Faruk Sami BilginNotes on quantum devices
A long read · Notes on quantum devices

The machine around a qubit

A transmon is a few square millimetres of patterned metal. Keeping it quantum takes a refrigerator taller than a person, dozens of cables that must not let heat in, and a small copper box. This is a tour of that machine, from room temperature down to the chip and back.

Faruk Sami Bilgin · November 2026

In the article on the Josephson junction I described the transmon, a superconducting circuit whose two lowest energy levels make a qubit. Its transition frequency is a few gigahertz, and that one number decides almost everything about the machine it has to live in. A microwave photon at \(f = 5\) GHz carries an energy \(hf \approx 3.3\times10^{-24}\) J, about 21 µeV. Thermal energies come in units of \(k_BT\), and the two are equal at

$$T = \frac{hf}{k_B} \approx 0.24\ \text{K}.$$

Well below that temperature the qubit relaxes to its ground state and stays there. Well above it, its surroundings keep knocking it into the excited state. In thermal equilibrium the ratio of the populations of the excited and the ground state is the Boltzmann factor

$$\frac{P_1}{P_0} = e^{-hf/k_BT}.$$

At 10 mK the exponent is 24 and \(e^{-24} \approx 4\times10^{-11}\): for every practical purpose the qubit is in its ground state. At 100 mK the ratio is already 9 per cent, which would be an error waiting in almost every operation. At 4 K, the temperature of liquid helium, it is 0.94; the qubit is then close to an even, random mixture of its two states and holds no information at all.

So a superconducting quantum processor is kept at about ten millikelvin, some 270 times colder than the microwave background that fills the universe. The machine that does this is a dilution refrigerator, and it is the object in nearly every photograph of a quantum computer: the gold chandelier. In this article I'll take it apart, from the plate at room temperature down to the copper box that holds the chip, and follow a microwave pulse all the way down and back up again.

Here is the whole machine. It hangs from a stainless-steel plate at room temperature. Below it, held by thin rods, are five plates, each colder than the one above: about 50 kelvin, about 4 kelvin, the still at under a kelvin, the cold plate at about a tenth of a kelvin, and finally the mixing-chamber plate at about ten millikelvin, under which the sample hangs. The upper stages are each wrapped in a radiation shield, and the whole stack sits inside a vacuum can. Drag the figure to turn it. The first slider slides the shields away, the second pulls the stages apart, and the third takes you down to the sample. The button sends a microwave pulse to the qubit and brings its answer back.

—
—
—
Pulse power— Where—
Figure 1. A dilution refrigerator modelled on the proportions of commercial units: the real stack is about 1.1 m from the top plate to the mixing-chamber plate, with plates some 40–70 cm across, and the model is drawn to that scale. Wiring is simplified to a few dozen lines, where real systems carry tens to hundreds. Connectors, attenuators and amplifiers are drawn at about their real size; the chip inside the package is shown at its real size but its features are enlarged. Temperatures are typical values; each manufacturer and each system differs. The pulse budget uses typical attenuations and cable losses, listed in the text.

Start at the top. The cylinder standing on the top plate is the head of a pulse-tube cooler. A compressor in a neighbouring room pushes helium gas back and forth through it roughly once a second, and the expansion of the gas cools two stages: the first to roughly 40–60 K, the second to about 3–4 K. Nothing moves in the cold part of a pulse tube, which keeps vibrations low. Its two stages are bolted through copper flanges to the 50 K and 4 K plates. A typical cooler delivers some tens of watts at the first stage and a watt or two at the second.

Everything colder than 4 K is the work of the dilution unit in the middle of the stack, which the next section takes apart: the still, the heat exchangers below it, and the mixing chamber on the bottom plate, the coldest place in the machine. Each plate has a cooling budget, and the budget shrinks at every step: about a watt at 4 K, milliwatts at the still, and only microwatts at the mixing chamber. Everything about the construction follows from that. The plates hang on thin-walled rods of stainless steel, a poor conductor of heat. The shields are there because a surface at 300 K radiates about 46 mW per square centimetre, while one at 50 K radiates about 0.035 mW: each shield catches the radiation from the one outside it and hands the heat to a stage that can afford it.

The lines on the near left carry microwave pulses down to the chip. Each one passes through every plate and is clamped to it, and on three plates it runs through an attenuator, the small dark cylinders: 20 dB at 4 K, 10 dB on the cold plate and 20 dB at the mixing chamber, which are typical values. The lines on the near right bring the signal back up. At the mixing chamber they pass two isolators and a Josephson travelling-wave parametric amplifier, the TWPA; at 4 K, a transistor amplifier, the HEMT. Between the two the cable is made of niobium–titanium, a superconductor. The brown bundles at the back are twisted pairs for DC bias and thermometers.

