Designing better semiconductor chips: NP hard problems and forever chemicals

1
00:00:07,919 --> 00:00:10,719
Hello, and welcome to the Physics World weekly

2
00:00:10,719 --> 00:00:12,740
podcast. I'm Hamish Johnston.

3
00:00:13,294 --> 00:00:15,695
In this episode, I'm joined by my colleague

4
00:00:15,695 --> 00:00:16,835
Margaret Harris,

5
00:00:17,135 --> 00:00:19,934
who's recorded a series of interviews for the

6
00:00:19,934 --> 00:00:20,434
podcast

7
00:00:20,734 --> 00:00:22,435
at this year's Heidelberg

8
00:00:22,894 --> 00:00:24,114
Laureate Forum.

9
00:00:24,574 --> 00:00:26,675
We're going to hear two of those conversations

10
00:00:27,054 --> 00:00:27,554
shortly,

11
00:00:27,969 --> 00:00:30,869
which both focus on the challenges of designing

12
00:00:31,089 --> 00:00:31,910
and developing

13
00:00:32,450 --> 00:00:33,509
computer chips.

14
00:00:34,049 --> 00:00:36,229
Hi, Margaret. Welcome to the podcast.

15
00:00:36,770 --> 00:00:37,590
Hi, Hamish.

16
00:00:38,289 --> 00:00:40,530
Margaret, before we hear those interviews, can you

17
00:00:40,530 --> 00:00:42,869
tell us a little bit about the Heidelberg

18
00:00:43,090 --> 00:00:44,075
Laureate Forum?

19
00:00:44,935 --> 00:00:47,094
Yeah. So the Heidelberg Laureate Forum is a

20
00:00:47,094 --> 00:00:49,435
scientific meeting that's held every year in Heidelberg,

21
00:00:49,575 --> 00:00:52,375
Germany, hence the name. And its aim is

22
00:00:52,375 --> 00:00:55,354
to foster connections between notable figures in computer

23
00:00:55,414 --> 00:00:56,554
science and mathematics,

24
00:00:57,469 --> 00:00:59,630
and then with early career researchers from around

25
00:00:59,630 --> 00:01:00,289
the world.

26
00:01:00,750 --> 00:01:03,229
It's modeled on the Lindau Nobel Laureate meeting,

27
00:01:03,229 --> 00:01:05,390
which does something similar for the Nobel Prize

28
00:01:05,390 --> 00:01:07,810
granting fields of medicine, physics, and chemistry,

29
00:01:08,269 --> 00:01:10,750
but there's no Nobel Prize for computer science

30
00:01:10,750 --> 00:01:13,185
or for mathematics. So instead, the laureates who

31
00:01:13,185 --> 00:01:16,064
come to Heidelberg have won major awards such

32
00:01:16,064 --> 00:01:18,224
as the Abel Prize and the Fields Medal

33
00:01:18,224 --> 00:01:18,804
in Mathematics

34
00:01:19,424 --> 00:01:22,144
or the ACM AM Turing Award for computer

35
00:01:22,144 --> 00:01:22,644
science.

36
00:01:23,504 --> 00:01:25,024
I I have to say, Margaret, I'm a

37
00:01:25,024 --> 00:01:27,379
bit ignorant of of both mathematics

38
00:01:27,840 --> 00:01:31,539
and computer science as sort of academic fields.

39
00:01:31,920 --> 00:01:34,159
Is is there a lot of crossover between

40
00:01:34,159 --> 00:01:35,759
those two? I mean I mean, I would

41
00:01:35,759 --> 00:01:38,974
have assumed so, but, I could be wrong.

42
00:01:38,974 --> 00:01:40,655
Yeah. I mean, you know, obviously, these are

43
00:01:40,655 --> 00:01:42,814
two fields that are very closely related to

44
00:01:42,814 --> 00:01:43,775
physics, and,

45
00:01:44,174 --> 00:01:46,194
so I would say that there is simultaneously

46
00:01:47,375 --> 00:01:49,534
less and more overlap between the two and

47
00:01:49,534 --> 00:01:51,155
with physics than you might think.

48
00:01:52,010 --> 00:01:54,010
So several of the computer scientists I spoke

49
00:01:54,010 --> 00:01:56,010
to during the format, which I attended back

50
00:01:56,010 --> 00:01:56,750
in September,

51
00:01:57,290 --> 00:01:59,129
they said that that being there and talking

52
00:01:59,129 --> 00:01:59,790
to mathematicians

53
00:02:00,250 --> 00:02:01,770
had actually convinced them that they need to

54
00:02:01,770 --> 00:02:02,750
learn more math

55
00:02:03,049 --> 00:02:05,290
before they can make a useful contribution to

56
00:02:05,290 --> 00:02:05,790
future

57
00:02:06,484 --> 00:02:07,844
discussions. And I would sort of back that

58
00:02:07,844 --> 00:02:08,965
up. You know, I spoke to a couple

59
00:02:08,965 --> 00:02:11,284
of pure mathematicians there who told me what

60
00:02:11,284 --> 00:02:12,645
they're working on, and I was like, I

61
00:02:12,645 --> 00:02:14,985
understand one word and three of those things.

62
00:02:15,284 --> 00:02:17,145
So, you know, you can't have

63
00:02:17,444 --> 00:02:18,724
much of a discussion if you don't have

64
00:02:18,724 --> 00:02:20,645
a common vocabulary. So I think there is

65
00:02:20,645 --> 00:02:21,509
some work there. And

66
00:02:22,229 --> 00:02:22,949
And I'm sure, you know, there are mathematicians

67
00:02:22,949 --> 00:02:24,870
who felt the same way about computer scientists.

68
00:02:24,870 --> 00:02:26,069
You know, it's not as if it's a

69
00:02:26,069 --> 00:02:27,289
one way street there.

70
00:02:27,829 --> 00:02:29,189
But that said, a lot of talks at

71
00:02:29,189 --> 00:02:31,509
this year's forum focused on machine learning and

72
00:02:31,509 --> 00:02:32,409
artificial intelligence

73
00:02:32,709 --> 00:02:33,530
or AI,

74
00:02:34,165 --> 00:02:36,645
which is having major impacts on so many

75
00:02:36,645 --> 00:02:38,805
areas of science, including physics as well as

76
00:02:38,805 --> 00:02:41,925
mathematics and computer science. And there's definitely some

77
00:02:41,925 --> 00:02:43,145
common ground there.

78
00:02:43,444 --> 00:02:44,885
And then, of course, there's a whole branch

79
00:02:44,885 --> 00:02:48,449
of mathematics specifically dedicated to computer science. You

80
00:02:48,610 --> 00:02:49,729
know, if you think back to the middle

81
00:02:49,729 --> 00:02:52,610
of the twentieth century when computer science was

82
00:02:52,610 --> 00:02:53,509
getting its start,

83
00:02:54,050 --> 00:02:56,129
you had work by scientists like John von

84
00:02:56,129 --> 00:02:58,229
Neumann and, of course, Alan Turing himself,

85
00:02:58,530 --> 00:03:00,209
you know, really laying the foundations of the

86
00:03:00,209 --> 00:03:01,989
field. So it's definitely a lot of crossover

87
00:03:02,050 --> 00:03:03,830
even if it's not complete overlap.

88
00:03:04,465 --> 00:03:06,245
So you you mentioned the Lindau

89
00:03:06,544 --> 00:03:07,044
meetings.

90
00:03:07,504 --> 00:03:10,064
I I went to one many years ago,

91
00:03:10,064 --> 00:03:12,385
and and I thought it was fantastic. And

92
00:03:12,385 --> 00:03:14,544
it was fantastic because of the people who

93
00:03:14,544 --> 00:03:16,724
were there. You had Nobel laureates,

94
00:03:17,185 --> 00:03:19,490
on one hand, and then you had lots

95
00:03:19,490 --> 00:03:22,710
of, well, students and early career people there

96
00:03:22,770 --> 00:03:25,170
sort of mixing and chatting. And, you know,

97
00:03:25,170 --> 00:03:26,849
I thought it was a really it was

98
00:03:26,849 --> 00:03:27,830
a really dynamic

99
00:03:28,129 --> 00:03:28,629
event.

100
00:03:29,170 --> 00:03:31,170
So so who is in in Heidelberg? Is

101
00:03:31,170 --> 00:03:32,930
it the the sort of same idea of

102
00:03:33,615 --> 00:03:35,615
Yeah. Very much so. So the laureates of

103
00:03:35,615 --> 00:03:38,254
this year's forum included, Vint Cerf, who's known

104
00:03:38,254 --> 00:03:40,094
as the father of the Internet, with his

105
00:03:40,094 --> 00:03:42,655
role in developing the TCP ISP system used

106
00:03:42,655 --> 00:03:43,395
in email.

107
00:03:44,094 --> 00:03:45,854
I wrote a blog post for Physics World

108
00:03:45,854 --> 00:03:48,700
about Cerf's talk at the forum on the

109
00:03:49,000 --> 00:03:51,240
on the longevity of digital information, and you

110
00:03:51,240 --> 00:03:52,939
can read it on the Physics World website.

111
00:03:53,879 --> 00:03:55,659
Other people I met there were the cryptographers,

112
00:03:55,879 --> 00:03:58,439
Whitfield Diffie and Martin Hellman, who pioneered the

113
00:03:58,439 --> 00:04:01,305
public key cryptography method that's used to keep

114
00:04:01,444 --> 00:04:03,844
all your data secure online, especially actually, there

115
00:04:03,924 --> 00:04:05,685
there's also the RSA algorithm, which is a

116
00:04:05,685 --> 00:04:07,525
big part of it. But Whitfield, Diffie, and

117
00:04:07,525 --> 00:04:08,884
Martin Hellman were the ones who were at

118
00:04:08,884 --> 00:04:09,944
this particular conference.

119
00:04:10,724 --> 00:04:12,164
And then as you said, there's the early

120
00:04:12,164 --> 00:04:15,050
career researchers. And at the Heidelberg Laureate Forum,

121
00:04:15,050 --> 00:04:17,129
they were a hugely diverse bunch. You know,

122
00:04:17,129 --> 00:04:19,290
they came from all over the world and

123
00:04:19,290 --> 00:04:21,310
ranging from undergraduate students to postdocs.

124
00:04:22,250 --> 00:04:24,009
They have to apply to attend, and it

125
00:04:24,009 --> 00:04:26,750
must be a pretty rigorous application process because

126
00:04:27,375 --> 00:04:29,774
everyone I met was just great fun to

127
00:04:29,774 --> 00:04:32,435
talk to. They're really impressive in their enthusiasm

128
00:04:32,654 --> 00:04:34,974
for the research and actually very good explain

129
00:04:35,055 --> 00:04:37,375
at explaining it, even to a physicist like

130
00:04:37,375 --> 00:04:39,134
me who doesn't come from a really strong

131
00:04:39,134 --> 00:04:41,074
computational or mathematical background.

132
00:04:41,910 --> 00:04:44,889
And Margaret, we've got two, of your conversations,

133
00:04:45,670 --> 00:04:47,529
queued up and ready to go,

134
00:04:48,389 --> 00:04:49,610
later in the podcast.

