Reversible computing could help solve AI’s looming energy crisis

1
00:00:08,960 --> 00:00:11,919
Hello, and welcome to the Physics World weekly

2
00:00:11,919 --> 00:00:12,419
podcast.

3
00:00:12,794 --> 00:00:15,535
In this episode, we meet Hannah Earley,

4
00:00:15,914 --> 00:00:17,774
a mathematician and physicist

5
00:00:18,155 --> 00:00:21,375
who's cofounder of a company that is commercializing

6
00:00:22,154 --> 00:00:22,654
reversible

7
00:00:22,954 --> 00:00:23,454
computing.

8
00:00:24,154 --> 00:00:25,935
This paradigm has the potential

9
00:00:26,300 --> 00:00:28,879
to use less energy than conventional

10
00:00:29,339 --> 00:00:29,839
computation,

11
00:00:30,539 --> 00:00:33,579
something that could prove very useful for power

12
00:00:33,579 --> 00:00:34,079
hungry

13
00:00:34,460 --> 00:00:35,600
AI applications.

14
00:00:36,619 --> 00:00:39,979
Hannah talks to Physics World's Margaret Harris about

15
00:00:39,979 --> 00:00:40,640
the physics,

16
00:00:41,164 --> 00:00:42,704
engineering, and commercialization

17
00:00:43,725 --> 00:00:45,265
of reversible computing.

18
00:00:54,200 --> 00:00:56,280
My guest today is Hannah Early, the chief

19
00:00:56,280 --> 00:00:58,920
technology officer and cofounder of a startup called

20
00:00:58,920 --> 00:01:01,000
VerComputing that aims to build a new type

21
00:01:01,000 --> 00:01:03,899
of computer architecture based on reversible operations.

22
00:01:04,280 --> 00:01:05,979
Hello, Hannah. Welcome to the podcast.

23
00:01:06,734 --> 00:01:08,594
Hi, Margaret. Thanks for having me on.

24
00:01:09,295 --> 00:01:11,295
We talk a lot about quantum computing on

25
00:01:11,295 --> 00:01:13,534
the Physics World Weekly podcast. But while I

26
00:01:13,534 --> 00:01:15,855
understand that Ver's work is in some sense

27
00:01:15,855 --> 00:01:17,234
based on quantum principles,

28
00:01:17,614 --> 00:01:19,855
you're not building a quantum computer in the

29
00:01:19,855 --> 00:01:20,834
usual sense.

30
00:01:21,269 --> 00:01:22,629
Maybe you could start out by giving our

31
00:01:22,629 --> 00:01:25,209
listeners just a quick introduction to what reversible

32
00:01:25,269 --> 00:01:26,409
computing is.

33
00:01:27,189 --> 00:01:29,670
Yeah. Of course. So, in fact, actually, there's

34
00:01:29,670 --> 00:01:32,469
not very much similarity to what we're doing

35
00:01:32,469 --> 00:01:34,629
in quantum computing at all except for the

36
00:01:34,629 --> 00:01:37,134
reversible principles at at the core of it.

37
00:01:37,134 --> 00:01:37,875
So, obviously, in

38
00:01:38,254 --> 00:01:41,055
quantum computers, it's very necessary that all the

39
00:01:41,055 --> 00:01:44,414
operations are unitary. And as a consequence, if

40
00:01:44,414 --> 00:01:47,055
you're doing any classical computation on a quantum

41
00:01:47,055 --> 00:01:47,555
computer,

42
00:01:48,174 --> 00:01:50,939
that means that it is logically reversible, which

43
00:01:50,939 --> 00:01:53,500
means that you can always from once they

44
00:01:53,500 --> 00:01:55,819
get back to the previous state. And that's

45
00:01:55,819 --> 00:01:58,060
very necessary in quantum computing because otherwise, you're

46
00:01:58,060 --> 00:01:59,040
going to get decoherence.

47
00:01:59,980 --> 00:02:02,799
In what we're doing in classical reversible computing,

48
00:02:03,340 --> 00:02:05,935
it's not kind of a physical necessity,

49
00:02:06,394 --> 00:02:08,235
but what it lets us do is,

50
00:02:08,634 --> 00:02:10,875
if we can not only make this computer

51
00:02:10,875 --> 00:02:11,854
logically reversible,

52
00:02:12,314 --> 00:02:14,414
but also also physically reversible,

53
00:02:14,715 --> 00:02:17,134
then that in principle lets us access

54
00:02:17,650 --> 00:02:19,909
significantly lower energy operations

55
00:02:20,210 --> 00:02:22,870
than than conventionally. So this dates back to,

56
00:02:23,250 --> 00:02:24,310
some work from

57
00:02:24,849 --> 00:02:26,849
Rolf Landau and if you trace it back

58
00:02:26,849 --> 00:02:29,189
even further to Leo Szilard and,

59
00:02:29,569 --> 00:02:30,710
Maxwell himself

60
00:02:31,055 --> 00:02:31,555
on

61
00:02:31,935 --> 00:02:34,594
what the energy costs or thermodynamics

62
00:02:35,694 --> 00:02:37,955
are when it comes to information processing.

63
00:02:38,574 --> 00:02:39,314
And so

64
00:02:39,694 --> 00:02:42,014
whilst we're not necessarily trying to break what

65
00:02:42,014 --> 00:02:43,854
is has become known as the Landau limit,

66
00:02:43,854 --> 00:02:46,360
it turns out that there is this limit

67
00:02:46,420 --> 00:02:47,640
when you want to

68
00:02:48,020 --> 00:02:50,580
erase information on on the energy cost of

69
00:02:50,580 --> 00:02:52,580
that. And this turns out to be a

70
00:02:52,580 --> 00:02:53,560
pretty small number,

71
00:02:54,260 --> 00:02:56,840
k t log two, k being Boltzmann's constant,

72
00:02:56,900 --> 00:02:58,040
and t being temperature,

73
00:02:58,504 --> 00:03:00,264
which is something like 10 to the minus

74
00:03:00,264 --> 00:03:02,824
21 joules. So it's a very small amount,

75
00:03:02,824 --> 00:03:05,224
but it turns out that, one, we're actually

76
00:03:05,224 --> 00:03:07,064
kind of closer than you might expect to

77
00:03:07,064 --> 00:03:08,284
this limit, but, two,

78
00:03:08,585 --> 00:03:10,840
even without trying to break this limit, by

79
00:03:10,840 --> 00:03:14,060
making your operations physically reversible, you can significantly

80
00:03:14,199 --> 00:03:18,219
reduce the energy cost of general operations even

81
00:03:18,360 --> 00:03:21,159
just when applying it to what we currently

82
00:03:21,159 --> 00:03:22,460
can do today in competing.

83
00:03:23,240 --> 00:03:25,180
Why is it that you have this

84
00:03:25,775 --> 00:03:26,275
incredible

85
00:03:26,574 --> 00:03:27,875
improvement in energy

86
00:03:28,175 --> 00:03:30,514
efficiency, I guess, if you have

87
00:03:30,814 --> 00:03:31,875
reversible operations?

