B1 Intermediate US 5448 Folder Collection
After playing the video, you can click or select the word to look it up in the dictionary.
Report Subtitle Errors
MALE SPEAKER: Good afternoon.
Thanks, everybody, for coming from the remote sites
to attend the talk by John Martinis about the design
of a superconducting quantum computer.
And we're very pleased to have John
here with us, just a short ride from UC Santa Barbara.
And the reason we are excited is John
is considered one of the world, if not THE world
authority, on superconducting qubits.
So since the current machine we're working on
is based on superconducting qubits, of course,
his opinion and advice would be very important
for the guidance of our project.
So John got his PhD in physics in 1987
from UC Berkeley in California.
But then went to France to the Commisiariat Energie
Atomic in Saclay.
Afterwards, he worked in NIST in Boulder.
And then in 2004, he settled where he is right now,
being full professor at UC Santa Barbara.
And then in 2010, nice achievement,
getting the AAAS Science Breakthrough of the Year
award for his work on a quantum mechanic oscillator.
So we are very curious to hear your--
Thank you very much.
MALE SPEAKER: Oh, one last thing I should say
is you remote sites, when the talks over, at this time
you guys will be able to ummute, and then you
can ask questions remotely.
Thank you.
JOHN MARTINIS: Thank you very much for the kind invitation
to come here.
I have a son who's a computer science major at UC Berkeley.
And I don't know if you have kids.
When you have kids and they're young,
the parents can do no wrong.
And then they turn into teenagers,
and their esteem of you goes down.
And then, as they get into the real world,
you suddenly become more and more intelligent
for some reason.
So coming to Google, for my son, is totally cool.
Makes me totally cool.
So I'm at a much higher esteem today after doing this.
I want to talk about our project now
to work on superconducting qubits.
And to talk about some recent, kind of amazing results here.
This is maybe one of the first times
we're talking about these results.
The ideas of quantum computing have been around
for 20, 25 years or so.
The idea here is you can do some kind of calculations
maybe much, much more powerfully than you can ever
do with a classical computer, taking
advantage of quantum states.
But it's been 20 years or so.
And you might ask, well, is it really
possible to actually build a quantum computer?
It's maybe a theorist's dream.
Or I've heard one paper call it a physics nightmare
to build a quantum computer.
It's really hard.
We've been going at it for 20 years.
Are we really going to get there?
Is it possible?
And what I want to do is talk today
about some theoretical understandings
in the last few years, and some recent results
in the last year.
Really coming up to data-- I'm going
to show data we've taken in the last few weeks.
Where we really think we can build a fault-tolerant quantum
And we can start down a road to really harvest,
to take advantage of the power of quantum computation.
So I'm going to talk about the theory.
I'm going to talk about our new superconducting qubits.
Basically, here, with the theory for fault-tolerant quantum
computer, you have to make your qubits well,
with an error per step of about 1%.
Then you can start building a quantum computer.
I'm going to show here that, in fact, we've done that.
To motivate this, I want to talk a little bit about D-Wave,
because people at Google and elsewhere
are thinking about that.
And exponential computing power.
And then a little bit more about the need
for fault-tolerant computer computation to do this.
So let's just start with the D-Wave Here's their machine.
Beautiful blue picture here.
They've been very clever in their market
to solve optimization problems, essentially
mapping it to physics of what's called a spin glass.
And one of the big conjectures of the D-Wave machine
is, because they're doing this energy minimization
optimization, mapping it to this physics problem,
maybe you don't have to build a quantum
computer with much coherence at all.
And in fact, their machine has about 10,000 times less
coherence then the kind of devices we're talking here.
So it's a different way of looking at it.
And the nice thing is, once you make that conjecture
and assumption, it's not too hard to go ahead and use
standard Josephson junction fabrication
and build a device to try to do that.
So it's an interesting conjecture.
The machine has superb engineering.
It really is a very, very nice piece of work,
with the low-temperature physics involved in all that.
The problem is, well, although they
think they could be useful, a lot of physicists
are very skeptical of whether it will have exponential computing
And I've been enjoying talking to people here at Google
and other places, because they've said, well,
what does nature have to say in this?
So they've actually taken the machine
and done some experiments.
And I'm just going to review the experiments here.
And this is basically the system size versus the time
that it takes for the D-Wave machine
to anneal to, effectively, the ground state.
You're doing the spin glass problem
with random couplings between the spins.
And they're plotting a typical mean execution time.
And with the D-Wave machine, initially
for small numbers up to maybe 100, it was pretty flat.
But now the latest results, up to 512.
It's starting to grow exponentially.
This exponential growth is actually
matched by some quantum-simulated annealing--
both to stimulated, classical annealing and other methods.
So the preliminary results here, maybe for this particular class
of problems, it's no faster than classical code.
Although people are looking at it.
That's not a firm conclusion yet.
And one has to do more work to see exactly what's going on
in the D-Wave and can you use it.
We're going to take an approach that's
very, very different than this D-Wave machine.
It's the conventional, classical approach where physicists
have proved theoretically-- it's still only theory--
but they have a very strong belief
they should be able to build a computer
with exponential power.
Let me just explain that briefly.
It's easy to understand.
You take a regular computer, and the classical computer
scales linearly with, say, the speed of the processor
or the number of processors.
It's very well understood.
The beauty of CMOS is that the growth of this power
actually goes exponentially in time because of the technology
But it's linearly with, say, speed or processor number.
However, in the quantum computer,
this power grows exponentially.
And the basic way to see this is, in a quantum computer,
it's not just a 0 or a 1 state.
You can put it in a superposition of a 0 and 1