135
00:04:49,990 --> 00:04:51,769
Who who did you speak to?

136
00:04:52,230 --> 00:04:54,470
So my goal at the Heidelberg Laureate Forum

137
00:04:54,470 --> 00:04:56,224
was to find out more about the aspects

138
00:04:56,224 --> 00:04:58,704
of computer science and mathematics that relate most

139
00:04:58,704 --> 00:05:00,404
closely to physics or,

140
00:05:00,865 --> 00:05:02,224
to put it in a slightly different way,

141
00:05:02,224 --> 00:05:03,745
the ones that are of most interest to

142
00:05:03,745 --> 00:05:04,245
physicists.

143
00:05:04,784 --> 00:05:06,224
And with that in mind, I spoke to

144
00:05:06,224 --> 00:05:09,425
two PhD students whose research focuses in different

145
00:05:09,425 --> 00:05:09,925
ways

146
00:05:10,579 --> 00:05:12,339
on kind of the nitty gritty factors that

147
00:05:12,339 --> 00:05:14,680
go in designing and developing computer chips.

148
00:05:15,220 --> 00:05:17,939
And these computer chips obviously play hugely important

149
00:05:17,939 --> 00:05:20,740
roles in many aspects of physicists' lives, whether

150
00:05:20,740 --> 00:05:22,500
that's making your own chips as part of

151
00:05:22,500 --> 00:05:25,375
a scientific prototyping process or even just, you

152
00:05:25,375 --> 00:05:27,154
know, owning consumer electronics.

153
00:05:27,694 --> 00:05:29,475
It goes the whole gamut there.

154
00:05:30,415 --> 00:05:31,855
And so who who are we going to

155
00:05:31,855 --> 00:05:32,834
hear from first?

156
00:05:33,375 --> 00:05:35,055
Okay. So first up, we'll be hearing from

157
00:05:35,055 --> 00:05:37,855
Maryam Elgamal, who's doing a PhD at Harvard

158
00:05:37,855 --> 00:05:39,154
University in The US.

159
00:05:39,569 --> 00:05:42,149
She's working on the design of environmentally sustainable

160
00:05:42,210 --> 00:05:43,189
computing systems,

161
00:05:43,649 --> 00:05:45,569
and I began by asking her what got

162
00:05:45,569 --> 00:05:47,029
her interested in this field.

163
00:05:55,384 --> 00:05:57,004
First, thank you for having me.

164
00:05:57,464 --> 00:05:58,985
I think what really got me into the

165
00:05:58,985 --> 00:06:01,485
field, I was I came into my PhD

166
00:06:01,785 --> 00:06:03,544
not really knowing what I wanted to work

167
00:06:03,544 --> 00:06:05,384
on. I just knew that I really liked

168
00:06:05,384 --> 00:06:05,884
computing,

169
00:06:06,264 --> 00:06:08,264
and I was just exploring different kind of

170
00:06:08,264 --> 00:06:11,089
projects. And I was lucky, I guess, because

171
00:06:11,089 --> 00:06:13,250
my advisers were just, like, giving me all

172
00:06:13,250 --> 00:06:15,009
of these different kind of projects that are

173
00:06:15,009 --> 00:06:17,329
available. And one of them really stuck out

174
00:06:17,329 --> 00:06:19,729
to me, which was the looking at, like,

175
00:06:19,729 --> 00:06:22,529
environmentally sustainable computing systems and and the the

176
00:06:22,529 --> 00:06:24,310
environmental impact. And I

177
00:06:24,714 --> 00:06:27,375
have always had an interest in environmental sustainability,

178
00:06:28,235 --> 00:06:30,095
even, like, prior to my undergraduate

179
00:06:30,555 --> 00:06:31,055
degree.

180
00:06:31,435 --> 00:06:33,834
And it was just this one project stuck

181
00:06:33,834 --> 00:06:35,194
out to me, and I started working on

182
00:06:35,194 --> 00:06:36,175
it with a senior

183
00:06:36,714 --> 00:06:38,714
student at the time. I was currently a

184
00:06:38,714 --> 00:06:39,214
professor,

185
00:06:40,020 --> 00:06:40,759
at Cornell,

186
00:06:41,220 --> 00:06:41,960
Udut Gupta.

187
00:06:42,660 --> 00:06:44,180
And so I started working with him and

188
00:06:44,180 --> 00:06:46,360
I really enjoyed that project and then just

189
00:06:46,660 --> 00:06:48,500
decided to continue in that area for the

190
00:06:48,500 --> 00:06:50,100
rest of my PhD. And now I'm in

191
00:06:50,100 --> 00:06:52,899
my fifth year still working on designing computing

192
00:06:52,899 --> 00:06:54,600
systems for environmental sustainability.

193
00:06:55,134 --> 00:06:57,134
I think what's really also exciting about this

194
00:06:57,134 --> 00:07:00,334
area is how it's an emerging research area.

195
00:07:00,334 --> 00:07:02,895
So it's pretty new compared to other areas

196
00:07:02,895 --> 00:07:05,154
in like hardware design or computer architecture.

197
00:07:05,615 --> 00:07:07,615
And so that also gets me really excited

198
00:07:07,615 --> 00:07:09,600
to just like, there's so much potential, so

199
00:07:09,839 --> 00:07:10,959
many things that we can do and so

200
00:07:10,959 --> 00:07:12,420
many things that we need to explore.

201
00:07:12,800 --> 00:07:16,000
What does sustainability mean for computing systems? What

202
00:07:16,000 --> 00:07:17,520
are the main things you need to consider

203
00:07:17,520 --> 00:07:20,420
when you're evaluating how sustainable a computer system

204
00:07:20,480 --> 00:07:20,980
is?

205
00:07:21,600 --> 00:07:23,060
So when it comes to

206
00:07:23,495 --> 00:07:26,134
sustainability for computing, there are multiple aspects for

207
00:07:26,134 --> 00:07:26,634
it.

208
00:07:27,334 --> 00:07:27,834
Traditionally,

209
00:07:28,214 --> 00:07:29,914
when we design computing systems,

210
00:07:30,774 --> 00:07:32,935
it has always been like the primary metrics

211
00:07:32,935 --> 00:07:34,694
that we would look at are like power,

212
00:07:34,694 --> 00:07:35,914
performance area.

213
00:07:36,329 --> 00:07:37,930
And, like, in the past two decades, there

214
00:07:37,930 --> 00:07:39,069
has been so much

215
00:07:39,449 --> 00:07:40,669
work and so much

216
00:07:40,970 --> 00:07:42,970
progress when it comes to the energy efficiency

217
00:07:42,970 --> 00:07:44,810
of computing systems. So how can we design

218
00:07:44,810 --> 00:07:46,110
computing systems that

219
00:07:46,410 --> 00:07:48,410
deliver the most work or the highest performance

220
00:07:48,410 --> 00:07:50,569
with the minimal energy possible? So there has

221
00:07:50,569 --> 00:07:52,029
been a lot of work in the area.

222
00:07:52,144 --> 00:07:53,985
But when it comes to the actual, like,

223
00:07:53,985 --> 00:07:56,785
sustainability of a computing system, it's not only

224
00:07:56,785 --> 00:07:58,004
about energy efficiency.

225
00:07:58,384 --> 00:08:00,225
It's also about, like, looking at the total

226
00:08:00,225 --> 00:08:02,064
carbon of the computing system. And what that

227
00:08:02,064 --> 00:08:04,705
means is that we're looking at the carbon

228
00:08:04,705 --> 00:08:07,019
footprint due to the use. So you can

229
00:08:07,019 --> 00:08:08,939
think that energy efficiency is a subset of

230
00:08:08,939 --> 00:08:10,779
that, but also we want to look at

231
00:08:10,779 --> 00:08:12,779
the carbon footprint due to the manufacturing of

232
00:08:12,779 --> 00:08:14,939
the system itself. And what we're finding is

233
00:08:14,939 --> 00:08:17,279
that the carbon footprint of computing systems

234
00:08:17,740 --> 00:08:20,220
from a manufacturing perspective can actually be a

235
00:08:20,220 --> 00:08:21,919
huge portion of the carbon footprint.

236
00:08:22,404 --> 00:08:23,925
It up it depends on the domain. So

237
00:08:23,925 --> 00:08:25,285
if you're looking at a data center, it

238
00:08:25,285 --> 00:08:27,544
could look like 50% of the carbon footprint

239
00:08:27,604 --> 00:08:31,125
is due to, like, the operational carbon, so

240
00:08:31,125 --> 00:08:34,289
use. And the other 50%, around 50% is

241
00:08:34,370 --> 00:08:36,629
going to be due to the actual manufacturing

242
00:08:36,690 --> 00:08:38,769
and the infrastructure of the data center. When

243
00:08:38,769 --> 00:08:41,169
you're looking at something like mobile devices or,

244
00:08:41,169 --> 00:08:44,129
like, more broad like, more broadly speaking, consumer

245
00:08:44,129 --> 00:08:47,350
devices, that would more look like 75%

246
00:08:47,735 --> 00:08:48,475
being manufacturing.

247
00:08:49,254 --> 00:08:52,294
And so that's something that hardware designers and

248
00:08:52,294 --> 00:08:55,574
computer engineers haven't typically looked at. And we're

249
00:08:55,574 --> 00:08:58,235
really trying to see how we can consider

250
00:08:58,454 --> 00:09:01,009
this total carbon and trade that off with

251
00:09:01,009 --> 00:09:03,409
all of these other conventional metrics of power,

252
00:09:03,409 --> 00:09:04,549
performance, area,

253
00:09:04,929 --> 00:09:07,429
and costs that have been traditionally done. So,

254
00:09:07,970 --> 00:09:09,830
this is how we're looking at the sustainability.

255
00:09:09,970 --> 00:09:12,529
And there are other aspects of sustainability beyond

256
00:09:12,529 --> 00:09:14,690
carbon footprint that some people had that we're

257
00:09:14,690 --> 00:09:16,664
starting to explore. And some people have also

258
00:09:16,985 --> 00:09:19,704
explored in other places, which is, like, water

259
00:09:19,704 --> 00:09:22,504
consumption, whether that's in the fabrication or, like,

260
00:09:22,504 --> 00:09:24,264
if you're using that for, like, cooling a

261
00:09:24,264 --> 00:09:26,985
data center. Another thing that I'm also looking

262
00:09:26,985 --> 00:09:29,644
at is, like, the forever chemicals,

263
00:09:30,480 --> 00:09:31,539
more like colloquially

264
00:09:32,159 --> 00:09:34,320
would be called forever chemicals, but it's also

265
00:09:34,320 --> 00:09:36,419
called PFAS, which is the pro and polyfluorocal

266
00:09:36,720 --> 00:09:37,220
substances.

267
00:09:38,399 --> 00:09:40,320
So the problem with these chemicals is that

268
00:09:40,320 --> 00:09:41,139
they're synthetic

269
00:09:41,519 --> 00:09:42,019
and

270
00:09:42,559 --> 00:09:44,445
that makes them very bioaccumulative.

