# Reversible computing could help solve AI’s looming energy crisis

An episode of [Physics World Weekly Podcast](https://invisiblenetwork.ai/podcast/1400686733), 26 Jun 2025, 27 min.

Audio: https://media.blubrry.com/physicsworldweeklypodcast/content.blubrry.com/physicsworldweeklypodcast/PWWeekly26JuneHannahEarleyVaireComputing.mp3

Hannah Earley of Vaire Computing is our podcast guest

## Transcript

[0:08] Hello, and welcome to the Physics World weekly
[0:11] podcast.
[0:12] In this episode, we meet Hannah Earley,
[0:15] a mathematician and physicist
[0:18] who's cofounder of a company that is commercializing
[0:22] reversible
[0:22] computing.
[0:24] This paradigm has the potential
[0:26] to use less energy than conventional
[0:29] computation,
[0:30] something that could prove very useful for power
[0:33] hungry
[0:34] AI applications.
[0:36] Hannah talks to Physics World's Margaret Harris about
[0:39] the physics,
[0:41] engineering, and commercialization
[0:43] of reversible computing.
[0:54] My guest today is Hannah Early, the chief
[0:56] technology officer and cofounder of a startup called
[0:58] VerComputing that aims to build a new type
[1:01] of computer architecture based on reversible operations.
[1:04] Hello, Hannah. Welcome to the podcast.
[1:06] Hi, Margaret. Thanks for having me on.
[1:09] We talk a lot about quantum computing on
[1:11] the Physics World Weekly podcast. But while I
[1:13] understand that Ver's work is in some sense
[1:15] based on quantum principles,
[1:17] you're not building a quantum computer in the
[1:19] usual sense.
[1:21] Maybe you could start out by giving our
[1:22] listeners just a quick introduction to what reversible
[1:25] computing is.
[1:27] Yeah. Of course. So, in fact, actually, there's
[1:29] not very much similarity to what we're doing
[1:32] in quantum computing at all except for the
[1:34] reversible principles at at the core of it.
[1:37] So, obviously, in
[1:38] quantum computers, it's very necessary that all the
[1:41] operations are unitary. And as a consequence, if
[1:44] you're doing any classical computation on a quantum
[1:47] computer,
[1:48] that means that it is logically reversible, which
[1:50] means that you can always from once they
[1:53] get back to the previous state. And that's
[1:55] very necessary in quantum computing because otherwise, you're
[1:58] going to get decoherence.
[1:59] In what we're doing in classical reversible computing,
[2:03] it's not kind of a physical necessity,
[2:06] but what it lets us do is,
[2:08] if we can not only make this computer
[2:10] logically reversible,
[2:12] but also also physically reversible,
[2:14] then that in principle lets us access
[2:17] significantly lower energy operations
[2:20] than than conventionally. So this dates back to,
[2:23] some work from
[2:24] Rolf Landau and if you trace it back
[2:26] even further to Leo Szilard and,
[2:29] Maxwell himself
[2:31] on
[2:31] what the energy costs or thermodynamics
[2:35] are when it comes to information processing.
[2:38] And so
[2:39] whilst we're not necessarily trying to break what
[2:42] is has become known as the Landau limit,
[2:43] it turns out that there is this limit
[2:46] when you want to
[2:48] erase information on on the energy cost of
[2:50] that. And this turns out to be a
[2:52] pretty small number,
[2:54] k t log two, k being Boltzmann's constant,
[2:56] and t being temperature,
[2:58] which is something like 10 to the minus
[3:00] 21 joules. So it's a very small amount,
[3:02] but it turns out that, one, we're actually
[3:05] kind of closer than you might expect to
[3:07] this limit, but, two,
[3:08] even without trying to break this limit, by
[3:10] making your operations physically reversible, you can significantly
[3:14] reduce the energy cost of general operations even
[3:18] just when applying it to what we currently
[3:21] can do today in competing.
[3:23] Why is it that you have this
[3:25] incredible
[3:26] improvement in energy
[3:28] efficiency, I guess, if you have
[3:30] reversible operations?