88
00:03:33,134 --> 00:03:35,634
Yeah. So, it really depends on

89
00:03:36,495 --> 00:03:39,294
the computational medium you're talking about. So for

90
00:03:39,294 --> 00:03:42,150
example, in in CMOS computing, computing, which is,

91
00:03:42,389 --> 00:03:44,090
what we're doing at Ver computing,

92
00:03:44,870 --> 00:03:45,610
when you

93
00:03:46,550 --> 00:03:49,510
do an irreversible operation in a CMOS logic

94
00:03:49,510 --> 00:03:50,010
circuit,

95
00:03:50,469 --> 00:03:52,490
what this corresponds to is,

96
00:03:52,870 --> 00:03:55,365
so you take your logical circuit, and you

97
00:03:55,365 --> 00:03:57,604
have some inputs that are currently supplied to

98
00:03:57,604 --> 00:03:59,764
it and some output. And then later when

99
00:03:59,764 --> 00:04:01,625
you want to change that input,

100
00:04:01,925 --> 00:04:04,165
you're pretty much just you know, there might

101
00:04:04,165 --> 00:04:06,344
be some latch upstream or or some register.

102
00:04:06,724 --> 00:04:07,944
You change that input,

103
00:04:08,270 --> 00:04:09,729
and it's going to then

104
00:04:10,030 --> 00:04:12,449
kind of propagate through that circuit.

105
00:04:12,909 --> 00:04:15,729
And as it propagates through, it's going to

106
00:04:16,110 --> 00:04:18,990
change transistor connectivity, and it's going to then

107
00:04:18,990 --> 00:04:21,709
end up effectively flushing all of these signal

108
00:04:21,709 --> 00:04:22,209
energies

109
00:04:22,555 --> 00:04:24,095
into ground or into VDD.

110
00:04:24,474 --> 00:04:25,454
And as a consequence,

111
00:04:25,834 --> 00:04:26,495
you get

112
00:04:26,954 --> 00:04:27,774
this characteristic

113
00:04:28,794 --> 00:04:30,254
c v squared dissipation.

114
00:04:30,794 --> 00:04:32,794
And so this is kind of just taken

115
00:04:32,794 --> 00:04:35,615
as most axiomatic in in CMOS computing.

116
00:04:36,474 --> 00:04:38,370
And for a very long time, this was

117
00:04:39,009 --> 00:04:41,410
kind of negligible. And for the amount of

118
00:04:41,410 --> 00:04:42,930
computation we wanted to do, we had more

119
00:04:42,930 --> 00:04:45,169
than enough energy, and so it wasn't really

120
00:04:45,169 --> 00:04:46,310
seen as an issue.

121
00:04:46,850 --> 00:04:49,649
But when you change, at least the CMOS

122
00:04:49,649 --> 00:04:51,189
logic circuit to operate reversibly,

123
00:04:52,204 --> 00:04:54,044
what that corresponds to is adding in a

124
00:04:54,044 --> 00:04:56,524
step before you change the input, and that

125
00:04:56,524 --> 00:04:57,664
step is to

126
00:04:57,964 --> 00:05:00,524
first recover that signal energy that is stored

127
00:05:00,524 --> 00:05:01,904
in the gates of transistors.

128
00:05:02,764 --> 00:05:04,685
And you so you recover that,

129
00:05:05,389 --> 00:05:07,949
probably storing it generally in in some kind

130
00:05:07,949 --> 00:05:08,689
of reservoir,

131
00:05:09,069 --> 00:05:12,269
maybe even in inductors magnetic field. And then

132
00:05:12,269 --> 00:05:14,430
once you've recovered that technology, you've put that

133
00:05:14,430 --> 00:05:16,110
circuit into a neutral state, and then you

134
00:05:16,110 --> 00:05:16,850
can supply

135
00:05:17,310 --> 00:05:19,329
new inputs into that circuit.

136
00:05:19,709 --> 00:05:22,085
And when you do that, you're no longer

137
00:05:22,464 --> 00:05:24,225
there's a little bit more complexity to it,

138
00:05:24,225 --> 00:05:26,545
but effectively, you're trying to avoid setting up

139
00:05:26,545 --> 00:05:27,045
these

140
00:05:27,665 --> 00:05:30,805
quite significant potential differences that lead to dissipation.

141
00:05:31,665 --> 00:05:33,425
This is dissipation of heat. Right? You know,

142
00:05:33,425 --> 00:05:35,264
this is this is heat heat that is

143
00:05:35,264 --> 00:05:36,629
generated when you erase

144
00:05:37,110 --> 00:05:39,509
a register of bits and just dump that

145
00:05:39,509 --> 00:05:40,009
information

146
00:05:40,389 --> 00:05:41,930
to the environment.

147
00:05:42,870 --> 00:05:43,689
Yes. Exactly.

148
00:05:44,230 --> 00:05:46,970
How does reversible computing solve that problem?

149
00:05:48,149 --> 00:05:50,230
Reversible computing really solves that problem by just

150
00:05:50,230 --> 00:05:52,085
not generating that heat to begin At the

151
00:05:52,085 --> 00:05:54,965
same signal energy is flowing through the circuit,

152
00:05:55,205 --> 00:05:57,705
at least when you're doing reversible CMOS.

153
00:05:58,165 --> 00:05:59,685
But because you're able to

154
00:06:00,485 --> 00:06:01,944
in principle, you can recover

155
00:06:02,405 --> 00:06:04,324
a large amount of that energy. The amount

156
00:06:04,324 --> 00:06:06,740
of that energy that you can recover depends

157
00:06:06,740 --> 00:06:08,759
on how much you slow down the computation.

158
00:06:09,779 --> 00:06:11,779
I I probably want to revisit that a

159
00:06:11,779 --> 00:06:13,779
bit later in what that means because it

160
00:06:13,779 --> 00:06:16,339
sounds like our computation is much, much slower,

161
00:06:16,339 --> 00:06:19,665
and, that's actually not necessarily the case.

162
00:06:20,125 --> 00:06:21,504
But depending on

163
00:06:21,805 --> 00:06:24,545
this factor by which you slow it down,

164
00:06:24,685 --> 00:06:26,384
that kind of linearly,

165
00:06:26,845 --> 00:06:29,884
proportionally reduces the amount of energy that gets

166
00:06:29,884 --> 00:06:31,264
dissipated in that operation.

167
00:06:32,439 --> 00:06:35,319
There's also other circuit components that enable this

168
00:06:35,319 --> 00:06:36,839
whole thing to work and those have their

169
00:06:36,839 --> 00:06:39,240
own dissipation. But as long as you optimize

170
00:06:39,240 --> 00:06:41,819
those and optimize the the slowdown,

171
00:06:42,520 --> 00:06:43,160
you can save

172
00:06:44,375 --> 00:06:45,355
I I won't say arbitrarily

173
00:06:46,055 --> 00:06:48,215
much energy, but really in CMOS, the limit

174
00:06:48,215 --> 00:06:50,154
seems to be about 4,000 times.