Just like you say that the electron is orbiting
around an atom, and it can be on one side
of the atom or the other.
There's an electron cloud.
At the same time, you can have these quantum bit states
that are both 0 and 1 at the same time.
So here, for example, we take three quantum bits, put it
as a superposition, a 0 and 1.
You write that out.
You have 8 possible states that the initial state can be in.
And you're in a quantum linear superposition
of all those states.
And the idea is you take this one state,
you run it through your quantum computer,
and that's basically taking all the 8 possible input states
and parallel processing them in one operation
through the computer.
So the quantum computer allows you
to do amazing parallel processing here
as 2 to the 3, 8 states, or in general 2 to the n states.
So if you have some quantum computer with 64 bits,
you're processing 2 to the 64 states at once.
To get a doubling in power, what do you do?
Here you would you double the size of it.
Here, to double the power with a quantum computer,
you just add 1 more bit.
And you've just double the parallel processing power.
And by the time you get something
like the 200 quantum bit quantum computer, the parallelization
that you're doing is greater than 2 to the 200,
is greater than the number of atoms in the universe.
So you're clearly doing something there that's amazing.
The problem, however, is that you're
doing this parallel processing.
But you only get, when you measure
the system, end bits of information.
And you have to encode the problem
and only can solve certain problems
to take advantage of that kind of optimization.
So I'm not going to go into that very much here.
But I am going to talk about one application of this.
It turns out the government's interested in this.
And that is factoring a large number
into its component primes.
For example, take the idea of factoring a 2000-bit number.
The algorithms for doing that scales exponentially.
And right now, if you take a 640-bit number,
that takes about 30 CPU years to factor that
into the composite primes.
And then, if you say, OK.
You take this and you exponentially
scale up to some number like 2000-bit,
which is something you might think of doing,
what do you have to do to get there?
So what I've drawn here is I think
this is some Google supercomputer here.
I put this especially for this talk.
What would you have to build to factor a 2000-bit number?
You would have to basically build
a computer farm almost the size of North America.
And you see, I put it up in Canada.
You get natural cooling.
Not too many people there.
I think the polar bears would be happy for that to be there,
because there'd be a lot of people to eat
and that'd be good for them.
But with that size, if you built something that size,
you could do this problem in a 10-year run time.
It's possible with that size.
However, that's maybe 10 to the 6 trillion dollars.
Which, even with quantitative easing,
I don't know if the federal government could even do that.
It takes about 10 to the 5 times the world's power consumption,
and it would consume all of the Earth's energy in one day.
So I know Google, you like to think big.
But I'm going to just conjecture you're not
going to want to do this.
This is not practical.
I get this example because I want to show you just
how reasonable a quantum computer might look.
And we don't quite know how to build that now.
We have a general idea.
We need about 200 million physical qubits.
100,000 let's call logical qubits.
You could probably put this in some building size.
Maybe even fit in this particular lecture room,
with a bunch of refrigerators and control electronics.
Maybe a small supercomputer.
A 24-hour run time.
I don't know what these facts are,
but it'd probably be the cost of a satellite or two
and certainly not consume that much power.
So it's something you could imagine possibly
doing, if you understood all the technology on how
to build this.
The basis of how to build this and the hardware
is what I want to talk about today.
So if you're building-- it's really great.
You have this potential exponential scaling,
exponential power to the quantum computer.
But the problem is that the qubit states are really, really
And it turns out that you have more power in fragility.
But you have to build this in the right way
to take advantage of it.
So I'm going to give an example here.
Just trying to understand qubit errors.
Take, for example, a coin.
We're talking about classical bit.
We're going to talk about a coin on a table.
This is a stable piece of classical information.
Why is that?
If I jiggle my hand, some air is going on there.
It stays in either the head or the tail state.
It stays as 0 or 1.
If I jiggle it hard enough, you can
imagine the tip of the coin lifting up a little bit.
But if it does so, the restoring for its dissipation
is going to push it down again.
And it'll stay in one state or the other.
And this is the basic idea of classical bits.
Is you can make them stable.
And they can be extremely stable,
and you don't have to worry about them having that error.
And if you do have errors, you can take care of it.
But they fundamentally can be made stable.
A quantum bit, in analogy, is not
stable like the classical bit.
So just using the coin analogy, you could say this is 0
and this is 1.
But 0 plus 1 is maybe the coin standing up on edge.
And in fact, with different phases it's going to have,
this coin can turn around.
You're going to have different angles.
You can have a wide variety of states here.
I think the right analogy to think
about that is a coin in space, where
there's no table holding it down to one state.
And you could set, initially, that coin with some angle
which would be some quantum state.
But you could see that any small perturbation, any small force--
a puff of air, whatever-- is going to start rotating
that coin and then giving you an error.
It's just fundamentally different situation
when you don't have this self-correcting mechanism
that you do with a classical bit.
So that's the problem.
Actually, when you go through the quantum physics,
it's really fundamental that you have this kind of problem.
And the idea is you can write a wave
function that's an amplitude.
How much 0 and 1 you have.
And also, there's a phase associated with, say, the one
state, which is like the coin turning around
in this direction.
And these two variables, amplitude and phase,
you have to worry about.
And you have to think about, will measurement fluctuations
cause these amplitude and phase to flip in some way?
Now, quantum mechanics says that there's
this thing called operators for the amplitude and phase.
This is a flip operator, which flips a 0 to 1 and 1 to 0.