[3:33] Yeah. So, it really depends on
[3:36] the computational medium you're talking about. So for
[3:39] example, in in CMOS computing, computing, which is,
[3:42] what we're doing at Ver computing,
[3:44] when you
[3:46] do an irreversible operation in a CMOS logic
[3:49] circuit,
[3:50] what this corresponds to is,
[3:52] so you take your logical circuit, and you
[3:55] have some inputs that are currently supplied to
[3:57] it and some output. And then later when
[3:59] you want to change that input,
[4:01] you're pretty much just you know, there might
[4:04] be some latch upstream or or some register.
[4:06] You change that input,
[4:08] and it's going to then
[4:10] kind of propagate through that circuit.
[4:12] And as it propagates through, it's going to
[4:16] change transistor connectivity, and it's going to then
[4:18] end up effectively flushing all of these signal
[4:21] energies
[4:22] into ground or into VDD.
[4:24] And as a consequence,
[4:25] you get
[4:26] this characteristic
[4:28] c v squared dissipation.
[4:30] And so this is kind of just taken
[4:32] as most axiomatic in in CMOS computing.
[4:36] And for a very long time, this was
[4:39] kind of negligible. And for the amount of
[4:41] computation we wanted to do, we had more
[4:42] than enough energy, and so it wasn't really
[4:45] seen as an issue.
[4:46] But when you change, at least the CMOS
[4:49] logic circuit to operate reversibly,
[4:52] what that corresponds to is adding in a
[4:54] step before you change the input, and that
[4:56] step is to
[4:57] first recover that signal energy that is stored
[5:00] in the gates of transistors.
[5:02] And you so you recover that,
[5:05] probably storing it generally in in some kind
[5:07] of reservoir,
[5:09] maybe even in inductors magnetic field. And then
[5:12] once you've recovered that technology, you've put that
[5:14] circuit into a neutral state, and then you
[5:16] can supply
[5:17] new inputs into that circuit.
[5:19] And when you do that, you're no longer
[5:22] there's a little bit more complexity to it,
[5:24] but effectively, you're trying to avoid setting up
[5:26] these
[5:27] quite significant potential differences that lead to dissipation.
[5:31] This is dissipation of heat. Right? You know,
[5:33] this is this is heat heat that is
[5:35] generated when you erase
[5:37] a register of bits and just dump that
[5:39] information
[5:40] to the environment.
[5:42] Yes. Exactly.
[5:44] How does reversible computing solve that problem?
[5:48] Reversible computing really solves that problem by just
[5:50] not generating that heat to begin At the
[5:52] same signal energy is flowing through the circuit,
[5:55] at least when you're doing reversible CMOS.
[5:58] But because you're able to
[6:00] in principle, you can recover
[6:02] a large amount of that energy. The amount
[6:04] of that energy that you can recover depends
[6:06] on how much you slow down the computation.
[6:09] I I probably want to revisit that a
[6:11] bit later in what that means because it
[6:13] sounds like our computation is much, much slower,
[6:16] and, that's actually not necessarily the case.
[6:20] But depending on
[6:21] this factor by which you slow it down,
[6:24] that kind of linearly,
[6:26] proportionally reduces the amount of energy that gets
[6:29] dissipated in that operation.
[6:32] There's also other circuit components that enable this
[6:35] whole thing to work and those have their
[6:36] own dissipation. But as long as you optimize
[6:39] those and optimize the the slowdown,
[6:42] you can save
[6:44] I I won't say arbitrarily
[6:46] much energy, but really in CMOS, the limit
[6:48] seems to be about 4,000 times.
[6:50] And, you know, you still have the same
[6:52] currents flowing through. It's just that instead of
[6:54] having those currents
[6:56] dumped to ground when when you start a
[6:58] new computational cycle, you much more carefully manage
[7:01] that energy flow.
[7:03] And you mentioned you wanted to talk a
[7:05] bit more about it's not that this is
[7:06] a really slow computation, because that's the traditional
[7:09] way you do things without,
[7:10] exchanging heat. You do things adiabatically. It's a
[7:13] really slow process. It doesn't sort of disturb
[7:15] the system in any way. Is that not
[7:16] what you're doing?
[7:18] It actually is, but I'll explain why that's
[7:20] not surprisingly slow.