175
00:06:50,615 --> 00:06:52,455
And, you know, you still have the same

176
00:06:52,455 --> 00:06:54,615
currents flowing through. It's just that instead of

177
00:06:54,615 --> 00:06:55,675
having those currents

178
00:06:56,134 --> 00:06:58,615
dumped to ground when when you start a

179
00:06:58,615 --> 00:07:01,459
new computational cycle, you much more carefully manage

180
00:07:01,459 --> 00:07:02,600
that energy flow.

181
00:07:03,620 --> 00:07:05,139
And you mentioned you wanted to talk a

182
00:07:05,139 --> 00:07:06,740
bit more about it's not that this is

183
00:07:06,740 --> 00:07:09,060
a really slow computation, because that's the traditional

184
00:07:09,060 --> 00:07:10,360
way you do things without,

185
00:07:10,660 --> 00:07:13,139
exchanging heat. You do things adiabatically. It's a

186
00:07:13,139 --> 00:07:15,245
really slow process. It doesn't sort of disturb

187
00:07:15,245 --> 00:07:16,764
the system in any way. Is that not

188
00:07:16,764 --> 00:07:17,664
what you're doing?

189
00:07:18,524 --> 00:07:20,925
It actually is, but I'll explain why that's

190
00:07:20,925 --> 00:07:22,704
not surprisingly slow.

191
00:07:23,084 --> 00:07:25,564
So adiabatic operations are kind of the core

192
00:07:25,564 --> 00:07:27,964
of what we're doing. So usually your signals

193
00:07:27,964 --> 00:07:30,629
in in regular CMOS are as close to

194
00:07:30,629 --> 00:07:32,790
a square wave as you can get. And

195
00:07:32,790 --> 00:07:34,790
we change those waves to be more what

196
00:07:34,790 --> 00:07:36,709
we like to call trapezoidal. So you have

197
00:07:36,709 --> 00:07:39,449
kind of flat regions where signals are stable,

198
00:07:39,509 --> 00:07:41,990
and then you have these linear ramps over

199
00:07:41,990 --> 00:07:45,134
a relatively long rise of full time. And

200
00:07:45,134 --> 00:07:47,394
that's where you kind of get the adiabaticness

201
00:07:47,615 --> 00:07:48,834
of what we're doing. So

202
00:07:49,375 --> 00:07:51,214
our approach is really a combination of both

203
00:07:51,214 --> 00:07:53,474
adiabatic computing and reversible computing.

204
00:07:54,495 --> 00:07:56,974
The reason why this is not horrendously slow,

205
00:07:56,974 --> 00:07:58,754
and this was really the worry

206
00:07:59,214 --> 00:07:59,954
back in

207
00:08:00,319 --> 00:08:02,720
the nineties when people were kind of first

208
00:08:02,720 --> 00:08:03,599
trying to build,

209
00:08:04,000 --> 00:08:06,319
these reversible computers. And and then it actually

210
00:08:06,319 --> 00:08:08,399
was more of a problem. The reason why

211
00:08:08,399 --> 00:08:10,579
this isn't a problem now is that

212
00:08:10,879 --> 00:08:12,019
this rise time

213
00:08:12,334 --> 00:08:14,115
is measured as a fraction of

214
00:08:14,574 --> 00:08:16,754
or as a sorry, as a multiple rather

215
00:08:17,134 --> 00:08:20,274
of the transistor's intrinsic switching time.

216
00:08:20,654 --> 00:08:23,055
And for modern processes, these can be on

217
00:08:23,055 --> 00:08:24,895
the order of picoseconds or even less than

218
00:08:24,895 --> 00:08:25,259
picoseconds,

219
00:08:26,139 --> 00:08:28,080
I e, terahertz frequencies.

220
00:08:29,020 --> 00:08:31,900
We obviously do not run our computers at

221
00:08:31,900 --> 00:08:33,980
terahertz frequencies, and there are very good reasons

222
00:08:33,980 --> 00:08:35,580
for that. And this kind of dates back

223
00:08:35,580 --> 00:08:37,580
to the end of de noid scaling back

224
00:08:37,580 --> 00:08:38,540
in 2005

225
00:08:38,540 --> 00:08:41,465
when before then, computational frequency seemed

226
00:08:41,924 --> 00:08:43,845
to double with a cadence similar to Moore's

227
00:08:43,845 --> 00:08:45,845
law, and and then afterwards, it kind of

228
00:08:45,845 --> 00:08:47,545
stagnates that a few gigahertz.

229
00:08:48,165 --> 00:08:50,884
But the transistor switching frequency kept kept going

230
00:08:50,884 --> 00:08:53,065
up. And so as long as we are

231
00:08:53,339 --> 00:08:56,459
significantly slower than this hundreds of gigahertz or

232
00:08:56,459 --> 00:08:57,199
even terahertz,

233
00:08:57,819 --> 00:08:59,360
then that's enough to get reversible

234
00:08:59,899 --> 00:09:02,319
efficiency. So we could operate in the gigahertz

235
00:09:02,379 --> 00:09:04,379
range, and that's still gonna give you 50

236
00:09:04,379 --> 00:09:06,639
or a 100 times energy saving

237
00:09:07,214 --> 00:09:09,855
in principle. I I did also mention that

238
00:09:09,855 --> 00:09:12,274
there are other circuit components that enable adiabatic

239
00:09:12,414 --> 00:09:13,394
switching, and

240
00:09:14,014 --> 00:09:16,274
those are pretty difficult to get good efficiencies

241
00:09:16,414 --> 00:09:18,654
on. And so those kind of end up

242
00:09:18,654 --> 00:09:19,154
dominating,

243
00:09:19,534 --> 00:09:21,920
but we can still get quite significant energy

244
00:09:21,920 --> 00:09:22,420
savings.

245
00:09:23,840 --> 00:09:26,100
But this sounds like, you know, really fascinating

246
00:09:26,240 --> 00:09:28,720
sort of concept in academic research. What is

247
00:09:28,720 --> 00:09:30,480
it that made you decide a few years

248
00:09:30,480 --> 00:09:32,399
ago now that now is the time to

249
00:09:32,399 --> 00:09:34,420
actually start commercializing this technology?

250
00:09:35,564 --> 00:09:38,684
Yeah. Great question. So it was kind of

251
00:09:38,684 --> 00:09:42,125
a almost serendipitous encounter between me and and

252
00:09:42,125 --> 00:09:44,225
my cofounder, Rodolfo Rustini.

253
00:09:46,605 --> 00:09:49,029
So I was doing a PhD in a

254
00:09:49,029 --> 00:09:52,549
number of unconventional computing topics, but, reversible computing

255
00:09:52,549 --> 00:09:54,169
being being one of the primary

256
00:09:54,789 --> 00:09:57,610
ones in that. And I had

257
00:09:58,070 --> 00:09:58,889
one of the

258
00:09:59,509 --> 00:10:01,684
topics I was looking at was what was

259
00:10:01,684 --> 00:10:04,245
the ultimate future of computing. And I kind

260
00:10:04,245 --> 00:10:05,865
of became very convinced

261
00:10:06,485 --> 00:10:06,985
that

262
00:10:07,365 --> 00:10:09,544
all future computers had to be

263
00:10:09,845 --> 00:10:12,264
at least involve a significant amount of reversibility

264
00:10:12,565 --> 00:10:15,440
if you wanted to keep increasing the performance

265
00:10:15,659 --> 00:10:17,839
of your larger and larger computers.