And a phase, which changes the phase of the wave
function from plus 1 and minus 1.
And these particular operations, we say they do not commute.
And it's basically saying, if we do an amplitude flip and then
do a phase flip, that's not the same as doing a phase flip
and then doing an amplitude flip.
And these two operations, the order matters.
That's like saying that when you have
an electron along the atom, the position and momentum don't
And if you try to measure position,
you would affect the momentum.
Things like that maybe you've heard in some basic physics
This happens with both this amplitude and phase
information, that they do not commute.
And in fact, if you look at it carefully--
and I hope people will go away and do this with your hand,
do an amplitude flip and then a phase flip, or a phase flip
and then an amplitude flip-- you're
going to say, hey, wait a second.
Those are doing the same thing.
Classically, they do the same thing.
But quantum mechanically, those two operations
are different, because there's a minus sign involved in that.
Now, you don't normally see that minus sign,
because the end probability of doing something
to quantum mechanics squares that minus
sign so it looks like the same thing.
But quantum mechanically, if you build a quantum computer,
these are fundamentally two different states,
and you would see that effect.
And this is talked about.
The minus sign is that the sum of these two operations
and 0, instead of the regular commutation [INAUDIBLE].
So this is a little bit mathematical.
But I wanted to bring up that mathematics to show you
how this problem is solved in quantum computation.
It's very simple and you can understand it.
It's very much like in error correction classically.
And what you do here is you now can
set a 1 bit, which doesn't work in this way.
You now take 2 bits.
And now you make a parity kind of measurement between 2 bits.
So there's an amplitude.
We call it a bit flip parity.
X1 and X2.
And then there's a phase flip parity.
So it's like having two coins.
And then we can flip both of them
or we can phase flip both of them at the same time.
Let's just do some math.
We're going to look at this commutation relation, which
describes the essential physics.
You now see that you have these pairs of these.
And I'm going to flip these around with a minus sign.
And then we use this amazing mathematical relationship--
minus squared is equal to one.
You see, there's a minus here and a minus here.
And that means this thing is equal to this.
And that's the commutation is 0.
So even though a single qubit has
this strange quantum mechanical behavior, when
you look at the relationship for 1 qubits,
they obey classically, both in amplitude and phase.
And thus, you build error detection protocols
that are based that you can do these essential classical
measurements on 1 bits on parity measurements.
So let's say we take 2 bits.
Because of these, we measure this phase parity.
It's plus 1.
And then we do an amplitude parity.
It's minus 1.
What does commutation relation equal to 0 mean?
Is I can continue to measure this over and over again.
I'm going to get a plus 1 for the blue and minus 1
for the red.
And it's stable.
And in fact, one measurement doesn't affect the other.
So now what we can do is take these two coins, if you like.
And we measure it in this way.
And then, if it just stays plus or minus 1, it never changes,
we know everything's OK.
However, if one of them changes, let's say plus 1 to minus 1,
then we know we had an error.
We can measure that.
Now of course, you're going to say, well,
how do you know which qubit was an error?
And it's very easy once you know about error correction.
What we can do is have 3 qubits right here.
We do the Z 1 2 measurement between here and here,
and the Z 2 3 between here and here.
And then, if this one was an error, then
these guys are going to change.
If this one flipped, this one only changed.
If this flips, both of these change.
And if this flips, this one here will change and this one won't.
So you see, by having the two measurements and 3 qubits,
we can figure out which one changed.
So you can identify the qubit errors.
So you can see that you can scale this up and make
it more and more accurate.
The problem here is if you have two qubits at the same time
that made an error, you can't detect it.
But if you make it longer and longer,
it's just like classical error correction.
You can fix that.
So now I want to talk a little bit about what the full quantum
computer would look like.
You basically take this idea and scale it up some.
You make a huge array of qubits.
We call these open circles the data qubits.
And then the closed circles here our measurement qubits.
And the measurement qubits are measuring
the four qubits around them.
And here's some circuit that does this.
This is basically the quantum version
of a CNOT or something called an XOR.
And this circuit basically measures the parity
of these four things here.
And same thing with this.
It measures what's called a phase parity in the normal way
that you would think about parity.
And then you just repeat this over and over again.
Repeat these measurements.
That's all that the surface code does.
So how does it work?
You have to realize, for these measurements here,
measuring these 4 here and measuring 4 qubits here are not
going to affect each other so that they're separate qubits.
The only time you have to worry about them affecting each other
is these qubits here and these qubits here.
But notice that there's a pair of qubits this
that identified with this and this.
And that means, because of that pairing
and because minus squared is 1, these two measurements commute
with each other, and you could simultaneously
know the answer here and the answer here.
And if you run the surface code, you'll
get a bunch of these measurement outcomes
that will be constant over time unless there's an error.
If there's error, you're going to get a bit flit somewhere.
And then you're going to measure that.
So for example, you might be running.
All these measurements are the same.
And then, at some point in time, you'll see that this plus 1
turns to a minus 1, and this minus 1 here turns to a plus 1.
So you'll get a pair of errors.
This error here says one of these 4 qubits flipped.
This error here says one of these 4 qubits flipped.
And you naturally say, OK, it was
this qubit that was in error.
And you can do the same thing down here.
This error here says 1 of these 4.
This said 1 of 3.
So you identify an error there.
You can do the same thing in case
there's a measurement error.
Instead of two pairs in space, it'll be two pairs in time.
So that's what you do, is you just run the surface code.
No errors, all these numbers come up at the same time.
