47:34,940 --> 00:47:35,940
 answering so many questions.

564
00:47:35,940 --> 00:47:35,940
 [ Applause ]

565
00:47:36,460 --> 00:47:37,460
 [ Music ]

566
00:47:37,500 --> 00:47:38,500
 [ Music ]

567
00:47:38,540 --> 00:47:39,540
 [ Applause ]

568
00:47:39,580 --> 00:47:40,580
 [ Music ]

569
00:47:40,620 --> 00:47:41,620
 [ Applause ]

570
00:47:41,620 --> 00:47:44,200
 (gentle music)

571
00:47:44,200 --> 00:47:45,980
 (electronic chimes)