271
00:09:45,544 --> 00:09:47,945
Sometimes toxic, sometimes not. And it's a very

272
00:09:47,945 --> 00:09:49,164
broad class of chemicals.

273
00:09:49,625 --> 00:09:51,804
And they're very much widely used in semiconductor

274
00:09:52,024 --> 00:09:54,745
manufacturing. They're pretty much almost in every like,

275
00:09:54,745 --> 00:09:57,779
used for every single integrated circuit manufacturing and

276
00:09:57,779 --> 00:09:58,839
fabrication process.

277
00:09:59,299 --> 00:10:00,899
Not in every single process, but it is

278
00:10:00,899 --> 00:10:03,559
used to make the actual integrated circuit.

279
00:10:03,940 --> 00:10:05,940
And so we're looking at how we can

280
00:10:05,940 --> 00:10:08,339
quantify that, which turns out to be even

281
00:10:08,339 --> 00:10:10,199
harder than cover for carbon footprint.

282
00:10:10,535 --> 00:10:12,295
And then also how we can start, like,

283
00:10:12,295 --> 00:10:14,695
looking at techniques to start optimizing that and

284
00:10:14,695 --> 00:10:15,674
reducing it.

285
00:10:15,975 --> 00:10:17,355
So that's like sustainability

286
00:10:17,735 --> 00:10:19,894
from a high level perspective when it comes

287
00:10:19,894 --> 00:10:20,554
to computing.

288
00:10:21,335 --> 00:10:23,915
You talk about making computing systems more sustainable.

289
00:10:24,295 --> 00:10:26,154
What would that actually look like in practice?

290
00:10:26,549 --> 00:10:28,789
What are some things that manufacturers or people

291
00:10:28,789 --> 00:10:31,350
who run data centers, even consumers who have

292
00:10:31,350 --> 00:10:33,190
phones, what are some of the things we

293
00:10:33,190 --> 00:10:35,370
could do to make that more sustainable?

294
00:10:36,389 --> 00:10:38,570
So I like to think about sustainability

295
00:10:39,029 --> 00:10:39,750
from two

296
00:10:40,504 --> 00:10:42,345
I like to classify it in two separate

297
00:10:42,345 --> 00:10:42,845
ways.

298
00:10:43,225 --> 00:10:45,464
One is the quantification aspect and one is

299
00:10:45,464 --> 00:10:47,644
the optimization aspect. So

300
00:10:47,945 --> 00:10:49,625
first tip for us to be able to

301
00:10:49,625 --> 00:10:51,625
know how much we need to reduce or

302
00:10:51,625 --> 00:10:53,544
optimize something, we need to measure it and

303
00:10:53,544 --> 00:10:55,559
we need to quantify it. So a lot

304
00:10:55,559 --> 00:10:57,480
of work has been happening in the past

305
00:10:57,480 --> 00:10:59,879
few years on, like, how can we quantify

306
00:10:59,879 --> 00:11:01,820
the carbon footprint of computing systems?

307
00:11:02,120 --> 00:11:03,720
So talking about what is like, what is

308
00:11:03,800 --> 00:11:05,639
how much is like going to manufacturing? How

309
00:11:05,639 --> 00:11:06,940
much is going to operational?

310
00:11:07,240 --> 00:11:09,240
And then the next part is, okay, how

311
00:11:09,240 --> 00:11:10,620
can we also optimize

312
00:11:11,215 --> 00:11:13,774
these systems? So maybe, like, should we be

313
00:11:13,774 --> 00:11:14,514
using, like,

314
00:11:14,894 --> 00:11:16,894
GPUs or should we be using CPUs in

315
00:11:16,894 --> 00:11:18,335
a data center and, like, in a more

316
00:11:18,335 --> 00:11:20,414
nuanced kind of context and depending obviously on

317
00:11:20,414 --> 00:11:22,254
the applications that you're running in your data

318
00:11:22,254 --> 00:11:22,754
center?

319
00:11:23,269 --> 00:11:26,070
Something, for example, for, like, PFAS, it could

320
00:11:26,070 --> 00:11:28,570
be more, can we reduce the number of

321
00:11:29,110 --> 00:11:31,210
metal layers that you're using to manufacture

322
00:11:31,669 --> 00:11:33,529
an integrated circuit? So

323
00:11:33,910 --> 00:11:36,054
that's, like, just for context, it's like you

324
00:11:36,295 --> 00:11:37,654
when you make a chip, you have your

325
00:11:37,654 --> 00:11:40,134
transistors, and then you start manufacturing, like, mental

326
00:11:40,134 --> 00:11:41,995
layers where you start doing the routing,

327
00:11:42,695 --> 00:11:44,855
to connect these transistors together so that you

328
00:11:44,855 --> 00:11:48,315
can your chip can actually work. And so,

329
00:11:48,375 --> 00:11:51,389
typically, like, these PFAS are included almost in

330
00:11:51,389 --> 00:11:53,470
every single step to do, like, the what

331
00:11:53,470 --> 00:11:54,450
is called the photolithography.

332
00:11:54,909 --> 00:11:57,389
So the patterning of the actual, like, routing

333
00:11:57,389 --> 00:12:00,589
in the circuit on the, chip. And so

334
00:12:00,589 --> 00:12:02,350
if we can reduce this number of metal

335
00:12:02,350 --> 00:12:04,269
layers, maybe that's one thing that we can

336
00:12:04,269 --> 00:12:07,184
also reduce the forever chemicals or PFAS with

337
00:12:07,184 --> 00:12:08,485
that are used in manufacturing.

338
00:12:09,264 --> 00:12:12,004
Fabrication facilities have started looking into

339
00:12:12,625 --> 00:12:15,264
how they can reduce, like, recycle maybe some

340
00:12:15,264 --> 00:12:17,345
of the water that they're using. Some have

341
00:12:17,345 --> 00:12:19,184
looked into recycling some of the metals that

342
00:12:19,184 --> 00:12:19,679
they

343
00:12:20,000 --> 00:12:22,820
use. So for example, I think, Taiwan Semiconductor

344
00:12:23,039 --> 00:12:23,539
Manufacturing,

345
00:12:24,320 --> 00:12:24,820
TSMC,

346
00:12:25,440 --> 00:12:28,639
they have worked in on basically recycling the

347
00:12:28,639 --> 00:12:30,899
copper that they're using in the manufacturing

348
00:12:31,440 --> 00:12:34,565
and then taking it, recycling it, not necessarily

349
00:12:34,705 --> 00:12:37,985
into electronic grade copper, which is a challenge

350
00:12:37,985 --> 00:12:41,125
because the materials that you typically use for

351
00:12:41,745 --> 00:12:44,144
electronics or semiconductor manufacturing need to be very

352
00:12:44,144 --> 00:12:44,884
high purity.

353
00:12:45,330 --> 00:12:47,410
But instead of having to recycle it back

354
00:12:47,410 --> 00:12:49,170
into the same fab like, to use it

355
00:12:49,170 --> 00:12:51,090
in the same fabrication facility, you can recycle

356
00:12:51,090 --> 00:12:52,850
it and then give it to some other

357
00:12:52,850 --> 00:12:54,450
industry that may not need that kind of

358
00:12:54,450 --> 00:12:57,269
higher purity copper that you may need.

359
00:12:57,730 --> 00:13:00,394
It can also be utilized elsewhere and still

360
00:13:00,695 --> 00:13:02,634
benefit the whole sustainable sustainability

361
00:13:02,935 --> 00:13:05,975
aspect. Making copper wires or copper roofs or

362
00:13:05,975 --> 00:13:08,535
something like that. Exactly. Yeah. And so these

363
00:13:08,535 --> 00:13:09,355
kind of things.

364
00:13:11,415 --> 00:13:13,370
I guess any change is going to involve

365
00:13:13,370 --> 00:13:15,370
some level of cost in terms of going

366
00:13:15,370 --> 00:13:17,230
from old technologies to new ones.

367
00:13:17,610 --> 00:13:20,809
And traditionally in engineering, you can basically there's

368
00:13:20,809 --> 00:13:23,470
this three way trade off between being cheap,

369
00:13:23,769 --> 00:13:26,009
being fast, and being high quality, you know,

370
00:13:26,009 --> 00:13:27,629
and you get to pick only two.

371
00:13:28,375 --> 00:13:30,134
Are those kinds of trade offs also going

372
00:13:30,134 --> 00:13:30,954
on in the sustainability

373
00:13:31,334 --> 00:13:34,534
sector? Like, is sustainability maybe one of those

374
00:13:34,534 --> 00:13:35,894
criteria that you have to pick just a

375
00:13:35,894 --> 00:13:38,294
limited number of them? It really depends on

376
00:13:38,294 --> 00:13:40,394
the domain and really depends on the application.

377
00:13:41,009 --> 00:13:43,730
There are cases when you find computing systems

378
00:13:43,730 --> 00:13:45,269
that like, the

379
00:13:45,649 --> 00:13:48,529
energy efficient and the highest performing system could

380
00:13:48,529 --> 00:13:50,309
also be the carbon efficient system.

381
00:13:51,009 --> 00:13:53,250
But in other cases, you may find that,

382
00:13:53,250 --> 00:13:54,309
oh, it's actually

383
00:13:54,690 --> 00:13:56,975
the opposite. So it's like, pick which one

384
00:13:56,975 --> 00:13:58,894
you want. And so this whole trade off

385
00:13:58,894 --> 00:14:00,815
space is something that is very exciting to

386
00:14:00,815 --> 00:14:02,654
me in in in my research area, which

387
00:14:02,654 --> 00:14:04,815
is how can we trade this off? And

388
00:14:04,815 --> 00:14:06,975
that really depends on what your target is

389
00:14:06,975 --> 00:14:07,714
as a designer.

390
00:14:08,654 --> 00:14:10,980
So for example, if you if as if

391
00:14:10,980 --> 00:14:13,860
you're designing something that, say, you just need

392
00:14:13,860 --> 00:14:15,879
to meet a specific performance criteria,

393
00:14:16,419 --> 00:14:18,259
once you meet that performance, you probably just

394
00:14:18,259 --> 00:14:20,179
want to lower the carbon footprint. But if

395
00:14:20,179 --> 00:14:22,259
you're trying to design a system that you

396
00:14:22,259 --> 00:14:24,019
want to make sure that you reach the

397
00:14:24,019 --> 00:14:25,879
highest speed and performance possible,

398
00:14:26,445 --> 00:14:26,945
then

399
00:14:27,404 --> 00:14:27,904
probably

400
00:14:28,365 --> 00:14:30,045
you want to figure out a trade off

401
00:14:30,045 --> 00:14:31,825
that would still give you a good performance,

402
00:14:32,045 --> 00:14:33,965
but is not too bad when it comes

403
00:14:33,965 --> 00:14:34,945
to sustainability.