Press Send a pulse. A readout pulse leaves the room-temperature electronics at −70 dBm, 100 picowatts. On the way down it loses 50 dB in the attenuators and 6 dB more in the cables and in a filter just above the sample, and it reaches the chip at about −126 dBm, some \(2.5\times10^{-16}\) W. At a readout frequency of 7 GHz that is about fifty photons per microsecond. The answer coming back is just as weak. It gains 20 dB in the TWPA and 40 dB in the HEMT, loses a few dB along the way, and leaves the refrigerator at about −70 dBm again. Why the line is built this way, cold attenuators going in and cold amplifiers coming out, is the subject of the second half of this article.

How a dilution refrigerator gets cold

Pumping on a liquid cools it. The most energetic atoms leave as vapour and take heat with them; the rest are colder. Pumping on liquid 4He reaches about 1 K, and on the rarer isotope 3He about 0.3 K. Below that the vapour pressure falls exponentially with temperature, there is almost nothing left to pump, and the cooling stops. A dilution refrigerator gets around this with a mixture of the two isotopes.

Cooled below about 0.87 K, a mixture with enough 3He separates into two liquid phases, a little like oil and water. One, the concentrated phase, is almost pure 3He; it is the lighter of the two and floats on top. The other, the dilute phase, is mostly 4He, which at these temperatures is a superfluid with almost no thermal excitations, so that it is thermally almost inert. But even as the temperature approaches zero, the dilute phase keeps about 6.6 per cent of 3He dissolved in it. A 3He atom is bound slightly more strongly in the dense 4He liquid than among other 3He atoms, and that is what keeps its solubility finite.

Those 6.6 per cent are the whole trick. The two phases meet in the mixing chamber. Its dilute phase is connected through a tube to the still, a pot higher up at about 0.7 K that is pumped by pumps at room temperature. At that temperature 3He has a far higher vapour pressure than 4He, so what leaves the still is almost pure 3He. As 3He is taken out of the dilute phase in the still, 3He from the mixing chamber moves up to replace it, and in the mixing chamber 3He atoms cross the phase boundary, from the concentrated phase into the dilute one, to keep its concentration at 6.6 per cent. Crossing costs energy, because 3He has a higher enthalpy in the dilute phase, and each atom that crosses takes heat from its surroundings, the way a molecule does when it evaporates. The concentrated phase plays the liquid, the dilute phase the vapour, and the 4He is simply the space in which that vapour lives.

The 3He pumped away is compressed at room temperature, cleaned and sent back down. It is precooled by the pulse tube, condensed, and cooled further in heat exchangers by the cold dilute stream on its way up to the still, before it re-enters the mixing chamber. It circulates. For ideal heat exchangers the cooling power is

$$\dot Q = 84\ \frac{\text{J}}{\text{mol K}^2}\;\dot n\,T^2,$$

where \(\dot n\) is the circulation rate of 3He in moles per second and \(T\) is the temperature of the mixing chamber. Because the concentration of the "vapour" stays at 6.6 per cent instead of falling exponentially, the cooling power falls only as \(T^2\). For a typical \(\dot n = 500\) µmol/s it gives about 17 µW at 20 mK and 420 µW at 100 mK. The mixing chamber settles where cooling power and heat load are equal, \(T = \sqrt{\dot Q/(84\,\dot n)}\).

The figure below shows the still, the heat exchanger and the mixing chamber, drawn larger than they are and with the liquids tinted so that you can see them. Cut the walls away with the last slider. The dark beads are 3He atoms, the pale ones 4He. Change the circulation rate and the heat load on the mixing-chamber plate, and the temperature on the thermometer follows the balance above.

—
—
—
T = √(Q̇ / 84ṅ)— Cooling power at 20 mK— at 100 mK—
Figure 2. A schematic of the dilution unit, not to scale. Real heat exchangers are a tube inside a tube followed by blocks of sintered silver; here they are drawn as two interleaved coils, one carrying returning 3He down and the other the dilute stream up. The number of atoms drawn is arbitrary; only the flow pattern and the make-up of each phase mean anything, with roughly one 3He for every fourteen 4He in the dilute phase. The temperature is a toy model: the ideal-exchanger balance \(\dot Q = 84\,\dot n T^2\), with all heat leaks lumped into the heat load. A real refrigerator's base temperature, typically 7–15 mK with nothing attached, is also limited by its heat exchangers.