Same thing every time.
If you see errors in that, you can figure out
what thing had the error.
Of course, the problem is, if you run this,
every once in a while you'll get a bunch of errors at one time.
And then the question is, can I back out
what really happened in the surface code?
Most the time you can.
But the errors come out when you can't figure that out.
And that's when it breaks down.
And I'll talk about that a little bit more mathematically
in a second.
So I've talked about how to pull out the errors.
But actually, how do you store information in this?
And it's actually stored in a very similar way
that you would see with classical codes, in that we see
we store the information in a parity way across all the bits.
So let's just look at this for a second.
We have 41 circles, which are the data qubits.
And 40 measurements, which are the closed circles.
And you might think that if there's one more data qubit
than measurements, you would think that there's
an extra degree of freedom to store the quantum state.
In fact that's true.
And the quantum state, in this case,
is stored by a string of data qubits going across this array.
And the bit part is stored in this way.
And the phase part is stored in this way.
And these particular, they're called operators,
that describe the state, they anti-commute with each other.
So they act like a qubit.
And all these commute with all the measurements,
so they're stabilized in the normal way.
I won't get into this in too much detail.
But you can make something look like a qubit because of that.
Just building a bigger and bigger space.
How big do you need to make it to make this accurate?
Well, that's done with some simulations that
look into the logical error rates.
And what we do there is we take the basic surface code cycle,
and then you put in a probability
to have some kind of quantum error
in each step of the surface code cycle.
And then you run the surface code cycle
when you have some algorithm, minimum weight matching
algorithm, that says, if we measure some errors,
what was the actual logical error?
And if it matches the errors that came up into here,
we say it was error corrected properly.
And then every once in a while, you
see that the logical error is not corrected.
And that will be a logical error.
And what this is is the logical error versus the error
probability P, put per step.
And you see basically, as the error probability goes down,
then the logical errors go down, as you would expect.
But then, as you make the array size bigger and bigger,
then the logical error rate goes down faster and faster.
As long as you're below some number of around 1%
in the error probability.
And this is called the threshold of about 1% error.
And as long as you're below that and you
have a big enough dimension, big enough surface code, then
the error will get exponentially small.
And that's how you store.
You can store a qubit state for a very long time without error.
You just make it good enough and make it big enough.
No different than classical error correction.
Just a little bit more complicated because
of the quantum physics here.
But the concepts are the same.
Now, it turns out you can understand
this behavior in a simple way.
This is high school statistics.
These kinds of concepts you use all the time
in classical computing.
Let's just take one row here of a surface code array
and say, at some point in time, I
had an error in measurement here and here and here.
And when you see this, you say, look.
If I have an error here and here,
that means you've got a data qubit error here.
There's an error here but not an error here.
So I'm going to associate this error
with the qubit at the end.
And this is a correct association
of a data qubit from here to here.
But it turns out that that backing out of the real error
is not unique.
You can also take the complement,
and the complement also solves this.
And your question is, of course, which one you take.
Well, obviously this has 2 errors.
It's going to go as this P squared.
This has 3 errors.
P cubed.
This is more likely than this.
So you're going to choose this and be right most the time.
But every so often, with probability P cubed,
you're going to get a logical error given by this.
And you can work out, this is high school statistics.
And then write down a formula for this.
And you see that this very simple description of this, it
fairly well matches this.
There's some subtlety that it doesn't pick up.
But you basically get the idea.
So that's how error correction works.
And it just means you need to have
small errors and a big enough size.
And this is just taking the formula we got here
and I say, let's hold our qubit state with a logical error
rate of 10 to minus 5, which is 1 second time.
10 to minus 10, a day.
And 10 to minus 20.
A little bit more is the lifetime of the universe.
And you see that if you can be at a 0.1% error
here and make a few thousand qubits,
you can hold a qubit state, this fragile quantum state,
for the lifetime of the universe.
That's cool.
And of course, that's kind of what you'd have to do.
If you have 100 million qubits doing some algorithm.
You need to have some kind of small, logical error rate
to run an algorithm properly.
But you can actually approach lifetimes estates
with this idea.
Like what you get for classical bits playing this game.
But it takes a lot of resources.
That's just what physics requires you.
I've talked about memory.
You need to do logical operations on it.
What's really beautiful about the surface code
is you just build this big code, and then you
can make additional qubits by essentially
what's called putting holes in it.
In The middle of this, where you turn off the surface
code measurement.
And then you have a bunch of states
that can then generate the qubit state.
And then you can do operations.
The most interesting is by taking one of these holes
and moving the hole around another one,
you then produce a logical CNOT or XOR operation.
You can do other things.
So basically, with this basic surface code,
you can build up and do logical operations
and do quantum computation without error,
if it's big enough.
So what I want to do now is I want
to talk about how we're going to implement this.
And we're using superconducting qubits.
You could think of these as atomic systems,
like an electron around a nucleus.
But in this case, we're building electrical circuits
where the quantum mechanical variables
are current and voltage.
So you have a wire and you have the current flowing
to the right and the current flowing to the left
at the same time with some quantum mechanical wave
function, just like an electron can
be on one side and the other side of the atom
at the same time.
So can current and voltages.