404
00:14:35,565 --> 00:14:37,884
So that trade off definitely makes it I

405
00:14:37,884 --> 00:14:39,804
think it's an extra dimension that we need

406
00:14:39,804 --> 00:14:41,450
to account for now when it comes to

407
00:14:41,450 --> 00:14:44,169
a sustainable design compared to just, like, energy

408
00:14:44,169 --> 00:14:46,409
efficient design or high performing design as it

409
00:14:46,409 --> 00:14:47,789
traditionally is. So

410
00:14:48,250 --> 00:14:50,089
In your research so far, what would you

411
00:14:50,089 --> 00:14:52,329
say has been something that's really surprised you?

412
00:14:52,329 --> 00:14:54,190
Something you didn't expect to find?

413
00:14:54,934 --> 00:14:57,115
Something you didn't expect to find. I

414
00:14:57,975 --> 00:15:00,475
was very surprised to see how

415
00:15:01,014 --> 00:15:04,714
the PFAS used in semiconductor manufacturing are actually

416
00:15:05,254 --> 00:15:06,235
they're pretty essential

417
00:15:06,909 --> 00:15:09,950
in many ways. And, the most surprising part

418
00:15:09,950 --> 00:15:11,870
to me was most of the PFAS used

419
00:15:11,870 --> 00:15:13,330
in semiconductor manufacturing

420
00:15:14,029 --> 00:15:16,750
do not actually stay on chip. They mostly

421
00:15:16,750 --> 00:15:17,570
go to wastewater,

422
00:15:18,029 --> 00:15:20,269
but it's something that you absolutely need to

423
00:15:20,269 --> 00:15:23,044
make that integrated circuit, and yet it's not

424
00:15:23,044 --> 00:15:24,264
on the actual chip.

425
00:15:24,644 --> 00:15:27,044
This whole PFAS thing makes it makes me

426
00:15:27,044 --> 00:15:28,105
realize that

427
00:15:28,804 --> 00:15:29,945
some of the sustainability

428
00:15:30,325 --> 00:15:32,325
things that we really need to think about

429
00:15:32,325 --> 00:15:34,169
in computing is not just

430
00:15:35,209 --> 00:15:36,029
what actually

431
00:15:36,809 --> 00:15:38,829
remains in the end product of computing,

432
00:15:39,289 --> 00:15:41,690
but the process itself, if we start looking

433
00:15:41,690 --> 00:15:42,970
into it, we might find that there are

434
00:15:42,970 --> 00:15:44,809
so many things that we that there are

435
00:15:44,809 --> 00:15:46,970
so many opportunities that we can do to

436
00:15:46,970 --> 00:15:49,470
improve the sustainability of computing in general.

437
00:15:50,294 --> 00:15:52,054
What's the next stage for this project? What

438
00:15:52,054 --> 00:15:53,414
are you working on now? What do you

439
00:15:53,414 --> 00:15:55,674
hope to achieve before you finish your PhD?

440
00:15:56,375 --> 00:15:58,534
So one of the things that really excite

441
00:15:58,534 --> 00:16:00,934
me and interest me in the sustainability space

442
00:16:00,934 --> 00:16:03,735
right now is thinking about what are better

443
00:16:03,735 --> 00:16:06,709
optimization techniques that we can do to actually

444
00:16:07,649 --> 00:16:09,350
help designers start

445
00:16:09,889 --> 00:16:11,029
doing actual

446
00:16:11,409 --> 00:16:12,949
sustainability aware design.

447
00:16:13,329 --> 00:16:15,089
And one of the main challenges in the

448
00:16:15,089 --> 00:16:17,649
sustainability area is that there's so much uncertainty

449
00:16:17,649 --> 00:16:20,004
in the data. And this uncertainty comes from

450
00:16:20,004 --> 00:16:22,004
so many, like, different shapes and forms. One

451
00:16:22,004 --> 00:16:24,085
of them is, like, one, there's lack of

452
00:16:24,085 --> 00:16:26,004
data. And two, there's just a lot of

453
00:16:26,004 --> 00:16:28,185
variability in the data that you have. Because

454
00:16:28,325 --> 00:16:30,644
as you can imagine, the semiconductor supply chain

455
00:16:30,644 --> 00:16:33,889
is so vast, and it's just so across

456
00:16:33,889 --> 00:16:36,289
so many countries, across so many grids, there's

457
00:16:36,289 --> 00:16:38,529
just so much happening in there that there's

458
00:16:38,529 --> 00:16:40,450
so much variability in it. And so I'm

459
00:16:40,450 --> 00:16:42,950
looking at optimization techniques that

460
00:16:43,554 --> 00:16:46,274
would still, despite all of this uncertainty, would

461
00:16:46,274 --> 00:16:48,514
still enable the designer to figure out what

462
00:16:48,514 --> 00:16:50,855
is the optimal design and what is, like,

463
00:16:50,915 --> 00:16:52,514
this is the design that you should be

464
00:16:52,514 --> 00:16:55,409
thinking about to have a more sustainable computing

465
00:16:55,409 --> 00:16:57,569
system, but also meet all of your other

466
00:16:57,569 --> 00:17:00,370
power performance area criteria that you need and

467
00:17:00,370 --> 00:17:01,809
what kind of trade offs that they can

468
00:17:01,809 --> 00:17:03,970
make. So that's that's the most exciting part

469
00:17:03,970 --> 00:17:05,829
for me right now in my project.

470
00:17:06,130 --> 00:17:07,650
Well, I look forward to hearing how that

471
00:17:07,650 --> 00:17:08,869
pans out in the future.

472
00:17:09,315 --> 00:17:11,315
Mariam Elgamel, thank you so much for appearing

473
00:17:11,315 --> 00:17:13,014
on the podcast. Thank you.

474
00:17:20,755 --> 00:17:23,075
Well, that was a really interesting discussion, Margaret.

475
00:17:23,075 --> 00:17:24,830
I mean, one thing that sort of struck

476
00:17:24,830 --> 00:17:25,970
me almost immediately

477
00:17:26,269 --> 00:17:26,769
is

478
00:17:27,309 --> 00:17:29,470
the, you know, the I I suppose the

479
00:17:29,470 --> 00:17:32,109
sort of pure economics of this. I mean,

480
00:17:32,109 --> 00:17:33,490
I can see if you

481
00:17:34,349 --> 00:17:35,950
you want to make a better chip, a

482
00:17:35,950 --> 00:17:36,930
chip that's more

483
00:17:37,294 --> 00:17:37,794
deficient,

484
00:17:38,095 --> 00:17:38,994
energy efficient,

485
00:17:39,454 --> 00:17:40,275
lower cost,

486
00:17:41,054 --> 00:17:42,575
you know, so I can see why you

487
00:17:42,654 --> 00:17:44,755
why a designer would want to do that.

488
00:17:44,894 --> 00:17:46,815
I can also understand why you'd want to

489
00:17:46,815 --> 00:17:47,315
eliminate,

490
00:17:48,654 --> 00:17:50,674
other sort of energy costs

491
00:17:51,109 --> 00:17:51,590
from,

492
00:17:51,910 --> 00:17:53,130
you know, from the process.

493
00:17:53,590 --> 00:17:54,090
But

494
00:17:54,549 --> 00:17:55,049
what,

495
00:17:56,070 --> 00:17:59,350
what are their economics that are driving things

496
00:17:59,350 --> 00:18:00,250
on the environmental

497
00:18:00,630 --> 00:18:02,809
side? I mean, particularly now,

498
00:18:03,350 --> 00:18:05,670
you know, when the the the tide almost

499
00:18:05,670 --> 00:18:08,384
seems to be going in a different direction,

500
00:18:08,605 --> 00:18:09,105
unfortunately,

501
00:18:09,884 --> 00:18:12,924
for making, well, just about anything greener. Did,

502
00:18:15,644 --> 00:18:17,484
was there any sort of talk about that

503
00:18:17,484 --> 00:18:19,825
at the conference, or did Maryam have any

504
00:18:20,045 --> 00:18:21,904
further insights into that?

505
00:18:22,330 --> 00:18:24,170
Yeah. Maryam, I didn't really get into that,

506
00:18:24,170 --> 00:18:25,529
but, you know, there was a little bit

507
00:18:25,529 --> 00:18:26,650
of a a talk or a lot of

508
00:18:26,650 --> 00:18:28,970
think people, sort of interested in in the

509
00:18:28,970 --> 00:18:31,850
environmental effects of AI, which you know are

510
00:18:31,850 --> 00:18:33,610
are quite severe in terms of how much

511
00:18:33,610 --> 00:18:36,430
energy is being used to create and run,

512
00:18:36,765 --> 00:18:39,484
the large language models that underlie things like

513
00:18:39,484 --> 00:18:42,464
chat GBT, Copilot, and various other AI systems.

514
00:18:43,404 --> 00:18:45,164
I think one of the the factors in

515
00:18:45,164 --> 00:18:48,305
terms of of computer chip design specifically and

516
00:18:48,649 --> 00:18:50,569
some of the chemicals that she talked about,

517
00:18:50,569 --> 00:18:51,230
you know,

518
00:18:51,609 --> 00:18:53,690
they are quite expensive and using less of

519
00:18:53,690 --> 00:18:55,849
them would be, you know, good, like, economically

520
00:18:55,849 --> 00:18:56,349
speaking

521
00:18:56,730 --> 00:18:58,730
as well as environmentally. And then I suppose

522
00:18:58,730 --> 00:19:00,349
if you want another economic perspective,

523
00:19:00,970 --> 00:19:03,904
if you end up dumping these chemicals into

524
00:19:03,904 --> 00:19:05,984
waterways, you can, in principle, get fined for

525
00:19:05,984 --> 00:19:07,424
that depending on where you are in the

526
00:19:07,424 --> 00:19:09,365
in in the world and how the local

527
00:19:09,904 --> 00:19:10,404
enforcement

528
00:19:10,865 --> 00:19:13,345
regime works. So I think that there are

529
00:19:13,345 --> 00:19:14,805
some some sort of cold

530
00:19:15,184 --> 00:19:16,085
cold hearted

531
00:19:16,464 --> 00:19:17,525
economic factors

532
00:19:18,200 --> 00:19:20,680
going towards making things greener, not just sort

533
00:19:20,680 --> 00:19:22,200
of, oh, well, it would be better for

534
00:19:22,200 --> 00:19:24,119
the environment. But, I mean, I'm not trying

535
00:19:24,119 --> 00:19:25,720
to sort of diminish the it would be

536
00:19:25,720 --> 00:19:28,119
better for the environment arguments because that's obviously

537
00:19:28,119 --> 00:19:29,640
important. We all live on this planet. We've

538
00:19:29,640 --> 00:19:31,000
all got to keep living on this planet.

539
00:19:31,000 --> 00:19:33,684
There's not really another option, is there? No.

