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