266
00:10:18,299 --> 00:10:18,959
Like, certainly

267
00:10:19,419 --> 00:10:20,940
in the long future, if you are thinking

268
00:10:20,940 --> 00:10:21,679
about building

269
00:10:22,459 --> 00:10:25,579
matrioshka brains and and other huge computers, then

270
00:10:25,579 --> 00:10:27,019
there's really no way to deal with the

271
00:10:27,019 --> 00:10:29,504
heat unless they use reversible computing. What's the

272
00:10:29,825 --> 00:10:31,024
sorry. I'm gonna stop you there. What's a

273
00:10:31,024 --> 00:10:32,164
matrioshka brain?

274
00:10:32,704 --> 00:10:34,945
Yeah. So I I might be mixing this

275
00:10:34,945 --> 00:10:37,184
up with Jupiter brains, but the idea is,

276
00:10:37,184 --> 00:10:40,084
you know, maybe these very advanced civilizations

277
00:10:40,625 --> 00:10:42,884
far beyond what what we are at

278
00:10:43,345 --> 00:10:44,404
might start to build

279
00:10:44,759 --> 00:10:45,740
computers the size

280
00:10:46,120 --> 00:10:48,860
of moons or planets or even larger astronomical

281
00:10:48,920 --> 00:10:49,420
systems.

282
00:10:50,120 --> 00:10:50,620
And

283
00:10:51,000 --> 00:10:53,160
the scaling laws turn out that if you

284
00:10:53,160 --> 00:10:54,220
just want to build

285
00:10:55,240 --> 00:10:58,084
a computer using irreversible techniques, you can only

286
00:10:58,084 --> 00:11:00,884
really cover the surface of some system in

287
00:11:00,884 --> 00:11:04,245
in computational matter. And that's purely because of

288
00:11:04,245 --> 00:11:05,065
the thermodynamics.

289
00:11:05,764 --> 00:11:06,664
If you're generating

290
00:11:07,204 --> 00:11:08,644
a certain amount of heat and you want

291
00:11:08,644 --> 00:11:10,644
to radiate that, then you're going to get

292
00:11:10,644 --> 00:11:13,160
some kind of area metric scaling law. If

293
00:11:13,160 --> 00:11:15,080
you want to go above that scaling law,

294
00:11:15,080 --> 00:11:17,320
then you really need to get this control

295
00:11:17,320 --> 00:11:18,620
over heat, and the only,

296
00:11:19,160 --> 00:11:21,500
approach that gives you that is is reversible

297
00:11:21,559 --> 00:11:22,059
computing.

298
00:11:22,680 --> 00:11:24,759
So I was looking maybe a little bit

299
00:11:24,759 --> 00:11:27,595
longer term, back then than what is maybe

300
00:11:27,595 --> 00:11:30,154
commercially practical right now. But I could also

301
00:11:30,154 --> 00:11:31,455
see that kind of

302
00:11:32,315 --> 00:11:33,835
it may be the case that in the

303
00:11:33,835 --> 00:11:36,235
nearer future, this might be relevant. And my

304
00:11:36,235 --> 00:11:36,735
cofounder

305
00:11:37,274 --> 00:11:39,115
was coming from a different direction. He was

306
00:11:39,115 --> 00:11:41,340
coming from very much looking at

307
00:11:42,139 --> 00:11:44,379
the growth of AI. And and this was

308
00:11:44,379 --> 00:11:45,899
back in 2021.

309
00:11:45,899 --> 00:11:48,860
So, you know, AI was becoming increasingly relevant,

310
00:11:48,860 --> 00:11:49,759
but we hadn't

311
00:11:50,300 --> 00:11:52,220
quite seen the explosion that we have in

312
00:11:52,220 --> 00:11:54,254
in the last couple years. It's been very

313
00:11:54,254 --> 00:11:56,254
fast how much this has increased. It feels

314
00:11:56,254 --> 00:11:58,595
like it's been around forever now. But

315
00:11:59,055 --> 00:12:00,274
he saw that

316
00:12:00,815 --> 00:12:02,975
we were potentially going to get a crisis

317
00:12:02,975 --> 00:12:05,795
in hardware, not least because there were

318
00:12:06,254 --> 00:12:08,195
increasing signs that Moore's Law

319
00:12:08,970 --> 00:12:10,409
is being predicted a number of times in

320
00:12:10,409 --> 00:12:12,169
the past, but perhaps this really was the

321
00:12:12,169 --> 00:12:13,450
time that Moore's law was going to come

322
00:12:13,450 --> 00:12:15,690
to an end. And so these combination of

323
00:12:15,690 --> 00:12:16,190
factors

324
00:12:16,730 --> 00:12:19,289
and a serendipitous meeting between us through a

325
00:12:19,289 --> 00:12:21,789
mutual friend led us realizing that maybe

326
00:12:22,605 --> 00:12:23,904
the solution to

327
00:12:24,285 --> 00:12:24,785
AI's

328
00:12:25,165 --> 00:12:28,524
upcoming energy problem would be reversible computing. And

329
00:12:28,524 --> 00:12:30,524
I wanna link back to what Feynman was

330
00:12:30,524 --> 00:12:33,165
saying in the nineteen eighties about reversible computing,

331
00:12:33,165 --> 00:12:34,065
which was that

332
00:12:35,085 --> 00:12:36,304
as long as you are

333
00:12:36,750 --> 00:12:39,470
significantly above the Landau limit, above a 100

334
00:12:39,470 --> 00:12:42,269
or 300 times the Landau limit, there's no

335
00:12:42,269 --> 00:12:44,669
need for you to ever consider reversible computing

336
00:12:44,669 --> 00:12:45,169
because

337
00:12:45,470 --> 00:12:47,470
we've got back in the eighties, we've got

338
00:12:47,470 --> 00:12:50,184
so much energy available, and our computation is

339
00:12:50,184 --> 00:12:52,985
already so inefficient, etcetera, and we're not really

340
00:12:52,985 --> 00:12:54,445
doing all that much computation.

341
00:12:55,225 --> 00:12:58,264
But now it's forty, fifty years later, and

342
00:12:58,264 --> 00:13:00,424
it turns out that we are actually at

343
00:13:00,424 --> 00:13:02,345
a few 100 times the landau limit. And

344
00:13:02,345 --> 00:13:02,845
so

345
00:13:03,490 --> 00:13:05,250
while it may not have been right in

346
00:13:05,250 --> 00:13:07,730
the nineties back when MIT built some reversible

347
00:13:07,730 --> 00:13:09,409
chips, it it looks like now might be

348
00:13:09,409 --> 00:13:10,069
the time.

349
00:13:10,929 --> 00:13:12,690
Okay. You talk about a chip. Right? Let's

350
00:13:12,690 --> 00:13:15,089
just get quite physical. What do logic gates

351
00:13:15,089 --> 00:13:17,144
look like in this technology? What do circuits

352
00:13:17,225 --> 00:13:19,004
look like in reversible computing?