It's possible to do that.
These circuits typically work in the microwave range,
5 gigahertz.
And they have the energy of these systems, which is HF.
To be greater than KT, we need to operate them
in 20 millikelvin ranges.
And that's not hard at all with something called a dilution
This is well-established technology.
Now, what happens is we can build these various qubit
If you, for example, take an inductor and capacitor--
or in this case, we have a transmission line
of a certain length, which has resonant modes that
look like piano string resonant modes.
This looks like a harmonic oscillator.
If you look at the quantum mechanics of that,
they have equally spaced energy levels.
And you would say, oh, let's just take the two lowest energy
levels and make that a qubit state.
And that's essentially what we do in our system.
The problem is that, for this linear harmonic oscillator,
these two energy levels are the same.
So you drive this, you drive this.
You drive this transition.
You drive this transition.
And the state just wanders all the way up and down here,
with many quantum states.
However, you can use a Josephson junction, which is basically
two metals separated by a very thin insulating barrier
so that electrons can tunnel through that barrier.
Then you get a non-linear inductance
from this particular quantum inductance device.
You then turn this quadratic potential
into what looks like a cosine potential.
This is now a non-linear potential.
So that when you look at the energy levels,
they are not equally spaced.
And now, when you drive this transition,
this is off-resonance, and then nothing happens there.
You stay within your qubit states.
And then you can build a quantum bit out of it.
So this is how we make them.
We build integrated circuits.
Right now, it's aluminum metal for the metal and the Josephson
What pink is in here is basically aluminum.
It's on a very low-loss sapphire substrate.
We just used standard IC fabrication technology.
There's quite detailed material issues you have to deal with,
which we've been working on for 10 years and 50 researchers,
just in my group.
There's a lot of other people working on this, too.
But nowadays, we know how to make it
so that these are really very well-made.
These little X straights here of structures
here are these called Xmon qubits bits.
They're capacitively coupled to each other.
The wires to control them are coming in from the bottom.
And then these wires here come from the top.
And then we can read out the qubit state
by putting microwave signals through this here.
And I'll explain how that works.
But the truly standard IC fabrication, kind of amazing.
You just have to choose the right materials
and make it in a particular way.
And this X1 qubit, we basically have a ground plane
to the outside.
And that just forms a capacitor in this X.
We have this Josephson junction that
forms an L. That non-linear LC resonance forms the qubit.
And then we have a loop here with a line coming in here,
and we can change the inductance.
We can change the frequency of the qubit.
We can also put microwaves in here, capacitively coupled.
Those microwaves electrically force current into the Xmon
and cause it to make transitions from the ground state
to the first excited state.
So by put it in microwaves, putting it
in a change in frequency, we can completely control the qubit.
This is a picture, a graduate student lying on the ground
as he's putting it together in the dilution refrigerator.
These chips go inside this aluminum box.
And then coming out of it are coax wires
through some filters and other structures.
And then we have a lot of coax that
goes from here to the top of the cryostat at room temperature,
and then through the electronics over here.
And this is when it's open, you put a bunch of infrared shields
and a vacuum jacket around this and cool it down
with liquid helium.
And you can get to 20 millikelvin,
so that you get rid of all the electrical noise in the system.
And then it's just all controlled with all
these microwave electronics here, a lot of test equipment.
But everything is controlled over here by computer
so that it's easy to set up the experiment and get it to work.
So this is just some simple way to think about the qubits.
The first one we called a Rabi oscillation.
In this particular case, we take our coin
and we have it in the ground state.
And then, with microwaves, we flip the coin,
we rotate the coin at a steady rate that's
proportional to the microwave amplitude.
At a certain time, we stop the rotation
and then measure whether it's 0 or 1 state.
Of course, that's probablistic.
If it's going on edge, half the time it'll be heads
and half the time it'll be tails.
But you can do the experiment many times
to get a probability.
And what you see here is you just rotate longer and longer.
You're just flipping from heads to tails, 0 to 1, up and down.
And you see that the magnitude of the oscillation
doesn't decrease in time, because we
have very good coherence of the system.
So the typical time scale that we can flip the system
is maybe 10, 20 nanoseconds.
And then the tip of the lifetime of the system, which
is given here, where we go from 0 to 1,
and then we measure if it's in the 1 state versus time,
it eventually decays and relaxes to the 0 state.
But that does that in, say, 30 microseconds.
And the ratio between this and this is a factor of 1,000.
So we should be getting roughly a 0.1% error per gate.
And that would be, in principle, good enough
to do this error-corrected quantum computer.
But that's, of course, only in principle.
Actually, how do you make the gates?
So I want briefly to talk about the gates and what we do.
And I want to show you that we can make very complex gates.
And this system works extremely well.
And what we have here is something called
randomized benchmarking, where we're
putting in a very long sequence of gates into the system
and seeing if we're controlling the state.
Now, in this particular case, with randomized benchmarking,
we're going from 0 to 1 or from 0 to 0 plus 1, or 4 phases.
So this is going 6 equally spaced points
on this, what's called the block sphere.
So it's reduced-static quantum states.
But the nice thing about going to these particular set
of states and rotating or gating them into those states
is this forms a gate set that you can calculate very easily
just with classical computation.
And forms a generic base that you
can calculate very carefully and know what's going to happen.
So what we do here is we just take
a bunch of these different rotations
to take the state all around the block sphere,
over and over again.
And at the end, we know where it should be.
And then we rotate it back to pointing this way in the 0.