[7:23] So adiabatic operations are kind of the core
[7:25] of what we're doing. So usually your signals
[7:27] in in regular CMOS are as close to
[7:30] a square wave as you can get. And
[7:32] we change those waves to be more what
[7:34] we like to call trapezoidal. So you have
[7:36] kind of flat regions where signals are stable,
[7:39] and then you have these linear ramps over
[7:41] a relatively long rise of full time. And
[7:45] that's where you kind of get the adiabaticness
[7:47] of what we're doing. So
[7:49] our approach is really a combination of both
[7:51] adiabatic computing and reversible computing.
[7:54] The reason why this is not horrendously slow,
[7:56] and this was really the worry
[7:59] back in
[8:00] the nineties when people were kind of first
[8:02] trying to build,
[8:04] these reversible computers. And and then it actually
[8:06] was more of a problem. The reason why
[8:08] this isn't a problem now is that
[8:10] this rise time
[8:12] is measured as a fraction of
[8:14] or as a sorry, as a multiple rather
[8:17] of the transistor's intrinsic switching time.
[8:20] And for modern processes, these can be on
[8:23] the order of picoseconds or even less than
[8:24] picoseconds,
[8:26] I e, terahertz frequencies.
[8:29] We obviously do not run our computers at
[8:31] terahertz frequencies, and there are very good reasons
[8:33] for that. And this kind of dates back
[8:35] to the end of de noid scaling back
[8:37] in 2005
[8:38] when before then, computational frequency seemed
[8:41] to double with a cadence similar to Moore's
[8:43] law, and and then afterwards, it kind of
[8:45] stagnates that a few gigahertz.
[8:48] But the transistor switching frequency kept kept going
[8:50] up. And so as long as we are
[8:53] significantly slower than this hundreds of gigahertz or
[8:56] even terahertz,
[8:57] then that's enough to get reversible
[8:59] efficiency. So we could operate in the gigahertz
[9:02] range, and that's still gonna give you 50
[9:04] or a 100 times energy saving
[9:07] in principle. I I did also mention that
[9:09] there are other circuit components that enable adiabatic
[9:12] switching, and
[9:14] those are pretty difficult to get good efficiencies
[9:16] on. And so those kind of end up
[9:18] dominating,
[9:19] but we can still get quite significant energy
[9:21] savings.
[9:23] But this sounds like, you know, really fascinating
[9:26] sort of concept in academic research. What is
[9:28] it that made you decide a few years
[9:30] ago now that now is the time to
[9:32] actually start commercializing this technology?
[9:35] Yeah. Great question. So it was kind of
[9:38] a almost serendipitous encounter between me and and
[9:42] my cofounder, Rodolfo Rustini.
[9:46] So I was doing a PhD in a
[9:49] number of unconventional computing topics, but, reversible computing
[9:52] being being one of the primary
[9:54] ones in that. And I had
[9:58] one of the
[9:59] topics I was looking at was what was
[10:01] the ultimate future of computing. And I kind
[10:04] of became very convinced
[10:06] that
[10:07] all future computers had to be
[10:09] at least involve a significant amount of reversibility
[10:12] if you wanted to keep increasing the performance
[10:15] of your larger and larger computers.
[10:18] Like, certainly
[10:19] in the long future, if you are thinking
[10:20] about building
[10:22] matrioshka brains and and other huge computers, then
[10:25] there's really no way to deal with the
[10:27] heat unless they use reversible computing. What's the
[10:29] sorry. I'm gonna stop you there. What's a
[10:31] matrioshka brain?
[10:32] Yeah. So I I might be mixing this
[10:34] up with Jupiter brains, but the idea is,
[10:37] you know, maybe these very advanced civilizations
[10:40] far beyond what what we are at
[10:43] might start to build
[10:44] computers the size
[10:46] of moons or planets or even larger astronomical
[10:48] systems.
[10:50] And
[10:51] the scaling laws turn out that if you
[10:53] just want to build
[10:55] a computer using irreversible techniques, you can only
[10:58] really cover the surface of some system in
[11:00] in computational matter. And that's purely because of
[11:04] the thermodynamics.