353
00:13:20,024 --> 00:13:20,504
Yeah.

354
00:13:20,904 --> 00:13:23,004
So this is a great question. And

355
00:13:23,545 --> 00:13:25,144
it can be kind of hard to figure

356
00:13:25,144 --> 00:13:27,804
out what this is from the literature. And

357
00:13:28,264 --> 00:13:30,329
in a sense, this is because it really

358
00:13:30,329 --> 00:13:32,570
depends on on what you're building. So, obviously,

359
00:13:32,570 --> 00:13:33,850
I'm going to talk a lot about what

360
00:13:33,850 --> 00:13:36,329
this means in CMOS. But, traditionally, when you're

361
00:13:36,329 --> 00:13:38,509
looking at reversible computing, you see that

362
00:13:38,889 --> 00:13:42,169
relevant gates are gates or gates. And it's

363
00:13:42,169 --> 00:13:43,389
asking if you're building

364
00:13:43,894 --> 00:13:46,855
classical computations in a quantum computer. These are

365
00:13:46,855 --> 00:13:48,615
the gates that make sense. What are those

366
00:13:48,615 --> 00:13:51,174
gates? Yes. Those those terminologies our listeners might

367
00:13:51,174 --> 00:13:53,754
not have heard of before. Sorry. Yes. In

368
00:13:53,815 --> 00:13:56,394
computers, when you're doing reversible classical computing,

369
00:13:57,509 --> 00:13:59,269
your gates need to be need to have

370
00:13:59,269 --> 00:14:01,129
the same number of inputs as outputs.

371
00:14:01,509 --> 00:14:02,250
And so

372
00:14:02,629 --> 00:14:03,529
whilst in

373
00:14:03,830 --> 00:14:04,970
irreversible computing,

374
00:14:05,509 --> 00:14:07,750
the universal gate might be, say, the NAND

375
00:14:07,750 --> 00:14:09,769
gate, which is two input, one output.

376
00:14:10,149 --> 00:14:10,649
In

377
00:14:11,134 --> 00:14:14,174
reversible computing, this universal gate might be something

378
00:14:14,174 --> 00:14:16,894
called the Tefoli gate. This was discovered in

379
00:14:16,894 --> 00:14:19,294
the nineteen eighties. It's well, you could argue

380
00:14:19,294 --> 00:14:21,054
it was discovered earlier, but it wasn't named

381
00:14:21,054 --> 00:14:22,975
in the nineteen eighties at least. And this

382
00:14:22,975 --> 00:14:25,879
is a three input, three output gate. And

383
00:14:25,879 --> 00:14:28,279
what it does is quite simple. Its first

384
00:14:28,279 --> 00:14:30,920
two inputs are just copies across, so let's

385
00:14:30,920 --> 00:14:32,600
call them a, b, and c. The first

386
00:14:32,600 --> 00:14:34,059
two outputs are a and b.

387
00:14:34,360 --> 00:14:36,379
The third output, we just XOR,

388
00:14:36,920 --> 00:14:39,184
the third input c with the product

389
00:14:39,965 --> 00:14:41,565
of both the logical and of a and

390
00:14:41,565 --> 00:14:42,065
b.

391
00:14:42,845 --> 00:14:45,404
And in quantum computing, you might also see

392
00:14:45,404 --> 00:14:48,045
this referred to as controlled controlled not or

393
00:14:48,045 --> 00:14:48,945
CC not.

394
00:14:49,565 --> 00:14:50,065
So

395
00:14:50,540 --> 00:14:52,379
a lot of the reversible circuits you see

396
00:14:52,379 --> 00:14:54,620
out there make use of of these kinds

397
00:14:54,620 --> 00:14:55,279
of gates.

398
00:14:55,740 --> 00:14:58,700
But seamless, it is interesting, and it actually

399
00:14:58,700 --> 00:14:59,200
deviates

400
00:14:59,580 --> 00:15:01,600
quite a bit from from this paradigm.

401
00:15:01,980 --> 00:15:03,679
And the reason for that is

402
00:15:04,985 --> 00:15:07,465
unlike in, say, a quantum gate where kind

403
00:15:07,465 --> 00:15:09,725
of as you put information

404
00:15:10,504 --> 00:15:12,205
into the inputs, it gets

405
00:15:12,585 --> 00:15:15,004
directly transformed in place to outputs.

406
00:15:15,384 --> 00:15:16,764
It's not really how

407
00:15:17,225 --> 00:15:18,809
switching based logic works.

408
00:15:19,210 --> 00:15:20,730
So if you look at, say, a NAND

409
00:15:20,730 --> 00:15:22,570
gate well, actually, let's take a NOT gate

410
00:15:22,570 --> 00:15:24,490
in CMOS. Right? Because that seems like an

411
00:15:24,490 --> 00:15:27,950
intrinsically reversible gate, and logically, it is. But

412
00:15:28,490 --> 00:15:30,350
physically, it's not necessarily

413
00:15:30,649 --> 00:15:32,634
reversible in CMOS. And the reason for this

414
00:15:32,634 --> 00:15:35,514
is you supply the input, to one side

415
00:15:35,514 --> 00:15:36,894
of of your not gate,

416
00:15:37,274 --> 00:15:39,295
and the output gets generated.

417
00:15:39,995 --> 00:15:42,394
But you haven't consumed the input. You haven't

418
00:15:42,394 --> 00:15:44,730
transformed the input into the output. And so,

419
00:15:44,730 --> 00:15:46,970
actually, you should more think of the not

420
00:15:46,970 --> 00:15:49,769
gate in CMOS as a one input, two

421
00:15:49,769 --> 00:15:52,590
output gate because it kind of intrinsically

422
00:15:52,970 --> 00:15:54,669
keeps around a copy of the input.

423
00:15:55,529 --> 00:15:56,684
And in that sense,

424
00:15:57,085 --> 00:15:58,544
actually, all CMOS gates

425
00:15:59,884 --> 00:16:00,945
just conventionally

426
00:16:02,284 --> 00:16:04,304
have the ability to be used reversibly.

427
00:16:04,845 --> 00:16:07,404
And so the gates we use at fair

428
00:16:07,404 --> 00:16:10,230
computing are not that different from the conventional

429
00:16:10,230 --> 00:16:12,389
gates you'd find in in any standard cell

430
00:16:12,389 --> 00:16:12,889
library.