And we see if it's in the 0 state or not.
And then we do that complicated sequence of pulses
as shown here.
We then do it for other kind of gates
that move it in a different sequence.
And then average all that and say, OK,
do we get into the ground state?
And we see, of course, that it's not in the ground state
But then for you to have an error-- that's here,
and this is 0.1 size.
So this is not a huge error.
We can make hundreds and hundreds of gates
here in arbitrary combination, and we more or less
get this right answer here.
And you can work out the statistics.
And this says that the fidelity of these operations are 99.93%.
So only one gate in 1,000 is going
to give you a significant error.
And in fact, you can understand this a little bit more.
You can interleave these with specific gates
here, and very much quantify what's going on here.
But the end result here is we can make these quantum gates
well beyond the 99% that we need to do the surface
code and the error-free correction.
That's 1 qubit.
We have to run 2 qubits at the same time
to do some parallel processing.
We take those two qubits, set them at different frequencies.
Even though they're coupling here capacitively,
when you put it at 2 frequencies,
it effectively turns off the interaction.
You run your Clifford gates.
99.9495 individually.
You then run them at the same time.
Because they're detuned, there's basically
no degradation in gate fidelity.
This number's smaller because you're
adding the errors of this and this in the way we do it.
So there's negligible crosstalk.
We should be able to operate these things in parallel.
We can also need to couple of them together.
We have to make this CNOT kind of gate
that I was talking about.
This, in fact, is the hard thing to do.
And this is what people have been trying to do for 20 years,
to get this gate good.
This is the hard gate.
And we think we've cracked this.
Conventional thinking-- you operate these qubits
in a very stable configuration so
that it's not frequency tunable.
It's like an atomic clock.
It gives the longest memory.
Then you connect them through some kind of quantum bus,
where that qubit connects to some resonator cavity,
connects to something else.
That gives you long-distance communication.
You then do some complex microwave or photon drive
to get all these things to interact and get it
to work [INAUDIBLE].
It's very complex And you get it to work.
Ion traps, for example, are at about 99%.
Superconducting qubits, when they do that,
these are slow gates.
10 times slower than what I've been talking about.
Fidelity's not so great.
What we've done here is a totally different design.
We've taken all the conventional theory, the thinking,
and turned it on its head.
We use an adjustable frequency qubit.
And that's actually good, because we
can move them in and out of resonance
and turn on and off the interaction.
We have direct qubit coupling, no intermediate quantum
bus that can give us de-coherence.
And then, instead of driving it with microwaves or photons
we just change it with the DC pulse to change the frequency.
You need to do that accurately, but it can be done.
Theory says this should be really good, acceptable.
Experimentally, we do this.
These are some tuneup procedures.
It's for a Controlled-Z that's equivalent to the CNOT.
We can get this pi phase shift, this minus 1 side.
That's shown here.
This shows with full quantum states,
it's acting in a way it should be.
I'm running out of time, so I'm going to go over this quickly.
But basically, things are working right.
You do randomized benchmarking.
These are the Conrolled-Z gates.
We get a fast gate.
That's a very accurate 99.45, as shown here.
And sorry I can't go into this much.
This is best in the world.
It is better than ion traps.
Better than other qubits.
We know how to improve it.
This basic idea works very well.
Let me talk about qubit measurements.
I'll be done in four or five slides or so.
You have to measure the qubit.
What we have here is this qubit, and then it's capacitively
coupled to a microwave resonator right here.
And then that is also capacitively coupled
to another circuit right here.
So these being capacitively coupled,
it turns out that there's no energy
exchange between the qubit and here.
But the frequency of this particular resonator
changes depending on whether this is a 0 or a 1 state.
So what we do is we put a microwave signal here
that's resonant with this frequency that couples to that.
And because this frequency changes because of this
being the 0 and 1 state, that will
introduce a delay in this microwave
depending on whether it's a 0 or 1 state.
You then measure that with a quantum limited pre-amplifier
and room temperature analog-to-digital converter,
an FPGA that can measure the phase shift.
And you can tell what's going on in the system.
So here's just more details of that.
Here's the drive signal.
You put about 100 photons into that one resonator,
that has a frequency shift.
Here is plotted the real and imaginary part of the signal
that you're measuring here.
If you're in the 0 state, you have the phase,
so it's over here.
If you're in the 1 state, the phase is over here.
And integrating over about 100 nanoseconds,
you see these two signals are super well-separated.
And then you just say, if it's on this side,
it's a 0, and this side's a 1.
These are plots that are basically
showing what the separation error is.
Because these have Gaussian tails,
there are small errors between this.
But it basically says, in a few times
the operation of our single- or two-qubit operations,
we can see separation errors that are 10 to minus 2 to 10
to minus 3.
So we can measure the states extremely accurately.
Finally, we need to measure more than one qubit.
We talked about this one here.
We also have another qubit here with another resonator.
These are at two different frequencies.
So you put in two tones here.
This tone here gets shifted depending on the state.
This tone here gets shifted depending on this state.
You amplify that all.
The FPGA can separate out these two frequencies.
Get the amplitude and phase.
And then tell whether it's a 0 or 1 state.
So this is just data coming from-- this
is the readout signal of one qubit versus the other.
If we put a 0, 0 in here, this ends up here.
If it's 0, 1, it ends up here.
1, 0 here.
0, 1 here with the other states.
These states are all separated very nicely from each other.
So you can accurately measure multiple qubits
in a very short amount of time.
So we know how to scale that up.
And again, this is above the threshold.
Everything works well.
Last thing.
This is maybe people here will understand.
When you're building these complex systems,
you have to abstract away the functions.