[11:05] If you're generating
[11:07] a certain amount of heat and you want
[11:08] to radiate that, then you're going to get
[11:10] some kind of area metric scaling law. If
[11:13] you want to go above that scaling law,
[11:15] then you really need to get this control
[11:17] over heat, and the only,
[11:19] approach that gives you that is is reversible
[11:21] computing.
[11:22] So I was looking maybe a little bit
[11:24] longer term, back then than what is maybe
[11:27] commercially practical right now. But I could also
[11:30] see that kind of
[11:32] it may be the case that in the
[11:33] nearer future, this might be relevant. And my
[11:36] cofounder
[11:37] was coming from a different direction. He was
[11:39] coming from very much looking at
[11:42] the growth of AI. And and this was
[11:44] back in 2021.
[11:45] So, you know, AI was becoming increasingly relevant,
[11:48] but we hadn't
[11:50] quite seen the explosion that we have in
[11:52] in the last couple years. It's been very
[11:54] fast how much this has increased. It feels
[11:56] like it's been around forever now. But
[11:59] he saw that
[12:00] we were potentially going to get a crisis
[12:02] in hardware, not least because there were
[12:06] increasing signs that Moore's Law
[12:08] is being predicted a number of times in
[12:10] the past, but perhaps this really was the
[12:12] time that Moore's law was going to come
[12:13] to an end. And so these combination of
[12:15] factors
[12:16] and a serendipitous meeting between us through a
[12:19] mutual friend led us realizing that maybe
[12:22] the solution to
[12:24] AI's
[12:25] upcoming energy problem would be reversible computing. And
[12:28] I wanna link back to what Feynman was
[12:30] saying in the nineteen eighties about reversible computing,
[12:33] which was that
[12:35] as long as you are
[12:36] significantly above the Landau limit, above a 100
[12:39] or 300 times the Landau limit, there's no
[12:42] need for you to ever consider reversible computing
[12:44] because
[12:45] we've got back in the eighties, we've got
[12:47] so much energy available, and our computation is
[12:50] already so inefficient, etcetera, and we're not really
[12:52] doing all that much computation.
[12:55] But now it's forty, fifty years later, and
[12:58] it turns out that we are actually at
[13:00] a few 100 times the landau limit. And
[13:02] so
[13:03] while it may not have been right in
[13:05] the nineties back when MIT built some reversible
[13:07] chips, it it looks like now might be
[13:09] the time.
[13:10] Okay. You talk about a chip. Right? Let's
[13:12] just get quite physical. What do logic gates
[13:15] look like in this technology? What do circuits
[13:17] look like in reversible computing?
[13:20] Yeah.
[13:20] So this is a great question. And
[13:23] it can be kind of hard to figure
[13:25] out what this is from the literature. And
[13:28] in a sense, this is because it really
[13:30] depends on on what you're building. So, obviously,
[13:32] I'm going to talk a lot about what
[13:33] this means in CMOS. But, traditionally, when you're
[13:36] looking at reversible computing, you see that
[13:38] relevant gates are gates or gates. And it's
[13:42] asking if you're building
[13:43] classical computations in a quantum computer. These are
[13:46] the gates that make sense. What are those
[13:48] gates? Yes. Those those terminologies our listeners might
[13:51] not have heard of before. Sorry. Yes. In
[13:53] computers, when you're doing reversible classical computing,
[13:57] your gates need to be need to have
[13:59] the same number of inputs as outputs.
[14:01] And so
[14:02] whilst in
[14:03] irreversible computing,
[14:05] the universal gate might be, say, the NAND
[14:07] gate, which is two input, one output.
[14:10] In
[14:11] reversible computing, this universal gate might be something
[14:14] called the Tefoli gate. This was discovered in
[14:16] the nineteen eighties. It's well, you could argue
[14:19] it was discovered earlier, but it wasn't named
[14:21] in the nineteen eighties at least. And this
[14:22] is a three input, three output gate. And
[14:25] what it does is quite simple. Its first
[14:28] two inputs are just copies across, so let's
[14:30] call them a, b, and c. The first
[14:32] two outputs are a and b.
[14:34] The third output, we just XOR,
[14:36] the third input c with the product
[14:39] of both the logical and of a and
[14:41] b.