431
00:16:13,590 --> 00:16:14,409
Rather, we

432
00:16:14,870 --> 00:16:17,750
operate them quite a bit differently from how

433
00:16:17,750 --> 00:16:18,570
they're conventionally

434
00:16:19,029 --> 00:16:21,509
driven. And so we take a lot greater

435
00:16:21,509 --> 00:16:23,750
care of how signals propagate, and we add

436
00:16:23,750 --> 00:16:25,705
a little bit of extra circuitry around so

437
00:16:25,705 --> 00:16:26,445
that we can

438
00:16:26,825 --> 00:16:29,465
kind of both compute a gate, so generate

439
00:16:29,465 --> 00:16:31,865
its output from its input, and also decompose

440
00:16:31,865 --> 00:16:34,264
a gate so that whilst holding its input,

441
00:16:34,264 --> 00:16:35,085
you can actually

442
00:16:35,465 --> 00:16:38,460
ungenerate the output. And so adding this control

443
00:16:38,600 --> 00:16:41,399
lets us transform pretty much any regular gate

444
00:16:41,399 --> 00:16:44,200
into a reversible gate. There's a little bit

445
00:16:44,200 --> 00:16:47,240
more implementation complexity, but but actually the gates

446
00:16:47,240 --> 00:16:49,160
and the logic itself are not not very

447
00:16:49,160 --> 00:16:51,384
much different from how you would build a

448
00:16:51,384 --> 00:16:51,884
conventional

449
00:16:52,264 --> 00:16:53,245
C West chip.

450
00:16:53,785 --> 00:16:55,144
And I think I read that you sort

451
00:16:55,144 --> 00:16:58,024
of store the energy in some sort of

452
00:16:58,024 --> 00:17:00,184
resonator in order to do the uncomputation to

453
00:17:00,184 --> 00:17:02,264
do the reverse operation to make the gate

454
00:17:02,264 --> 00:17:03,965
reversible. How does that work?

455
00:17:04,690 --> 00:17:05,170
Yeah.

456
00:17:05,570 --> 00:17:08,470
So this is the other critical component. So,

457
00:17:08,529 --> 00:17:10,210
you know, it's not enough to just make

458
00:17:10,210 --> 00:17:10,869
your circuit

459
00:17:11,170 --> 00:17:13,650
theoretically logically reversible. You need to add in

460
00:17:13,650 --> 00:17:16,470
some extra circuitry so that you can actually

461
00:17:17,424 --> 00:17:19,924
operatively reverse the the operations.

462
00:17:20,464 --> 00:17:23,184
So one of the simplest approaches is so

463
00:17:23,184 --> 00:17:24,644
as you mentioned, you know,

464
00:17:25,105 --> 00:17:27,825
most of these implementations use a resonator. So

465
00:17:27,825 --> 00:17:29,744
one of the simplest could be, say, an

466
00:17:29,744 --> 00:17:33,029
LC resonator like you might encounter in first

467
00:17:33,029 --> 00:17:34,950
year physics, so just an inductor and a

468
00:17:34,950 --> 00:17:35,450
capacitor.

469
00:17:36,150 --> 00:17:37,610
Now one of the key

470
00:17:39,269 --> 00:17:40,809
developments of the,

471
00:17:41,269 --> 00:17:44,230
MOSFET transistor of, the kind of transistors that

472
00:17:44,230 --> 00:17:45,285
came before is that

473
00:17:46,484 --> 00:17:49,705
its gate is effectively a capacitor.

474
00:17:50,404 --> 00:17:54,105
And so when you supply inputs to a

475
00:17:54,244 --> 00:17:55,625
seamless logic cell,

476
00:17:55,924 --> 00:17:58,164
what you are doing is storing energy on

477
00:17:58,164 --> 00:18:00,720
the capacitors of the gates of of that

478
00:18:00,720 --> 00:18:03,839
logic cell, and then the logic cell will

479
00:18:03,839 --> 00:18:06,259
then generate outputs, and those outputs will

480
00:18:06,720 --> 00:18:09,059
then drive additional capacitive gates.

481
00:18:09,920 --> 00:18:12,579
And so if you can arrange your circuit

482
00:18:13,105 --> 00:18:13,684
so that

483
00:18:14,065 --> 00:18:15,444
these capacitive gates

484
00:18:16,304 --> 00:18:16,884
are effectively

485
00:18:17,265 --> 00:18:18,484
one big capacitor

486
00:18:19,024 --> 00:18:21,505
and then you tie that to an inductor,

487
00:18:21,505 --> 00:18:22,005
then

488
00:18:22,625 --> 00:18:24,544
you're already most of the way there because

489
00:18:24,544 --> 00:18:27,444
now you've built an LC circuit. And so

490
00:18:27,720 --> 00:18:29,319
at some points in time, that,

491
00:18:30,119 --> 00:18:33,319
inductor will have all of the energy stored

492
00:18:33,319 --> 00:18:35,079
within it, and the capacitors will be in

493
00:18:35,079 --> 00:18:37,400
some well, they will have no energy stored,

494
00:18:37,400 --> 00:18:39,339
and they will be computationally neutral.

495
00:18:40,085 --> 00:18:42,404
And then at a later time, that energy

496
00:18:42,404 --> 00:18:44,484
from that inductor can move onto the capacitive

497
00:18:44,484 --> 00:18:45,625
gates of those transistors.

498
00:18:46,005 --> 00:18:48,884
And then those gates are computationally active and

499
00:18:48,884 --> 00:18:50,265
you can generate an output.

500
00:18:50,805 --> 00:18:51,464
And then

501
00:18:52,190 --> 00:18:54,029
because this is an oscillatory circuit, they can

502
00:18:54,029 --> 00:18:55,549
then pull that energy back. And so that's

503
00:18:55,549 --> 00:18:57,170
kind of the fundamental principle.

504
00:18:58,509 --> 00:19:01,630
That's not enough because, you know, you need

505
00:19:01,630 --> 00:19:02,130
to

506
00:19:02,589 --> 00:19:05,309
then control how the outputs are generated. And

507
00:19:05,309 --> 00:19:06,990
so really you have kind of a number

508
00:19:06,990 --> 00:19:09,125
of these, not too many, but a number

509
00:19:09,125 --> 00:19:11,785
of these LC circuits. And so kind of

510
00:19:11,845 --> 00:19:13,704
you divide your computation into

511
00:19:14,085 --> 00:19:15,304
a few different stages,

512
00:19:15,605 --> 00:19:17,525
and then each of those stages has their

513
00:19:17,525 --> 00:19:18,025
own.

514
00:19:18,404 --> 00:19:19,684
Yeah. We we do it a little bit

515
00:19:19,684 --> 00:19:22,500
more compactly, but as a first order, you

516
00:19:22,500 --> 00:19:24,259
could imagine that each of these stages has

517
00:19:24,259 --> 00:19:25,240
their own inductor.

518
00:19:26,579 --> 00:19:29,299
Okay. What stage is Vericomputing at now? I

519
00:19:29,299 --> 00:19:30,899
think you're in the process of building a

520
00:19:30,899 --> 00:19:31,399
chip?