540
00:19:33,684 --> 00:19:35,444
That that that's right. Yeah. And I suppose

541
00:19:35,444 --> 00:19:37,444
if, you know, if your factory is working

542
00:19:37,444 --> 00:19:38,424
to very high,

543
00:19:39,284 --> 00:19:42,085
environmental and health and safety standards, then if

544
00:19:42,085 --> 00:19:44,345
you can eliminate as many dangerous,

545
00:19:45,210 --> 00:19:48,509
nasty chemicals and materials from that process,

546
00:19:48,890 --> 00:19:50,650
you can you can save a lot of

547
00:19:50,650 --> 00:19:53,289
money, can't you? And, you know, sort of,

548
00:19:54,329 --> 00:19:56,569
make your employees happier because they don't have

549
00:19:56,569 --> 00:19:59,375
to work with these, you know, particularly nasty

550
00:19:59,375 --> 00:20:01,375
materials. So Yeah. Yeah. You you have health

551
00:20:01,375 --> 00:20:03,214
effects also for the people working with it.

552
00:20:03,214 --> 00:20:04,974
That's true. And I think it's a little

553
00:20:04,974 --> 00:20:06,595
bit like, other sort

554
00:20:07,134 --> 00:20:09,875
of economic, environmental configurations. I mean,

555
00:20:10,200 --> 00:20:11,500
not flying to conferences

556
00:20:12,119 --> 00:20:14,140
as much as we used to, perhaps,

557
00:20:14,679 --> 00:20:16,599
is good for the environment, but it's also

558
00:20:16,599 --> 00:20:17,659
good for organizations'

559
00:20:18,039 --> 00:20:20,359
bottom line. You know, not sending researchers around

560
00:20:20,359 --> 00:20:22,214
the world quite as much is, you know,

561
00:20:22,214 --> 00:20:24,214
beneficial for that, which I guess is is

562
00:20:24,214 --> 00:20:25,894
one problem with the things like the Heidelberg

563
00:20:25,894 --> 00:20:27,575
Laurier form, because people do come from all

564
00:20:27,575 --> 00:20:29,734
over the world. But I met some people

565
00:20:29,734 --> 00:20:31,414
there who said, you know, we the reason

566
00:20:31,414 --> 00:20:33,974
we're here, particularly some of the there's a

567
00:20:33,974 --> 00:20:35,815
group of alumni who can kind of come

568
00:20:35,815 --> 00:20:36,634
to the conference,

569
00:20:37,095 --> 00:20:37,595
occasionally.

570
00:20:38,289 --> 00:20:40,630
And they said, well, we have a collaboration

571
00:20:40,769 --> 00:20:42,369
going, and we're we're having it at the

572
00:20:42,369 --> 00:20:44,450
Heidelberg Laureate Forum because it saves us sort

573
00:20:44,450 --> 00:20:46,849
of travelling around various different places. So it

574
00:20:46,849 --> 00:20:48,609
saves us in that respect. So I think

575
00:20:48,609 --> 00:20:50,950
there are these sort of kind of parallel

576
00:20:51,009 --> 00:20:52,630
concerns that can feed into,

577
00:20:53,255 --> 00:20:55,835
things becoming greener in a slightly indirect way.

578
00:20:56,134 --> 00:20:57,914
Yeah. And I I suppose, you know,

579
00:20:58,295 --> 00:21:00,934
solar panels are a classic example of something

580
00:21:00,934 --> 00:21:01,434
where,

581
00:21:01,975 --> 00:21:03,894
yeah, I suppose they were brought in to

582
00:21:03,894 --> 00:21:06,875
reduce carbon emissions, but it turns out that,

583
00:21:07,420 --> 00:21:08,880
you know, you can produce electricity

584
00:21:09,259 --> 00:21:10,779
really cheaply with them. Yeah.

585
00:21:11,339 --> 00:21:13,599
You know, much cheaper than with fossil fuels.

586
00:21:14,059 --> 00:21:15,980
And so who's the, the second person that

587
00:21:15,980 --> 00:21:17,820
we're gonna hear from, Margaret? I have a

588
00:21:17,820 --> 00:21:20,460
funny feeling it's a fellow Canadian. We love

589
00:21:20,460 --> 00:21:22,934
to have Canadians on the podcast. Actually, Mariam's

590
00:21:22,934 --> 00:21:25,095
Canadian as well. Although although she's doing her

591
00:21:25,095 --> 00:21:27,575
PhD at Harvard, she is Canadian. So yeah.

592
00:21:27,575 --> 00:21:29,734
That's interesting because I was listening to her,

593
00:21:29,734 --> 00:21:30,554
and I thought

594
00:21:30,934 --> 00:21:33,174
she she sounds like she's from Montreal. Do

595
00:21:33,174 --> 00:21:34,990
you know if she's from Montreal? I don't

596
00:21:34,990 --> 00:21:36,509
think so. No. I think I think she's

597
00:21:36,509 --> 00:21:38,829
more more more Toronto, probably from near your

598
00:21:38,829 --> 00:21:41,230
area. But, she's she's sort of lived between

599
00:21:41,230 --> 00:21:43,069
Canada and Egypt and The US, so it's,

600
00:21:43,069 --> 00:21:44,289
you know Wow.

601
00:21:44,589 --> 00:21:47,309
So both Canadians. Both Canadians. Yeah. Yeah. Well,

602
00:21:47,309 --> 00:21:48,990
well done, Margaret. I should say that I

603
00:21:48,990 --> 00:21:50,204
had nothing to do with,

604
00:21:50,765 --> 00:21:53,085
with this. That's great. Well, let let let's

605
00:21:53,085 --> 00:21:55,024
move on to, to the next Canadian.

606
00:21:55,325 --> 00:21:57,484
What what's his name? So he's Andrew Gunter.

607
00:21:57,484 --> 00:21:59,644
He's finishing his PhD at the University of

608
00:21:59,644 --> 00:22:02,125
British Columbia in Canada, and he works in

609
00:22:02,125 --> 00:22:04,065
designing circuits for computer chips.

610
00:22:04,390 --> 00:22:06,230
And I asked him to start by explaining

611
00:22:06,230 --> 00:22:09,130
what that design process usually looks like.

612
00:22:18,549 --> 00:22:19,609
Designing chips

613
00:22:20,035 --> 00:22:22,855
and creating chips, let's say, has three aspects.

614
00:22:23,154 --> 00:22:25,714
First, there's the design entry, where you determine

615
00:22:25,714 --> 00:22:26,454
the functionality

616
00:22:26,994 --> 00:22:28,914
of the chip that you want. But you

617
00:22:28,914 --> 00:22:30,434
don't know anything about how it's going to

618
00:22:30,434 --> 00:22:31,095
be implemented.

619
00:22:31,809 --> 00:22:33,170
You don't know what it's going to look

620
00:22:33,170 --> 00:22:35,330
like, how well it's going to perform. You

621
00:22:35,330 --> 00:22:37,269
just know the functions. That's step one.

622
00:22:38,130 --> 00:22:39,109
Step two

623
00:22:39,410 --> 00:22:42,609
is converting that functional description of the chip

624
00:22:42,609 --> 00:22:43,990
to a physical description.

625
00:22:44,664 --> 00:22:47,464
That physical description now, it tells you where

626
00:22:47,464 --> 00:22:50,204
the components go, where the wires are laid,

627
00:22:50,265 --> 00:22:51,644
how it's going to be manufactured.

628
00:22:52,505 --> 00:22:53,944
Once you have that, then you go to

629
00:22:53,944 --> 00:22:54,444
fabrication.

630
00:22:55,545 --> 00:22:56,045
You

631
00:22:56,505 --> 00:22:59,325
ship your design off to TSMC in Taiwan.

632
00:22:59,710 --> 00:23:01,549
They do a whole bunch of very complicated

633
00:23:01,549 --> 00:23:04,430
things, and the output is a physical chip.

634
00:23:04,430 --> 00:23:06,509
And you would have this regardless of whether

635
00:23:06,509 --> 00:23:08,269
you're ordering a couple of chips for a

636
00:23:08,269 --> 00:23:09,650
very specialist application

637
00:23:10,430 --> 00:23:12,750
all the way through mass manufacturing. Is that

638
00:23:12,750 --> 00:23:13,970
right? Yeah. So

639
00:23:14,315 --> 00:23:15,835
for what we'd say the high end tech

640
00:23:15,835 --> 00:23:17,855
nodes are, those are very expensive.

641
00:23:18,795 --> 00:23:21,035
You really only want to use high end

642
00:23:21,035 --> 00:23:22,815
technology if it's mass market

643
00:23:23,595 --> 00:23:25,694
production. If there's a lower volume,

644
00:23:26,154 --> 00:23:29,295
you might use an older technology, cheaper and

645
00:23:29,619 --> 00:23:30,119
more

646
00:23:30,579 --> 00:23:32,519
reliable. A big issue with chip fabrication

647
00:23:32,899 --> 00:23:35,480
is errors in the fabrication process.

648
00:23:35,940 --> 00:23:37,880
The newer and more complex the processes,

649
00:23:38,259 --> 00:23:40,179
the more errors you have, the lower your

650
00:23:40,179 --> 00:23:41,799
yield rate is. So maybe

651
00:23:42,419 --> 00:23:44,339
you try to manufacture, let's just say, a

652
00:23:44,339 --> 00:23:47,184
100 chips on a new technology node, maybe

653
00:23:47,184 --> 00:23:49,365
you only get 80 out of the 100.

654
00:23:49,664 --> 00:23:50,804
That drives up cost.

655
00:23:51,585 --> 00:23:53,525
If you have really low,

656
00:23:53,904 --> 00:23:55,904
let's say, volume requirements for how many chips

657
00:23:55,904 --> 00:23:58,085
you need, you might look at field programmable

658
00:23:58,224 --> 00:24:00,630
gateways, a form of reconfigurable chip, which is

659
00:24:00,630 --> 00:24:01,769
actually my expertise.

660
00:24:02,630 --> 00:24:04,069
And we can talk about that more if

661
00:24:04,069 --> 00:24:05,609
you want, but that's the spectrum.

662
00:24:05,990 --> 00:24:08,089
Okay. So tell me what a field programmable

663
00:24:08,230 --> 00:24:09,289
gate array is.

664
00:24:09,669 --> 00:24:10,169
Yes.

665
00:24:10,470 --> 00:24:10,970
So

666
00:24:11,515 --> 00:24:13,595
as opposed to what we call an ASIC,

667
00:24:13,595 --> 00:24:16,015
an application specific integrated circuit,

668
00:24:16,954 --> 00:24:19,054
where so for the ASIC, you

669
00:24:19,515 --> 00:24:21,035
determine the design that you want and you

670
00:24:21,035 --> 00:24:22,575
go off and you manufacture it.

671
00:24:23,434 --> 00:24:26,335
An FPGA, a Field Programmable Gate Array,

672
00:24:26,700 --> 00:24:28,799
is a chip with fixed resources

673
00:24:29,500 --> 00:24:31,279
which has already been manufactured,

674
00:24:31,819 --> 00:24:33,440
and after the manufacturing

675
00:24:33,740 --> 00:24:35,679
process, you then determine

676
00:24:36,539 --> 00:24:38,859
the design that you want to implement in

677
00:24:38,859 --> 00:24:39,599
the FPGA.

678
00:24:40,174 --> 00:24:42,654
So let me clarify what that means. Let's

679
00:24:42,654 --> 00:24:44,034
say you're creating a CPU.