[14:42] And in quantum computing, you might also see
[14:45] this referred to as controlled controlled not or
[14:48] CC not.
[14:49] So
[14:50] a lot of the reversible circuits you see
[14:52] out there make use of of these kinds
[14:54] of gates.
[14:55] But seamless, it is interesting, and it actually
[14:58] deviates
[14:59] quite a bit from from this paradigm.
[15:01] And the reason for that is
[15:04] unlike in, say, a quantum gate where kind
[15:07] of as you put information
[15:10] into the inputs, it gets
[15:12] directly transformed in place to outputs.
[15:15] It's not really how
[15:17] switching based logic works.
[15:19] So if you look at, say, a NAND
[15:20] gate well, actually, let's take a NOT gate
[15:22] in CMOS. Right? Because that seems like an
[15:24] intrinsically reversible gate, and logically, it is. But
[15:28] physically, it's not necessarily
[15:30] reversible in CMOS. And the reason for this
[15:32] is you supply the input, to one side
[15:35] of of your not gate,
[15:37] and the output gets generated.
[15:39] But you haven't consumed the input. You haven't
[15:42] transformed the input into the output. And so,
[15:44] actually, you should more think of the not
[15:46] gate in CMOS as a one input, two
[15:49] output gate because it kind of intrinsically
[15:52] keeps around a copy of the input.
[15:55] And in that sense,
[15:57] actually, all CMOS gates
[15:59] just conventionally
[16:02] have the ability to be used reversibly.
[16:04] And so the gates we use at fair
[16:07] computing are not that different from the conventional
[16:10] gates you'd find in in any standard cell
[16:12] library.
[16:13] Rather, we
[16:14] operate them quite a bit differently from how
[16:17] they're conventionally
[16:19] driven. And so we take a lot greater
[16:21] care of how signals propagate, and we add
[16:23] a little bit of extra circuitry around so
[16:25] that we can
[16:26] kind of both compute a gate, so generate
[16:29] its output from its input, and also decompose
[16:31] a gate so that whilst holding its input,
[16:34] you can actually
[16:35] ungenerate the output. And so adding this control
[16:38] lets us transform pretty much any regular gate
[16:41] into a reversible gate. There's a little bit
[16:44] more implementation complexity, but but actually the gates
[16:47] and the logic itself are not not very
[16:49] much different from how you would build a
[16:51] conventional
[16:52] C West chip.
[16:53] And I think I read that you sort
[16:55] of store the energy in some sort of
[16:58] resonator in order to do the uncomputation to
[17:00] do the reverse operation to make the gate
[17:02] reversible. How does that work?
[17:04] Yeah.
[17:05] So this is the other critical component. So,
[17:08] you know, it's not enough to just make
[17:10] your circuit
[17:11] theoretically logically reversible. You need to add in
[17:13] some extra circuitry so that you can actually
[17:17] operatively reverse the the operations.
[17:20] So one of the simplest approaches is so
[17:23] as you mentioned, you know,
[17:25] most of these implementations use a resonator. So
[17:27] one of the simplest could be, say, an
[17:29] LC resonator like you might encounter in first
[17:33] year physics, so just an inductor and a
[17:34] capacitor.
[17:36] Now one of the key
[17:39] developments of the,
[17:41] MOSFET transistor of, the kind of transistors that
[17:44] came before is that
[17:46] its gate is effectively a capacitor.
[17:50] And so when you supply inputs to a
[17:54] seamless logic cell,
[17:55] what you are doing is storing energy on
[17:58] the capacitors of the gates of of that
[18:00] logic cell, and then the logic cell will
[18:03] then generate outputs, and those outputs will
[18:06] then drive additional capacitive gates.
[18:09] And so if you can arrange your circuit
[18:13] so that
[18:14] these capacitive gates
[18:16] are effectively
[18:17] one big capacitor
[18:19] and then you tie that to an inductor,
[18:21] then
[18:22] you're already most of the way there because
[18:24] now you've built an LC circuit. And so
[18:27] at some points in time, that,
[18:30] inductor will have all of the energy stored
[18:33] within it, and the capacitors will be in
[18:35] some well, they will have no energy stored,
[18:37] and they will be computationally neutral.