521
00:19:32,419 --> 00:19:35,444
Yeah. So we last year, we got our

522
00:19:35,444 --> 00:19:37,684
seed funding. We built our team, and we

523
00:19:37,684 --> 00:19:38,184
actually

524
00:19:38,644 --> 00:19:41,204
fully taped out our first well, we taped

525
00:19:41,204 --> 00:19:43,125
out our first test chip. It's not come

526
00:19:43,125 --> 00:19:45,704
back yet. So hopefully, we'll be able to

527
00:19:45,924 --> 00:19:48,164
announce that and its results in in the

528
00:19:48,164 --> 00:19:50,009
future. But yeah. So we've

529
00:19:50,650 --> 00:19:52,910
developed our first test chip. And what

530
00:19:53,369 --> 00:19:54,670
that chip does is

531
00:19:55,049 --> 00:19:55,549
really

532
00:19:56,089 --> 00:19:57,710
kind of bring together

533
00:19:58,329 --> 00:19:59,630
all of the different aspects

534
00:20:00,170 --> 00:20:03,769
of what a potentially commercially viable reversible chip

535
00:20:03,769 --> 00:20:05,994
would look like. So back in the nineties

536
00:20:05,994 --> 00:20:09,115
and more recently, people have made purely reversible

537
00:20:09,115 --> 00:20:11,515
chips and that the logic is reversible, but

538
00:20:11,515 --> 00:20:14,335
it doesn't actually have the capability to recover

539
00:20:14,394 --> 00:20:16,875
signal energy. And so our test chip, when

540
00:20:16,875 --> 00:20:18,710
we get it back and announce it, should

541
00:20:18,710 --> 00:20:20,710
be able to actually recover that energy within

542
00:20:20,710 --> 00:20:21,369
the system.

543
00:20:21,910 --> 00:20:23,830
And then we should be able to, well,

544
00:20:23,830 --> 00:20:25,930
measure how much how much better it is.

545
00:20:26,070 --> 00:20:28,710
And then what we're doing now is so

546
00:20:28,710 --> 00:20:31,190
so we want to commercialize this technology as

547
00:20:31,190 --> 00:20:32,570
quickly as possible. And

548
00:20:33,234 --> 00:20:35,234
really within the next few years, we think

549
00:20:35,234 --> 00:20:36,214
that there's

550
00:20:36,595 --> 00:20:39,154
a urgent demand for more energy efficient computing,

551
00:20:39,154 --> 00:20:41,954
but also energy efficient computing that doesn't look

552
00:20:41,954 --> 00:20:43,554
that different. Obviously, there are a lot of

553
00:20:43,554 --> 00:20:44,054
different

554
00:20:44,434 --> 00:20:47,650
various approaches to kind of making computing more

555
00:20:47,650 --> 00:20:50,929
energy efficient, but reversible computing's advantage, and this

556
00:20:50,929 --> 00:20:53,329
was actually something it was criticized for in

557
00:20:53,329 --> 00:20:54,309
the past that

558
00:20:54,690 --> 00:20:56,549
changes have been made, is that

559
00:20:56,929 --> 00:20:59,984
so reversible computing as we implement it has

560
00:20:59,984 --> 00:21:00,644
the same

561
00:21:01,505 --> 00:21:02,005
programming

562
00:21:02,625 --> 00:21:04,085
approach as conventional.

563
00:21:04,384 --> 00:21:06,464
And we, you know, have a little bit

564
00:21:06,464 --> 00:21:06,964
of

565
00:21:07,345 --> 00:21:09,744
extra complexity in in the logic to account

566
00:21:09,744 --> 00:21:11,125
for this. But effectively,

567
00:21:12,065 --> 00:21:14,404
you know, if you build, say, an inference

568
00:21:14,545 --> 00:21:15,045
accelerator,

569
00:21:15,390 --> 00:21:17,630
you should be able to just plug this

570
00:21:17,630 --> 00:21:21,230
into a server, launch PyTorch, and have it

571
00:21:21,230 --> 00:21:23,630
run. And so what we're doing this year

572
00:21:23,630 --> 00:21:24,130
is

573
00:21:24,430 --> 00:21:25,650
we are trying to

574
00:21:26,190 --> 00:21:28,430
take the print source we implemented in our

575
00:21:28,430 --> 00:21:31,170
test chip and really make them

576
00:21:31,505 --> 00:21:34,065
scalable so that we can build powerful chips

577
00:21:34,065 --> 00:21:34,565
and

578
00:21:34,865 --> 00:21:37,345
improve the efficiency of all the individual components.

579
00:21:37,585 --> 00:21:39,745
So last year was more making a proof

580
00:21:39,745 --> 00:21:42,384
of concept or proof of viability, and and

581
00:21:42,384 --> 00:21:44,005
now we're trying to get it to be

582
00:21:44,144 --> 00:21:46,679
actually something that people would want to buy.

583
00:21:47,299 --> 00:21:49,139
And what does success look like for you

584
00:21:49,139 --> 00:21:50,279
in the next few years?

585
00:21:51,299 --> 00:21:54,359
Yeah. So in the next few years, particularly

586
00:21:54,500 --> 00:21:56,259
by, you know, 2027,

587
00:21:56,259 --> 00:21:57,399
2028,

588
00:21:57,460 --> 00:21:58,359
starting to

589
00:21:58,819 --> 00:21:59,720
sell actual

590
00:22:00,194 --> 00:22:00,694
reversible

591
00:22:01,315 --> 00:22:01,815
chips

592
00:22:02,194 --> 00:22:02,694
that,

593
00:22:03,234 --> 00:22:04,615
you know, you could, say,

594
00:22:05,075 --> 00:22:05,815
put in

595
00:22:06,274 --> 00:22:08,994
a data center or put in a mobile

596
00:22:08,994 --> 00:22:10,515
device. You know, there are lots of different

597
00:22:10,515 --> 00:22:12,839
applications. Right? You can make this much more,

598
00:22:13,319 --> 00:22:15,980
obviously, it's intrinsically more energy efficient. But

599
00:22:16,839 --> 00:22:18,200
what that means, you know, whether you want

600
00:22:18,200 --> 00:22:19,720
this in a low power device or whether

601
00:22:19,720 --> 00:22:21,400
that just means you want to do even

602
00:22:21,400 --> 00:22:23,480
more computation for the same amount of energy,

603
00:22:23,480 --> 00:22:25,659
that's that's something you can play around with.

604
00:22:25,960 --> 00:22:27,765
So success would look like we want to

605
00:22:27,765 --> 00:22:29,765
be selling products in a few years. We

606
00:22:29,765 --> 00:22:31,065
want this to actually be

607
00:22:31,605 --> 00:22:33,445
to move out of the lab, out of

608
00:22:33,445 --> 00:22:35,384
academia. We want this to be something

609
00:22:36,005 --> 00:22:37,465
that is commercially viable.

610
00:22:38,404 --> 00:22:40,164
And then we want to build off that

611
00:22:40,164 --> 00:22:42,805
and keep improving the energy efficiency of this

612
00:22:42,805 --> 00:22:45,420
and really get to a new scaling law,

613
00:22:45,420 --> 00:22:45,920
something

614
00:22:46,460 --> 00:22:48,380
akin to Moore's law, something not going to

615
00:22:48,380 --> 00:22:51,599
be identical. You know, scaling down of transistor

616
00:22:51,660 --> 00:22:55,200
size pretty much hitting its its limit, but

617
00:22:55,914 --> 00:22:57,914
we think that, you know, at least energy

618
00:22:57,914 --> 00:23:01,355
efficiency, you can keep doubling that every cell

619
00:23:01,355 --> 00:23:01,855
phone,

620
00:23:02,234 --> 00:23:04,634
maybe all throughout there every two years, but,

621
00:23:04,955 --> 00:23:07,214
that could be ambitious, could be unambitious.