680
00:24:44,734 --> 00:24:47,454
Right? You could design the CPU and manufacture

681
00:24:47,454 --> 00:24:48,815
it, and that will cost a bunch of

682
00:24:48,815 --> 00:24:49,315
money.

683
00:24:49,934 --> 00:24:52,755
Or you design the CPU and you implement

684
00:24:52,815 --> 00:24:54,595
the design on an FPGA.

685
00:24:55,099 --> 00:24:56,559
And now this FPGA

686
00:24:57,580 --> 00:24:59,119
implements the functionality

687
00:24:59,580 --> 00:25:01,440
of the CPU that you specified.

688
00:25:02,140 --> 00:25:04,299
So the FPGA has a fixed set of

689
00:25:04,299 --> 00:25:04,799
resources.

690
00:25:05,500 --> 00:25:07,359
If the CPU you've designed

691
00:25:07,785 --> 00:25:08,525
only requires

692
00:25:09,144 --> 00:25:11,884
less than those resources, you can't implement it.

693
00:25:11,945 --> 00:25:13,785
If it requires more, either you get a

694
00:25:13,785 --> 00:25:15,884
bigger FPGA or you go to an ASIC.

695
00:25:16,585 --> 00:25:18,105
That sounds a little bit like it's a

696
00:25:18,105 --> 00:25:20,025
breadboard type of stage. If you're thinking about

697
00:25:20,025 --> 00:25:22,125
traditional physically wiring circuits,

698
00:25:22,549 --> 00:25:24,150
you put something on the breadboard, then you

699
00:25:24,150 --> 00:25:25,669
see if it works, and then only then

700
00:25:25,669 --> 00:25:27,269
do you go off and solder the real

701
00:25:27,269 --> 00:25:27,769
thing.

702
00:25:28,470 --> 00:25:30,950
Exactly. So a breadboard would be useful for

703
00:25:30,950 --> 00:25:33,190
a lot of analog circuitry or very simple

704
00:25:33,190 --> 00:25:33,690
circuits.

705
00:25:34,150 --> 00:25:36,950
Once you get to complicated digital circuitry and

706
00:25:36,950 --> 00:25:38,169
you want to do prototyping,

707
00:25:38,524 --> 00:25:39,825
then you go to an FPGA.

708
00:25:40,204 --> 00:25:42,605
In fact, and I'm not an industry person,

709
00:25:42,605 --> 00:25:44,065
but from what I've been told,

710
00:25:44,605 --> 00:25:46,684
one of the big revenue drivers for the

711
00:25:46,684 --> 00:25:49,184
FPGA industry is prototyping

712
00:25:49,484 --> 00:25:49,984
ASICs.

713
00:25:50,845 --> 00:25:52,944
So they do get used in that production

714
00:25:53,005 --> 00:25:55,730
loop as well. What specific problem with this

715
00:25:55,730 --> 00:25:57,409
process have you been trying to solve in

716
00:25:57,409 --> 00:25:58,150
your PhD?

717
00:25:58,529 --> 00:26:01,569
Yeah. So maybe multiple problems, but I will

718
00:26:01,569 --> 00:26:02,069
give

719
00:26:02,529 --> 00:26:04,710
the let's say, kind of the big picture

720
00:26:04,769 --> 00:26:05,269
problem.

721
00:26:05,809 --> 00:26:07,569
So I told you step one, step two,

722
00:26:07,569 --> 00:26:08,710
and step three previously.

723
00:26:09,184 --> 00:26:10,644
Step one is design entry.

724
00:26:11,105 --> 00:26:14,085
Step two is the conversion from a functional

725
00:26:14,224 --> 00:26:16,384
to a physical design, and we refer to

726
00:26:16,384 --> 00:26:19,424
that as electronic design automation, and step three

727
00:26:19,424 --> 00:26:20,325
is the fabrication.

728
00:26:21,105 --> 00:26:23,444
So my research deals with step two primarily,

729
00:26:24,000 --> 00:26:27,039
electronic design automation. We have the functional design.

730
00:26:27,039 --> 00:26:28,980
We want to produce a physical design.

731
00:26:29,680 --> 00:26:31,519
Now the way that we do this is

732
00:26:31,519 --> 00:26:33,839
in several stages and we'll say that at

733
00:26:33,839 --> 00:26:36,400
each stage, we have usually an NP hard

734
00:26:36,400 --> 00:26:38,480
problem to solve and we use a heuristic

735
00:26:38,480 --> 00:26:39,734
algorithm to solve it.

736
00:26:40,214 --> 00:26:41,194
So stage

737
00:26:41,654 --> 00:26:43,974
one within this, then stage two, stage three,

738
00:26:43,974 --> 00:26:45,595
it's a sequential process.

739
00:26:46,454 --> 00:26:47,755
At each stage,

740
00:26:48,454 --> 00:26:49,595
things can go wrong.

741
00:26:49,894 --> 00:26:51,434
We're using heuristic algorithms.

742
00:26:52,069 --> 00:26:52,569
So

743
00:26:52,950 --> 00:26:55,529
they can fail to produce a viable solution.

744
00:26:56,390 --> 00:26:57,529
If they succeed,

745
00:26:58,069 --> 00:26:59,369
they can take an unexpectedly

746
00:26:59,750 --> 00:27:00,490
long time

747
00:27:00,789 --> 00:27:02,410
to succeed and produce a solution,

748
00:27:02,869 --> 00:27:04,630
and the solution may be of a low

749
00:27:04,630 --> 00:27:05,130
quality.

750
00:27:05,775 --> 00:27:08,595
So my research deals with predicting in advance

751
00:27:09,055 --> 00:27:10,434
before you run the algorithms

752
00:27:11,134 --> 00:27:12,894
if any of those three things will become

753
00:27:12,894 --> 00:27:13,555
an issue.

754
00:27:14,095 --> 00:27:14,595
So

755
00:27:15,535 --> 00:27:16,755
prior to this research,

756
00:27:17,295 --> 00:27:20,355
that electronic design automation process, it's very uncertain,

757
00:27:20,849 --> 00:27:22,289
and I like to say that my research

758
00:27:22,289 --> 00:27:25,029
is adding certainty to electronic design automation.

759
00:27:25,650 --> 00:27:27,730
That sounds like the equivalent of if you're

760
00:27:27,730 --> 00:27:29,490
having a software update, you might want to

761
00:27:29,490 --> 00:27:31,589
know that the software update is gonna take,

762
00:27:31,650 --> 00:27:33,410
I don't know, three hours and forty five

763
00:27:33,410 --> 00:27:35,554
minutes to run that thing over time as

764
00:27:35,554 --> 00:27:37,654
opposed to taking, you know, two minutes.

765
00:27:38,035 --> 00:27:39,954
Yeah. So I normally don't market this way,

766
00:27:39,954 --> 00:27:41,714
but when I talk to others in the

767
00:27:41,714 --> 00:27:44,914
domain, they say, oh, you're creating a progress

768
00:27:44,914 --> 00:27:47,315
bar for electronic design automation. And I go,

769
00:27:47,315 --> 00:27:48,595
yeah. You know what? You can think of

770
00:27:48,595 --> 00:27:50,275
it that way. There's a little bit more,

771
00:27:50,275 --> 00:27:52,369
but you could create a progress bar with

772
00:27:52,369 --> 00:27:55,089
this. Whereas today, there's nothing. You just kind

773
00:27:55,089 --> 00:27:56,369
of sit there and you wait and you

774
00:27:56,369 --> 00:27:58,369
see what happens. I want to get a

775
00:27:58,369 --> 00:28:00,210
little bit into the mathematics of this. You

776
00:28:00,210 --> 00:28:03,029
mentioned something as being an NP hard process.

777
00:28:03,650 --> 00:28:04,549
Yeah. So

778
00:28:05,484 --> 00:28:07,804
without going into the technical details of what

779
00:28:07,804 --> 00:28:10,785
NP hardness is, I think the important intuition

780
00:28:11,565 --> 00:28:14,065
is that we're looking at optimization problems,

781
00:28:14,845 --> 00:28:15,345
and

782
00:28:16,044 --> 00:28:19,565
there are many, many different possible solutions to

783
00:28:19,565 --> 00:28:21,500
this. So we would typically speak to the

784
00:28:21,500 --> 00:28:22,799
complexity of the problem,

785
00:28:23,339 --> 00:28:23,839
and

786
00:28:24,220 --> 00:28:26,400
the problems that we're dealing with have combinatorial

787
00:28:26,859 --> 00:28:27,359
complexity,

788
00:28:28,059 --> 00:28:28,559
where

789
00:28:28,940 --> 00:28:29,839
for every

790
00:28:30,380 --> 00:28:32,480
element of the output solution,

791
00:28:33,045 --> 00:28:35,365
you can choose from, let's say, a set

792
00:28:35,365 --> 00:28:36,025
of possibilities.

793
00:28:37,125 --> 00:28:38,184
And for each

794
00:28:38,725 --> 00:28:39,225
possibility

795
00:28:39,684 --> 00:28:41,465
for a given element of the solution,

796
00:28:42,325 --> 00:28:44,664
any combination of those is

797
00:28:45,380 --> 00:28:47,380
viable in the output space. So we say

798
00:28:47,380 --> 00:28:48,120
it's combinatorially

799
00:28:48,500 --> 00:28:51,080
hard, and those combinatorially hard problems

800
00:28:51,620 --> 00:28:54,019
are they fall into the complexity class of

801
00:28:54,019 --> 00:28:55,799
NP hardness in computer science.

802
00:28:56,580 --> 00:28:58,279
So very, very hard stuff.

803
00:28:59,275 --> 00:29:01,115
And what was the outcome of your process?

804
00:29:01,115 --> 00:29:03,615
What stage is this business project at currently?

805
00:29:03,994 --> 00:29:06,075
Yeah. So I'm toward the end of my

806
00:29:06,075 --> 00:29:06,575
PhD.

807
00:29:06,875 --> 00:29:09,355
Many years have gone into it, and I've

808
00:29:09,355 --> 00:29:10,954
completed a lot, but there's a lot more

809
00:29:10,954 --> 00:29:13,259
to go. So what I've done so far

810
00:29:13,400 --> 00:29:15,019
is I've looked at the

811
00:29:15,320 --> 00:29:15,820
hardest

812
00:29:16,279 --> 00:29:16,779
problem

813
00:29:17,160 --> 00:29:19,799
within electronic design automation, which is a routing

814
00:29:19,799 --> 00:29:20,299
problem.

815
00:29:21,000 --> 00:29:21,500
Now

816
00:29:22,200 --> 00:29:25,080
routing is hard, not just because of the

817
00:29:25,080 --> 00:29:27,194
complexity of the task, but the scale at

818
00:29:27,194 --> 00:29:28,255
which we do it.

819
00:29:28,875 --> 00:29:29,375
So

820
00:29:30,154 --> 00:29:33,275
for routing, we typically think of the chip

821
00:29:33,275 --> 00:29:35,115
as a graph. And when I say graph,

822
00:29:35,115 --> 00:29:37,755
I mean in the computer science domain where

823
00:29:37,755 --> 00:29:39,054
the graph has nodes

824
00:29:39,434 --> 00:29:40,734
and nodes have edges.