You have a lot of complicated things going on here.
In our system, we can scale with all this stuff
with good control using software distraction, which
includes calibration of the hardware and waveform
and non-idealities.
Specific qubit calibrations.
So you basically calibrate the whole system.
And that takes, maybe, program 100,000 lines of code.
You understand that.
And then once you do all that, if you
want to do some complicated algorithm here, it's, what?
7 lines of code.
You just say, I want to do these particular gates.
And all the calibrations are done for you.
You just put in the gates, run it, you're done.
So at this point, running the programs
are really essentially trivial, as it's all just calibrating it
The amazing thing is that we can calibrate this up
and we run it, and it runs super well.
It runs with the errors that I showed you.
So it is possible to build this hardware system to abstract it
away as you would need to do.
So I think my 50 minutes is up.
I want to summarize and talk about the outlook.
People have been wanting to build a quantum
computer, a fault-tolerant quantum computer that
would potentially, eventually give you
enough exponential power.
We've been looking at this for 20
years in the experimental realm.
We think that our particular technology is now good enough
to do fault-tolerant computation.
This would be very hard to scale up.
We have a lot of technical challenges.
But the basic ingredients to do this are there.
It's at least good enough that we really
have to start doing this seriously.
No more playing around, writing physics papers-- although
were going to do that, too.
It's time to get serious and build this quantum computer.
The surface codes needs 99% fidelity.
We have 99.3, 99.5.
Measurement's good enough.
We think this is scalable.
Improvements are likely here so we can do well.
So the numbers are there.
It's time to get started.
What I'm looking at, based on what I've talked to here,
I would like to start what I think
is roughly a five-year project.
Although we can have problems.
It may take a little bit longer.
But I think we understand the basic technology.
And is basically to scale up to 100s,
maybe 1,000s of qubits using the surface code architecture.
And then try to do one with a logical error rate 10
to minus 15.
Hold the qubit state.
These incredible fragile quantum states and hold it
for 100 years or 1,000 years.
A really long time.
Showing that it would be OK.
And then this would be big enough so
that you can start doing these [INAUDIBLE] operation
or whatever to do logic operations at 10
to minus 6 errors.
I think this particular science project
is what's needed right now to show that all these ideas are
correct in a way that we understand
that the power is there.
And then, if this works, you would then go ahead
and, if you've got all the technology right,
you would try to build something that
was useful and could do something.
But we really want to focus on getting the science right
and understanding it in the next five years
and we really think that's doable.
Not just me.
All the graduate students and post-docs in my lab.
They're doing the work.
They really think this is possible along with me.
We look at the technology.
It really looks doable.
It looks like something we should be working hard on.
So let me end right there.
Here's our group at UC Santa Barbara.
It really takes a lot of people working together to do that
and we have a larger collaboration
of about 50 people with theorists
another experimentalists to get this done.
It's quite a lot of work, really takes a lot of teamwork.
But we think the technology's there.
So thank you very much.
MALE SPEAKER: Thank you, John.
It was a very nice talk.
I appreciate that you did it nicely in time,
so that leaves time for some questions.
I was wondering if you could--
MALE SPEAKER: Could everybody use a microphone
so that people on the remote sites can hear it as well?
I was wondering if you could compare your surface code
architecture with, for example, the toric code?
What are the advantages and disadvantages?
The surface code architecture has the highest threshold
that we know of.
And that's incredibly important, because it's
hard to make good qubits.
We've been struggling with that.
Typically, initially, people talked about codes
with you needed 99.99% fidelities
to get it to threshold.
That, to me, looks really hard.
But at two 9s, that's something we can do.
The other nice thing about the surface code
is it only requires nearest neighbor interactions.
And if you're building that on the integrated circuit,
that's great.
So I think those two things are really
the key advantages of a surface code.
But people are looking at different codes
and different things.
And if something gets better, we can do that.
But surface code looks really quite ideal
for building integrated circuits.
AUDIENCE: Thank you.
JOHN MARTINIS: There's a question there.
MALE SPEAKER: Pass the microphone over there.
AUDIENCE: Can you discuss how far along you
are towards a surface code architecture,
and what is it going to take to get from 2 to 41?
How far along?
So let's just look at the surface code.
Come on.
Slow computer.
You have to make a big array, OK?
Here, this is a couple hundred qubits.
There are some simple versions of the simple surface
code we can do at 5 or 9 qubits to test
if it's working properly.
And we're starting to design the chip.
And we hope to have some error detection, whatever,
working in about three to six months.
No one else is even thinking about doing that.
We think we can make quite rapid progress.
We really want to show that this simple surface
code is working right.
And then, at that point, I think people
will get on board that this is possible.
Everything is working great, so we really
think in three to six months we may have that.
And then we have to figure out how to make lots of qubits.
But we have some ideas.
We really want to demonstrate a simple version of that code.
MALE SPEAKER: I'm kind of scared to step
in front of the loudspeaker.
But connecting to this, I actually had one question.
You mentioned scaling it up would be really hard.
Can you list, a little bit, the main challenges?
JOHN MARTINIS: We know how to build, more or less,
the integrated circuit, and we know the materials.
But when you build something like this,
you have to get control lines in to all of those qubits.
Now, if you're talking about atoms
that are microns apart or less, it's
hard to get those control lines in.
But here, they're separated by hundreds of microns.
And we can IC fabricate control lines to get into that.
So we think we know how to do that.
We have an idea on how to do the processing and all that.
And then we have to bring 100 or 1,000 control lines