622
00:23:07,994 --> 00:23:09,295
And we want to then,

623
00:23:09,950 --> 00:23:13,150
effectively in twenty years time, have computation that's

624
00:23:13,150 --> 00:23:15,869
a few thousand times more energy efficient than

625
00:23:15,869 --> 00:23:16,690
it is today.

626
00:23:17,150 --> 00:23:20,190
And the ultimate success would be if this

627
00:23:20,190 --> 00:23:23,309
becomes kind of a fundamental part of how

628
00:23:23,309 --> 00:23:23,970
you build

629
00:23:24,484 --> 00:23:26,484
most computing systems. There's always gonna be a

630
00:23:26,484 --> 00:23:29,924
need for traditional irreversible computing, and the reason

631
00:23:29,924 --> 00:23:30,744
for that is

632
00:23:31,125 --> 00:23:32,744
that reversible computing excels

633
00:23:33,045 --> 00:23:36,244
at parallel tasks. But for very serial tasks,

634
00:23:36,244 --> 00:23:36,569
it's

635
00:23:37,529 --> 00:23:38,669
this slowdown

636
00:23:39,210 --> 00:23:41,609
becomes more more significant. So there's always gonna

637
00:23:41,609 --> 00:23:43,549
be a need for kind of a CPU

638
00:23:43,609 --> 00:23:47,069
type architecture, but perhaps anything which is parallel,

639
00:23:47,450 --> 00:23:49,769
might lead to a more reversible architecture. And

640
00:23:49,769 --> 00:23:53,125
maybe in ten, twenty years' time, every computer

641
00:23:53,125 --> 00:23:54,484
you buy might have a little bit of

642
00:23:54,484 --> 00:23:56,964
reversibility or maybe a lot of reversibility in

643
00:23:56,964 --> 00:23:57,464
it.

644
00:23:57,845 --> 00:23:59,525
That's fascinating. We'll have to check back with

645
00:23:59,525 --> 00:24:01,045
you in a few years. It'll be interesting

646
00:24:01,045 --> 00:24:03,444
to see particularly how this develops in parallel

647
00:24:03,444 --> 00:24:05,924
with quantum computing, which I think, likewise, people

648
00:24:05,924 --> 00:24:07,065
who work in that field

649
00:24:07,410 --> 00:24:09,410
recognize that quantum computers are not gonna do

650
00:24:09,410 --> 00:24:09,910
everything,

651
00:24:10,289 --> 00:24:11,970
but they may do some things well. It'll

652
00:24:11,970 --> 00:24:13,029
be interesting to see

653
00:24:13,570 --> 00:24:16,150
an evolution maybe beyond the sort of monolithic

654
00:24:16,369 --> 00:24:20,130
CMOS technology and monolithic irreversible CMOS technology that

655
00:24:20,130 --> 00:24:22,365
we have at the moment towards this quantum

656
00:24:22,424 --> 00:24:23,625
area and towards this,

657
00:24:24,184 --> 00:24:25,404
reversible computing

658
00:24:25,865 --> 00:24:26,365
paradigm.

659
00:24:27,304 --> 00:24:29,224
Yeah. It'd be really interesting to see where

660
00:24:29,224 --> 00:24:29,724
heterogeneous

661
00:24:30,184 --> 00:24:30,684
architectures

662
00:24:31,065 --> 00:24:32,904
end up going. And so obviously, we've had

663
00:24:32,904 --> 00:24:34,924
a huge amount of success in the traditional

664
00:24:35,650 --> 00:24:37,750
digital programming model, and

665
00:24:38,529 --> 00:24:40,049
there were a lot of advances that and

666
00:24:40,049 --> 00:24:42,470
that will probably never go never go away.

667
00:24:42,690 --> 00:24:43,190
But

668
00:24:43,650 --> 00:24:45,250
we've made a lot of progress in the

669
00:24:45,250 --> 00:24:45,750
last

670
00:24:46,609 --> 00:24:50,244
few decades in quantum and analog and photonic

671
00:24:50,384 --> 00:24:52,644
and and all of these other computational paradigms.

672
00:24:52,944 --> 00:24:53,684
And so

673
00:24:54,464 --> 00:24:56,625
I I can very much see that maybe

674
00:24:56,625 --> 00:24:58,085
in in the future, you'll

675
00:24:58,384 --> 00:25:00,865
not just have a CPU and a GPU,

676
00:25:00,865 --> 00:25:02,464
but maybe all of these other,

677
00:25:02,865 --> 00:25:04,369
types of computation

678
00:25:04,910 --> 00:25:07,869
embedded. Maybe not into your into your cell

679
00:25:07,869 --> 00:25:09,329
phone, but maybe into

680
00:25:09,710 --> 00:25:10,210
supercomputing

681
00:25:10,509 --> 00:25:12,369
clusters and data centers at least.

682
00:25:13,390 --> 00:25:15,390
Hannah Early, thank you much for joining us

683
00:25:15,390 --> 00:25:16,210
in the podcast.

684
00:25:16,829 --> 00:25:17,809
Thank you, Margaret.

685
00:25:27,164 --> 00:25:30,525
That was Margaret Harris in conversation with Hannah

686
00:25:30,525 --> 00:25:31,025
Earley,

687
00:25:31,349 --> 00:25:31,849
cofounder

688
00:25:32,230 --> 00:25:32,970
of VerComputing.

689
00:25:34,069 --> 00:25:36,650
You can find out more about Hannah's journey

690
00:25:36,789 --> 00:25:40,789
from getting a PhD in applied mathematics and

691
00:25:40,789 --> 00:25:42,089
theoretical physics

692
00:25:42,549 --> 00:25:44,089
to becoming the cofounder

693
00:25:44,549 --> 00:25:46,089
of a start up company

694
00:25:46,585 --> 00:25:47,884
in the career section

695
00:25:48,184 --> 00:25:49,485
of Physics World.

696
00:25:50,025 --> 00:25:51,325
Just look for the headline,

697
00:25:51,705 --> 00:25:52,924
Ask Me Anything.

698
00:25:53,305 --> 00:25:56,904
Hannah Earley. I love theory, but seeing an

699
00:25:56,904 --> 00:25:59,965
idea get closer and closer to reality

700
00:26:00,345 --> 00:26:01,164
is great.

701
00:26:01,730 --> 00:26:03,570
I'm afraid that's all the time we have

702
00:26:03,570 --> 00:26:06,369
for this week's podcast. Thanks to Hannah and

703
00:26:06,369 --> 00:26:08,070
Margaret for a fascinating

704
00:26:08,450 --> 00:26:10,549
introduction to reversible computing,

705
00:26:10,930 --> 00:26:13,730
and a special thanks to our producer, Fred

706
00:26:13,730 --> 00:26:14,230
Isles.

707
00:26:14,690 --> 00:26:17,585
We'll be back again next week. See you

708
00:26:21,484 --> 00:26:21,984
soon.