825
00:29:41,359 --> 00:29:42,580
The nodes represent

826
00:29:43,119 --> 00:29:45,840
wiring resources in the chip and the edges

827
00:29:45,840 --> 00:29:46,900
represent the connectivity

828
00:29:47,440 --> 00:29:49,539
between the wires in the chip.

829
00:29:50,240 --> 00:29:51,539
Now, the routing problem

830
00:29:52,240 --> 00:29:53,460
involves connecting,

831
00:29:54,825 --> 00:29:55,325
communicating

832
00:29:55,785 --> 00:29:58,105
components in the design. So if I have

833
00:29:58,105 --> 00:30:00,585
component a in the chip and component b

834
00:30:00,585 --> 00:30:02,904
and my functional design says these two things

835
00:30:02,904 --> 00:30:04,585
need to talk to each other, we need

836
00:30:04,585 --> 00:30:06,904
to lay a wire that communicate that connects

837
00:30:06,904 --> 00:30:07,884
the two of them.

838
00:30:08,265 --> 00:30:09,980
Now, that's really simple.

839
00:30:10,519 --> 00:30:12,839
However, it's not usually just component a and

840
00:30:12,839 --> 00:30:14,460
b. There's not normally two.

841
00:30:14,839 --> 00:30:17,500
The typical order of magnitude in my work

842
00:30:17,640 --> 00:30:18,460
is about

843
00:30:19,799 --> 00:30:21,079
10,000,000

844
00:30:21,079 --> 00:30:21,579
components,

845
00:30:23,054 --> 00:30:25,315
and the typical number of wires

846
00:30:25,934 --> 00:30:27,154
is about a billion.

847
00:30:28,014 --> 00:30:31,794
And so this graph then has about 10,000,000

848
00:30:31,855 --> 00:30:33,875
nodes and a billion edges.

849
00:30:34,734 --> 00:30:37,054
And so we have to use algorithms to

850
00:30:37,054 --> 00:30:37,554
solve

851
00:30:38,089 --> 00:30:39,549
this NP hard problem

852
00:30:40,410 --> 00:30:42,830
on a billion scale chip.

853
00:30:43,930 --> 00:30:46,490
And so my research is predicting whether the

854
00:30:46,490 --> 00:30:49,630
algorithms can do that for a specific problem,

855
00:30:49,930 --> 00:30:51,914
predict how long it's going to take, and

856
00:30:51,914 --> 00:30:53,695
predict the quality of the final solution.

857
00:30:54,315 --> 00:30:56,315
But if your process said, actually, this is

858
00:30:56,315 --> 00:30:58,075
going to either take a very long time

859
00:30:58,075 --> 00:30:59,515
or it might take a long time, might

860
00:30:59,515 --> 00:31:01,515
take a short time, but the solution it's

861
00:31:01,515 --> 00:31:02,795
going to come up with is not gonna

862
00:31:02,795 --> 00:31:03,695
be very good,

863
00:31:04,315 --> 00:31:05,375
what happens then?

864
00:31:06,075 --> 00:31:08,730
Yep. So a few different things can happen.

865
00:31:09,349 --> 00:31:11,670
It depends on exactly how your design flow

866
00:31:11,670 --> 00:31:14,089
looks, but I'll give you two examples.

867
00:31:15,029 --> 00:31:15,849
One example

868
00:31:16,630 --> 00:31:18,309
that is a state of the art for

869
00:31:18,309 --> 00:31:19,625
many people is

870
00:31:20,805 --> 00:31:21,305
they

871
00:31:22,325 --> 00:31:22,825
perform

872
00:31:23,125 --> 00:31:25,605
what we call the placement task, which happens

873
00:31:25,605 --> 00:31:26,505
before routing,

874
00:31:27,285 --> 00:31:28,825
with multiple different instances.

875
00:31:29,125 --> 00:31:31,125
So what this means is that in placement,

876
00:31:31,125 --> 00:31:32,565
we have components of the chip, and we

877
00:31:32,565 --> 00:31:34,259
need to figure out where they go. We

878
00:31:34,259 --> 00:31:36,119
have to place them at fixed locations.

879
00:31:36,980 --> 00:31:40,200
We use pseudo random algorithms for this, which

880
00:31:40,580 --> 00:31:43,320
give different outputs based on a different random

881
00:31:43,700 --> 00:31:46,065
number generator seed. So So let's say we'll

882
00:31:46,065 --> 00:31:48,304
do 10 different placements with 10 different random

883
00:31:48,304 --> 00:31:48,804
seeds.

884
00:31:49,184 --> 00:31:51,125
And then now we have 10 different inputs

885
00:31:51,184 --> 00:31:52,325
to the routing problem.

886
00:31:53,025 --> 00:31:55,825
We will then run 10 different instances of

887
00:31:55,825 --> 00:31:56,804
the routing problem,

888
00:31:57,345 --> 00:32:00,325
and we can use my research to predict

889
00:32:00,470 --> 00:32:03,450
which of those is most likely to succeed,

890
00:32:03,910 --> 00:32:05,910
take a short amount of time, and then

891
00:32:05,910 --> 00:32:07,529
give you a good result at the end.

892
00:32:07,670 --> 00:32:09,349
And so you would kill nine of the

893
00:32:09,349 --> 00:32:11,910
instances and keep just one. So that's one

894
00:32:11,910 --> 00:32:14,045
way it could be used. Another way that

895
00:32:14,045 --> 00:32:14,944
it could be used

896
00:32:15,644 --> 00:32:17,184
in tandem with that actually

897
00:32:17,644 --> 00:32:19,265
is you have a routing problem

898
00:32:19,805 --> 00:32:20,704
and you predict

899
00:32:21,085 --> 00:32:22,684
it's not going to work out or it's

900
00:32:22,765 --> 00:32:24,765
there's a low probability that is going to

901
00:32:24,765 --> 00:32:26,625
work out and give a successful solution.

902
00:32:27,130 --> 00:32:27,630
You

903
00:32:28,169 --> 00:32:30,589
can kill the routing run early,

904
00:32:31,289 --> 00:32:33,289
go to an earlier stage of the design

905
00:32:33,289 --> 00:32:33,789
process,

906
00:32:34,250 --> 00:32:36,829
make some tweaks based on your engineering expertise

907
00:32:37,289 --> 00:32:39,950
that you believe are going to improve the

908
00:32:40,434 --> 00:32:43,555
likelihood that the chip will be routable, as

909
00:32:43,555 --> 00:32:44,695
we like to say it.

910
00:32:45,075 --> 00:32:46,674
You've worked out that you can't get from

911
00:32:46,674 --> 00:32:48,434
point a to point b by subway, and

912
00:32:48,434 --> 00:32:49,715
you've got to work out how to do

913
00:32:49,715 --> 00:32:50,775
it on foot instead.

914
00:32:51,154 --> 00:32:53,970
Exactly. Right? During this project, was there anything

915
00:32:53,970 --> 00:32:55,570
that you thought would be hard that turned

916
00:32:55,570 --> 00:32:56,549
out to be easy

917
00:32:56,930 --> 00:32:59,490
or vice versa? Anything you thought that would

918
00:32:59,490 --> 00:33:01,330
be easy turned out to be really, really

919
00:33:01,330 --> 00:33:01,830
difficult?

920
00:33:02,369 --> 00:33:04,210
I can't say anything that I thought was

921
00:33:04,210 --> 00:33:05,650
going to be hard turned out to be

922
00:33:05,650 --> 00:33:07,045
easy, but but I can say that there

923
00:33:07,045 --> 00:33:08,244
are things which I thought was going to

924
00:33:08,244 --> 00:33:09,945
be easy that turned out to be hard.

925
00:33:10,484 --> 00:33:13,144
I'll give you one example of that. So

926
00:33:14,164 --> 00:33:16,404
something which I've personally tried to look at,

927
00:33:16,404 --> 00:33:18,744
which prior research hasn't really, is

928
00:33:19,125 --> 00:33:20,184
the the probabilities

929
00:33:20,644 --> 00:33:21,945
of what's going to happen.

930
00:33:22,289 --> 00:33:22,789
So

931
00:33:23,090 --> 00:33:25,190
I apply machine learning to do these predictions.

932
00:33:26,130 --> 00:33:28,049
There's been prior work on using machine learning

933
00:33:28,049 --> 00:33:30,710
for similar aspects of electronic design automation

934
00:33:31,330 --> 00:33:34,789
and they typically predict the expected outcome.

935
00:33:35,250 --> 00:33:36,230
So for example,

936
00:33:37,414 --> 00:33:38,795
we could say that

937
00:33:39,255 --> 00:33:41,174
we expect that the chip is going to

938
00:33:41,174 --> 00:33:44,375
come out with an operating frequency of 500

939
00:33:44,375 --> 00:33:44,875
megahertz.

940
00:33:45,734 --> 00:33:46,234
Sure.

941
00:33:47,255 --> 00:33:49,494
You can take a more nuanced view to

942
00:33:49,494 --> 00:33:50,795
this. You could say

943
00:33:51,734 --> 00:33:55,170
our average expected result is 500 megahertz,

944
00:33:55,549 --> 00:33:57,230
but within a 5%

945
00:33:57,230 --> 00:33:58,509
to 95%

946
00:33:58,509 --> 00:33:59,650
confidence interval,

947
00:34:00,029 --> 00:34:02,049
we expect it to be be between

948
00:34:02,750 --> 00:34:05,089
400 megahertz and 600 megahertz.

949
00:34:06,109 --> 00:34:08,655
As it turns out, this is very useful

950
00:34:08,655 --> 00:34:09,954
to do, but it's not

951
00:34:10,335 --> 00:34:10,835
straightforward

952
00:34:11,295 --> 00:34:13,795
to do that. Why is that so useful?

953
00:34:14,255 --> 00:34:14,755
So

954
00:34:15,375 --> 00:34:17,474
it ends up being useful because

955
00:34:17,855 --> 00:34:18,994
we often have,

956
00:34:19,860 --> 00:34:20,360
let's

957
00:34:20,820 --> 00:34:21,320
say,

958
00:34:22,980 --> 00:34:24,680
some room for tolerance

959
00:34:25,140 --> 00:34:26,200
in these things.

960
00:34:26,740 --> 00:34:28,820
Right? So it it's nice to know that

961
00:34:28,820 --> 00:34:30,920
we expect a 500 megahertz chip,

962
00:34:31,220 --> 00:34:34,755
but maybe we don't actually need that. Maybe

963
00:34:34,755 --> 00:34:35,635
450

964
00:34:35,635 --> 00:34:36,934
megahertz would be enough.

965
00:34:38,034 --> 00:34:38,534
So

966
00:34:39,074 --> 00:34:41,155
only knowing that we expect 500 is nice.