to the outside of a wafer, then wire-bond that up
to electronics at room temperature.
You just have to think like a high-energy physicist.
You just build a lot of wires and do all that.
We think we can do that.
From technology we have, or maybe
we just have to modestly invent something.
But that's the basic idea.
Just bring out those control wires
to the outside of the chip.
Wire-bond them.
All these cables going up to racks of electronics.
And for doing the scientific demonstration,
we think we can do that.
Eventually, if you want to go beyond the thousand qubits,
you have to put the control either right down in the chip.
And that there is the technology of classical Josephson junction
computing, which people have been working on
for years and years.
And we actually have a program to start
trying to figure out how to do that.
So as we're building up this brute-force way,
at the same time, we wanted to be developing
the classical control circuitry to do that.
Going back to D-Wave, one of the impressive things
D-Wave has done is they built that classical control.
It's not exactly what we want.
But when I look at what they invented,
it gives me a lot of hope that we can figure that out.
Because that's both a combination
of analog and digital.
We have to do the research.
But I'm optimistic that that can all be done.
It's just hard.
But OK.
This is what you have to do.
And in fact, the hard part of building a quantum computer--
making good qubits, DiVincenzo criteria-- yeah,
it's really hard to get 99.45% fidelity.
The hard part is the control circuitry.
You have millions of qubits.
How do you get all that control within each qubit?
Because it's basically analog control.
I think you can do it here.
But that's going to be a super challenge.
Again, to do some physics, we don't have to crack that yet.
MALE SPEAKER: Another immediate thought.
That if you could borrow some of the control electronics from
D-Wave and apply it here--
JOHN MARTINIS: Their control electronics
is a different mode than this.
But there could be a lot of commonality.
And for me, it's more that they've
shown that you can mix digital and analog, in their way.
And you might want to borrow some of the ideas
or be inspired by those ideas to do it.
But I really feel that, given people working hard on that,
we can crack that problem.
But it's something eventually we do.
However, if we want to show the science works well,
and to have a fragile qubit state
and hold it for 100 years, I think you can brute force that.
Which is one path we want to take.
And then, at the same time, work on the other things.
That's my view of how things should go.
MALE SPEAKER: There was one more question earlier,
but I think we--
AUDIENCE: I have a question.
AUDIENCE: So how small can you make this, practically,
if you wanted to have-- and you show a homogeneous matrix here.
But if you wanted to have a bunch of matrices,
maybe with some space between them for control circuitry.
This is 10 by 10, the minimum size.
JOHN MARTINIS: So that's what I'm talking about here.
Right now, we're thinking the cell size
is going to be eventually between 100
microns to a millimeter on a size.
And remember, it can't be too small,
because you have to pack all that control circuitry in it.
So at 100 microns in a millimeter,
you can put a significant amount of control circuitry.
And then, if you do that, say 100 microns,
it's maybe meters across in this direction.
It'd have to be a big thing.
But those are the numbers.
Everyone thinks, from modern microelectronics,
that you have to make everything small.
But as soon as you do that, you have
to make your control circuitry that small.
And the control circuitry is not two transistors or something.
It's complicated.
So that's why you need it kind of big.
But these numbers, I think, you can imagine, given enough time,
you can solve these problems.
They're not easy problems.
But I think it's possible.
AUDIENCE: About, basically, the oral architecture
of the computer.
So suppose you placed 100 qubits on the chip and all the control
Does that mean that you already have a 100-qubit computer?
So is this a device for practical computations?
Or, basically, the difference between physical and logical
qubits here.
What is that?
JOHN MARTINIS: It depends if you're
worried about error correction in your algorithm.
And that's a question we're talking
about today, as we did here.
I'm talking about building an error-corrected device.
So if you build 1,000 qubits, your error rate
is going to be 100 years.
But then you could start making smaller qubits in it,
where their error rate may be one per second.
But then you could do logical operations
with those qubits and test things.
So I'm not sure if you could do anything practical
at that point.
But you can certainly test the science.
And that's what I'm thinking right now.
Like with the D-Wave, the question is the science of it.
So if we were to test out the science
and make sure that everything was OK,
that would give us a lot of confidence
that we can move forward in doing it.
Because there's a lot of theoretical assumptions
here that we have to deal with.
But you might be able to use such an array
without error-corrected mode in some interesting, useful way.
And then we would, of course, do that,
if someone had a good plan.
But the error correction forces you into an architecture.
But once we have the technology, we
can do other things, for sure.
For example, part of our group is
looking at quantum simulation for a physics problem.
And we're thinking we can do some interesting things there
MALE SPEAKER: Maybe just to quickly check
whether any of the remote sites may have a question?
There don't seem to be.
AUDIENCE: Sorry, ask the question again?
MALE SPEAKER: I was wondering if the remote sites, was there
any questions from there?
Any last question from here?
Thanks one more time--
JOHN MARTINIS: Thank you very much.
MALE SPEAKER: --for the very interesting talk.
And very upbeat information.
Thank you.
    You must  Log in  to get the function.
Tip: Click on the article or the word in the subtitle to get translation quickly!


Tech Talk: John Martinis, "Design of a Superconducting Quantum Computer"

5448 Folder Collection
Daniel Yang published on November 21, 2014
More Recommended Videos
  1. 1. Search word

    Select word on the caption to look it up in the dictionary!

  2. 2. Repeat single sentence

    Repeat the same sentence to enhance listening ability

  3. 3. Shortcut


  4. 4. Close caption

    Close the English caption

  5. 5. Embed

    Embed the video to your blog

  6. 6. Unfold

    Hide right panel

  1. Listening Quiz

    Listening Quiz!

  1. Click to open your notebook

  1. UrbanDictionary 俚語字典整合查詢。一般字典查詢不到你滿意的解譯,不妨使用「俚語字典」,或許會讓你有滿意的答案喔