[18:40] And then at a later time, that energy
[18:42] from that inductor can move onto the capacitive
[18:44] gates of those transistors.
[18:46] And then those gates are computationally active and
[18:48] you can generate an output.
[18:50] And then
[18:52] because this is an oscillatory circuit, they can
[18:54] then pull that energy back. And so that's
[18:55] kind of the fundamental principle.
[18:58] That's not enough because, you know, you need
[19:01] to
[19:02] then control how the outputs are generated. And
[19:05] so really you have kind of a number
[19:06] of these, not too many, but a number
[19:09] of these LC circuits. And so kind of
[19:11] you divide your computation into
[19:14] a few different stages,
[19:15] and then each of those stages has their
[19:17] own.
[19:18] Yeah. We we do it a little bit
[19:19] more compactly, but as a first order, you
[19:22] could imagine that each of these stages has
[19:24] their own inductor.
[19:26] Okay. What stage is Vericomputing at now? I
[19:29] think you're in the process of building a
[19:30] chip?
[19:32] Yeah. So we last year, we got our
[19:35] seed funding. We built our team, and we
[19:37] actually
[19:38] fully taped out our first well, we taped
[19:41] out our first test chip. It's not come
[19:43] back yet. So hopefully, we'll be able to
[19:45] announce that and its results in in the
[19:48] future. But yeah. So we've
[19:50] developed our first test chip. And what
[19:53] that chip does is
[19:55] really
[19:56] kind of bring together
[19:58] all of the different aspects
[20:00] of what a potentially commercially viable reversible chip
[20:03] would look like. So back in the nineties
[20:05] and more recently, people have made purely reversible
[20:09] chips and that the logic is reversible, but
[20:11] it doesn't actually have the capability to recover
[20:14] signal energy. And so our test chip, when
[20:16] we get it back and announce it, should
[20:18] be able to actually recover that energy within
[20:20] the system.
[20:21] And then we should be able to, well,
[20:23] measure how much how much better it is.
[20:26] And then what we're doing now is so
[20:28] so we want to commercialize this technology as
[20:31] quickly as possible. And
[20:33] really within the next few years, we think
[20:35] that there's
[20:36] a urgent demand for more energy efficient computing,
[20:39] but also energy efficient computing that doesn't look
[20:41] that different. Obviously, there are a lot of
[20:43] different
[20:44] various approaches to kind of making computing more
[20:47] energy efficient, but reversible computing's advantage, and this
[20:50] was actually something it was criticized for in
[20:53] the past that
[20:54] changes have been made, is that
[20:56] so reversible computing as we implement it has
[20:59] the same
[21:01] programming
[21:02] approach as conventional.
[21:04] And we, you know, have a little bit
[21:06] of
[21:07] extra complexity in in the logic to account
[21:09] for this. But effectively,
[21:12] you know, if you build, say, an inference
[21:14] accelerator,
[21:15] you should be able to just plug this
[21:17] into a server, launch PyTorch, and have it
[21:21] run. And so what we're doing this year
[21:23] is
[21:24] we are trying to
[21:26] take the print source we implemented in our
[21:28] test chip and really make them
[21:31] scalable so that we can build powerful chips
[21:34] and
[21:34] improve the efficiency of all the individual components.
[21:37] So last year was more making a proof
[21:39] of concept or proof of viability, and and
[21:42] now we're trying to get it to be
[21:44] actually something that people would want to buy.
[21:47] And what does success look like for you
[21:49] in the next few years?
[21:51] Yeah. So in the next few years, particularly
[21:54] by, you know, 2027,
[21:56] 2028,
[21:57] starting to
[21:58] sell actual
[22:00] reversible
[22:01] chips
[22:02] that,
[22:03] you know, you could, say,
[22:05] put in
[22:06] a data center or put in a mobile
[22:08] device. You know, there are lots of different
[22:10] applications. Right? You can make this much more,
[22:13] obviously, it's intrinsically more energy efficient. But
[22:16] what that means, you know, whether you want
[22:18] this in a low power device or whether
[22:19] that just means you want to do even
[22:21] more computation for the same amount of energy,
[22:23] that's that's something you can play around with.