967
00:34:41,155 --> 00:34:42,755
That kind of tells us most likely we'll

968
00:34:42,755 --> 00:34:44,934
get what we want. But what if

969
00:34:45,315 --> 00:34:47,355
there's a a significant probability of it going

970
00:34:47,355 --> 00:34:48,835
to be below 450

971
00:34:48,835 --> 00:34:49,335
megahertz?

972
00:34:49,719 --> 00:34:52,760
Then that single prediction of just 500 doesn't

973
00:34:52,760 --> 00:34:54,539
give you any information about that.

974
00:34:55,160 --> 00:34:55,660
Now

975
00:34:56,199 --> 00:34:57,660
if you don't have that information,

976
00:34:58,599 --> 00:34:59,820
how do you make a decision?

977
00:35:00,119 --> 00:35:01,559
How do you know whether this is going

978
00:35:01,559 --> 00:35:02,460
to be good enough?

979
00:35:02,835 --> 00:35:04,275
You don't, and so you have to guess,

980
00:35:04,275 --> 00:35:06,675
and that's what people have done. People have

981
00:35:06,675 --> 00:35:08,375
guessed in the face of uncertainty.

982
00:35:08,914 --> 00:35:11,394
But now I'm giving you a probabilistic view

983
00:35:11,394 --> 00:35:12,855
of this. I can tell you

984
00:35:13,234 --> 00:35:14,054
the probability

985
00:35:14,675 --> 00:35:16,934
based on prior data that we've collected

986
00:35:18,219 --> 00:35:20,780
indicates that there's an x percent chance that

987
00:35:20,780 --> 00:35:22,940
the frequency of the chip will be greater

988
00:35:22,940 --> 00:35:25,739
than your minimum threshold of 450

989
00:35:25,739 --> 00:35:26,239
megahertz.

990
00:35:26,860 --> 00:35:29,260
Are there any implications of these results beyond

991
00:35:29,260 --> 00:35:30,940
chip design? I mean, I can imagine that

992
00:35:30,940 --> 00:35:32,719
other industries will also have routing

993
00:35:33,974 --> 00:35:35,974
problems and will also be using the algorithms

994
00:35:35,974 --> 00:35:37,894
to try to solve them. Yes. So I

995
00:35:37,894 --> 00:35:38,554
think this

996
00:35:39,094 --> 00:35:39,674
is an

997
00:35:40,135 --> 00:35:40,635
early

998
00:35:41,414 --> 00:35:45,599
but fertile area for research. So thinking thinking

999
00:35:45,599 --> 00:35:47,920
more fundamentally about what my research is doing,

1000
00:35:47,920 --> 00:35:48,900
yes, we're predicting

1001
00:35:49,440 --> 00:35:50,820
outcomes in chip design,

1002
00:35:51,360 --> 00:35:53,780
but we're really predicting the outcomes of algorithms.

1003
00:35:54,320 --> 00:35:57,135
Algorithms get used beyond the chip design world.

1004
00:35:57,375 --> 00:35:58,815
They get used in more places than I

1005
00:35:58,815 --> 00:36:00,114
could ever imagine counting.

1006
00:36:00,655 --> 00:36:01,394
And so

1007
00:36:02,175 --> 00:36:04,815
researching the outcomes of algorithms has a more

1008
00:36:04,815 --> 00:36:05,875
global implication,

1009
00:36:06,255 --> 00:36:07,775
an implication that I think a lot of

1010
00:36:07,775 --> 00:36:09,235
other researchers and practitioners

1011
00:36:09,775 --> 00:36:11,074
should pay attention to

1012
00:36:11,460 --> 00:36:11,960
because

1013
00:36:12,340 --> 00:36:14,199
it it gives us new applications

1014
00:36:14,500 --> 00:36:16,500
for this research, but I think it also

1015
00:36:16,500 --> 00:36:18,519
drives new research areas. So

1016
00:36:19,219 --> 00:36:21,059
one of the hot topics in machine learning

1017
00:36:21,059 --> 00:36:22,840
these days is foundation models.

1018
00:36:23,775 --> 00:36:24,275
Now

1019
00:36:25,534 --> 00:36:26,914
I'm telling you that

1020
00:36:27,295 --> 00:36:29,534
there are many different algorithms that we could

1021
00:36:29,534 --> 00:36:30,514
apply this to.

1022
00:36:30,974 --> 00:36:31,954
Here's an idea.

1023
00:36:32,414 --> 00:36:34,275
What if we collected data

1024
00:36:34,815 --> 00:36:35,315
from

1025
00:36:35,614 --> 00:36:36,835
several different algorithms

1026
00:36:37,329 --> 00:36:40,449
to train one foundation model, which is used

1027
00:36:40,449 --> 00:36:43,670
to predict the outcomes of many different algorithms,

1028
00:36:44,210 --> 00:36:45,730
and then we fine tune this as we

1029
00:36:45,730 --> 00:36:46,230
want.

1030
00:36:46,530 --> 00:36:48,369
This is a paradigm that we've seen successful

1031
00:36:48,369 --> 00:36:50,769
in other areas of machine learning, but not

1032
00:36:50,769 --> 00:36:52,389
really in the algorithms world.

1033
00:36:52,704 --> 00:36:55,025
And something that's really key to note here

1034
00:36:55,025 --> 00:36:57,424
is that we can collect high quality data

1035
00:36:57,424 --> 00:37:00,644
from algorithms. It's really easy to extract information

1036
00:37:01,184 --> 00:37:04,085
because we're not dealing with physical processes with

1037
00:37:04,385 --> 00:37:06,704
noisy sensor data that you might have in,

1038
00:37:06,704 --> 00:37:08,579
say, the physics world. We're in the world

1039
00:37:08,579 --> 00:37:10,760
of software. And in the world of software,

1040
00:37:10,900 --> 00:37:13,220
we can collect near perfect information to train

1041
00:37:13,220 --> 00:37:15,480
our models. That does sound exciting.

1042
00:37:16,420 --> 00:37:18,739
Okay. Final question. What are you planning to

1043
00:37:18,739 --> 00:37:19,400
do next?

1044
00:37:19,700 --> 00:37:21,160
That's a good question. So

1045
00:37:22,464 --> 00:37:24,304
I do everything that I said in my

1046
00:37:24,304 --> 00:37:25,925
academic life, in my PhD.

1047
00:37:26,385 --> 00:37:28,224
I also have an industry life on the

1048
00:37:28,224 --> 00:37:30,385
side. So I work at part time at

1049
00:37:30,385 --> 00:37:33,284
a startup called Singulos Research in Vancouver, Canada,

1050
00:37:33,905 --> 00:37:36,065
where we do nothing related to anything I

1051
00:37:36,065 --> 00:37:38,199
just said. We are looking at

1052
00:37:38,819 --> 00:37:39,719
mixed reality

1053
00:37:40,019 --> 00:37:43,639
experiences using computer vision for real time applications

1054
00:37:43,859 --> 00:37:44,920
on mobile devices.

1055
00:37:45,859 --> 00:37:46,359
And

1056
00:37:46,659 --> 00:37:47,159
I'm

1057
00:37:47,539 --> 00:37:49,400
hoping that things go well with the start

1058
00:37:49,460 --> 00:37:51,139
up, and if they do, I'm planning to

1059
00:37:51,139 --> 00:37:53,984
continue research there. Maybe I'll go between that

1060
00:37:53,984 --> 00:37:56,224
and the chip design world. I have several

1061
00:37:56,224 --> 00:37:56,724
interests.

1062
00:37:57,105 --> 00:37:58,724
So it's up in there at the moment.

1063
00:37:59,105 --> 00:38:00,085
Well, good luck.

1064
00:38:00,385 --> 00:38:01,664
Thank you very much for appearing on the

1065
00:38:01,664 --> 00:38:03,184
podcast. It's been good to talk to you.

1066
00:38:03,184 --> 00:38:04,724
Thank you. It's been lovely, Margaret.

1067
00:38:12,800 --> 00:38:15,860
That was Andrew Gunther speaking to Margaret Harris

1068
00:38:16,000 --> 00:38:20,099
at the Heidelberg Laureate Forum in Heidelberg, Germany.

1069
00:38:20,894 --> 00:38:23,954
Thanks to Andrew and our other guest, Mariam

1070
00:38:24,094 --> 00:38:24,594
Elgemel,

1071
00:38:25,135 --> 00:38:28,195
for enlightening us about what goes into designing

1072
00:38:28,335 --> 00:38:29,315
computer chips,

1073
00:38:29,695 --> 00:38:32,255
the chips that we all use just about

1074
00:38:32,255 --> 00:38:34,355
every hour of every day.

1075
00:38:35,070 --> 00:38:37,329
And, Margaret, thanks for coming on the podcast.

1076
00:38:37,469 --> 00:38:39,230
And, can can you give us a little

1077
00:38:39,230 --> 00:38:41,630
preview of some other interviews that you've got

1078
00:38:41,630 --> 00:38:43,469
from Heidelberg that are coming up? I think

1079
00:38:43,469 --> 00:38:45,550
we've got two more, don't we? Yeah. I

1080
00:38:45,550 --> 00:38:47,965
spoke to two laureates at the Huddlberg Laureate

1081
00:38:47,965 --> 00:38:49,804
Forum who both have a background in physics

1082
00:38:49,804 --> 00:38:51,965
before they went into computer science, which is

1083
00:38:51,965 --> 00:38:53,644
where they've made their names and and won

1084
00:38:53,644 --> 00:38:54,304
their prizes.

1085
00:38:55,005 --> 00:38:57,885
Patrick Hanrahan is now best known for,

1086
00:38:58,684 --> 00:39:01,425
his work on at Pixar on computer animation.

1087
00:39:02,059 --> 00:39:05,260
And Amanda Randalls is working on computational sort

1088
00:39:05,260 --> 00:39:06,639
of health health science,

1089
00:39:07,099 --> 00:39:09,500
in the engineering department at Duke University. We

1090
00:39:09,500 --> 00:39:10,320
actually overlapped

1091
00:39:10,940 --> 00:39:12,780
there as physics students, so it was nice

1092
00:39:12,780 --> 00:39:14,219
to talk to her for that reason as

1093
00:39:14,219 --> 00:39:15,954
well. And so, yeah, they'll be coming up

1094
00:39:15,954 --> 00:39:17,635
on the podcast over the next couple of

1095
00:39:17,635 --> 00:39:19,394
months. We don't want to have all computer

1096
00:39:19,394 --> 00:39:21,394
science all the time, but, back to our

1097
00:39:21,394 --> 00:39:23,875
regular scheduled physics next week, I think. That's

1098
00:39:23,875 --> 00:39:26,195
right. Well, I look forward to, to, to

1099
00:39:26,195 --> 00:39:28,614
getting those, interviews onto the podcast.

1100
00:39:29,170 --> 00:39:31,889
Thanks, Margaret, for, coming on the podcast and

1101
00:39:31,889 --> 00:39:33,190
for doing those interviews,

1102
00:39:33,650 --> 00:39:36,050
and a special thanks to our producer, Fred

1103
00:39:36,050 --> 00:39:36,550
Isles.

1104
00:39:36,929 --> 00:39:39,349
We'll be back again next week. Bye.