[22:25] So success would look like we want to
[22:27] be selling products in a few years. We
[22:29] want this to actually be
[22:31] to move out of the lab, out of
[22:33] academia. We want this to be something
[22:36] that is commercially viable.
[22:38] And then we want to build off that
[22:40] and keep improving the energy efficiency of this
[22:42] and really get to a new scaling law,
[22:45] something
[22:46] akin to Moore's law, something not going to
[22:48] be identical. You know, scaling down of transistor
[22:51] size pretty much hitting its its limit, but
[22:55] we think that, you know, at least energy
[22:57] efficiency, you can keep doubling that every cell
[23:01] phone,
[23:02] maybe all throughout there every two years, but,
[23:04] that could be ambitious, could be unambitious.
[23:07] And we want to then,
[23:09] effectively in twenty years time, have computation that's
[23:13] a few thousand times more energy efficient than
[23:15] it is today.
[23:17] And the ultimate success would be if this
[23:20] becomes kind of a fundamental part of how
[23:23] you build
[23:24] most computing systems. There's always gonna be a
[23:26] need for traditional irreversible computing, and the reason
[23:29] for that is
[23:31] that reversible computing excels
[23:33] at parallel tasks. But for very serial tasks,
[23:36] it's
[23:37] this slowdown
[23:39] becomes more more significant. So there's always gonna
[23:41] be a need for kind of a CPU
[23:43] type architecture, but perhaps anything which is parallel,
[23:47] might lead to a more reversible architecture. And
[23:49] maybe in ten, twenty years' time, every computer
[23:53] you buy might have a little bit of
[23:54] reversibility or maybe a lot of reversibility in
[23:56] it.
[23:57] That's fascinating. We'll have to check back with
[23:59] you in a few years. It'll be interesting
[24:01] to see particularly how this develops in parallel
[24:03] with quantum computing, which I think, likewise, people
[24:05] who work in that field
[24:07] recognize that quantum computers are not gonna do
[24:09] everything,
[24:10] but they may do some things well. It'll
[24:11] be interesting to see
[24:13] an evolution maybe beyond the sort of monolithic
[24:16] CMOS technology and monolithic irreversible CMOS technology that
[24:20] we have at the moment towards this quantum
[24:22] area and towards this,
[24:24] reversible computing
[24:25] paradigm.
[24:27] Yeah. It'd be really interesting to see where
[24:29] heterogeneous
[24:30] architectures
[24:31] end up going. And so obviously, we've had
[24:32] a huge amount of success in the traditional
[24:35] digital programming model, and
[24:38] there were a lot of advances that and
[24:40] that will probably never go never go away.
[24:42] But
[24:43] we've made a lot of progress in the
[24:45] last
[24:46] few decades in quantum and analog and photonic
[24:50] and and all of these other computational paradigms.
[24:52] And so
[24:54] I I can very much see that maybe
[24:56] in in the future, you'll
[24:58] not just have a CPU and a GPU,
[25:00] but maybe all of these other,
[25:02] types of computation
[25:04] embedded. Maybe not into your into your cell
[25:07] phone, but maybe into
[25:09] supercomputing
[25:10] clusters and data centers at least.
[25:13] Hannah Early, thank you much for joining us
[25:15] in the podcast.
[25:16] Thank you, Margaret.
[25:27] That was Margaret Harris in conversation with Hannah
[25:30] Earley,
[25:31] cofounder
[25:32] of VerComputing.
[25:34] You can find out more about Hannah's journey
[25:36] from getting a PhD in applied mathematics and
[25:40] theoretical physics
[25:42] to becoming the cofounder
[25:44] of a start up company
[25:46] in the career section
[25:48] of Physics World.
[25:50] Just look for the headline,
[25:51] Ask Me Anything.
[25:53] Hannah Earley. I love theory, but seeing an
[25:56] idea get closer and closer to reality
[26:00] is great.
[26:01] I'm afraid that's all the time we have
[26:03] for this week's podcast. Thanks to Hannah and
[26:06] Margaret for a fascinating
[26:08] introduction to reversible computing,
[26:10] and a special thanks to our producer, Fred
[26:13] Isles.
[26:14] We'll be back again next week. See you
[26:21] soon.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.