Follow one dark bead. It comes down the coil from the left, already liquid, and enters the top of the mixing chamber into the concentrated phase. It sinks to the boundary and crosses it, and the ring that flashes there is the moment it absorbs heat. In the dilute phase it is one 3He among fourteen or so 4He atoms. It drifts to the outlet near the bottom, climbs the other coil, passing the incoming stream and cooling it, reaches the still, and evaporates into the pumping line.

Two things are worth trying. Double the circulation rate and the temperature falls only by a factor of \(\sqrt2\), because the cooling power is proportional to \(\dot n\) and to \(T^2\). Then raise the heat load: a few microwatts keep the mixing chamber near 10 mK, but a few hundred microwatts warm it towards 100 mK. Everything that touches the coldest plate, every cable, every support, every stray photon, is paid for out of this budget of microwatts.

Why the input lines are attenuated

A coaxial cable is a pipe for microwaves, and at room temperature it is full of thermal photons. The mean number of photons in a mode of frequency \(f\) at temperature \(T\) is the Bose–Einstein occupation

$$n_{\rm th}(T) = \frac{1}{e^{hf/k_BT}-1},$$

which at 5 GHz and 300 K is about \(k_BT/hf \approx 1250\). If a cable ran straight from room temperature to the chip, the qubit's readout resonator would sit in a 300 K bath. As we will see, every stray photon in that resonator shifts the qubit frequency.

An attenuator is a small resistive network that absorbs a fixed fraction of the power passing through it. At temperature \(T\) it does two things. It passes a fraction \(1/A\) of the photons arriving from above, where \(A\) is its attenuation as a power ratio, and, since anything that absorbs also emits, it adds thermal photons of its own:

$$n_{\rm out} = \frac{n_{\rm in}}{A} + \left(1-\frac{1}{A}\right) n_{\rm th}(T).$$

A 20 dB attenuator, \(A = 100\), at 4 K turns 1250 photons into \(12.5 + 0.99\times16.2 \approx 28.5\). However strong it is, it cannot go below its own \(n_{\rm th}\) of 16.2. The way down is to attenuate a little at each of several stages, so that every attenuator absorbs the noise of the warmer stages above it and replaces it with its own, colder noise.

Why not put all the attenuation at the mixing chamber? Because an attenuator absorbs the signal as well, and absorbed power is heat. A drive line carrying −40 dBm, 100 nW, of pulses into a 20 dB attenuator at the mixing chamber dissipates 99 nW there, for every line, in a stage whose whole budget is some tens of microwatts. Dissipating at 4 K is far cheaper. The attenuators also have a second job: they connect the centre conductor of the coax, which its dielectric otherwise insulates, to the plate, so that the inner conductor is cooled stage by stage too.

The figure shows one input line from the top plate to the sample, with an attenuator on each plate. The glowing cloud between two plates stands for the thermal photons in the line at that point: the number of dots grows with the logarithm of the occupation, so every dot that disappears is a factor of about 1.1 fewer photons. The numbers next to the line are exact, from the formula above. Change the attenuators and the frequency.

—
—
—
—
—
—
n at the chip— Teff = hf / kB ln(1 + 1/n)— Dephasing limit Tφ—
Figure 3. Thermal photons in one input line, from \(n_{\rm out} = n_{\rm in}/A + (1-1/A)\,n_{\rm th}(T)\) applied stage by stage, starting from \(n_{\rm th}(300\ \text{K})\). Stage temperatures are typical: 50 K, 4 K, 0.8 K, 0.1 K and 10 mK. Each attenuator is assumed to be at the temperature of its plate, and cable loss, which acts as a weak distributed attenuator, is left out. Right: the occupation after each stage on a logarithmic scale, with each stage's own \(n_{\rm th}(T)\), the floor its attenuator cannot go below, marked by a grey tick. The dephasing limit assumes the readout resonator is in equilibrium with this line alone, with a typical \(\kappa/2\pi = 1\) MHz and a dispersive shift \(2\chi = \kappa\).

With the typical values, 20 dB at 4 K, 10 dB on the cold plate and 20 dB at the mixing chamber, the line delivers about \(n = 0.029\) photons per mode at 5 GHz. That is equivalent to a bath at 68 mK, a thousand times better than room temperature but still much warmer than the mixing chamber. Notice how little the stages without attenuators do, and how each attenuator meets a floor: no attenuation at 4 K can push the occupation below 16. Add 10 dB at the still, or another 10 dB on the cold plate, and the occupation falls to \(6\times10^{-3}\) or \(4\times10^{-3}\).

Does 0.03 of a photon matter? In the dispersive regime used for readout, each photon in the readout resonator shifts the qubit frequency by \(2\chi\). Thermal photons arrive and leave at random, so the shift jumps at random and the qubit's phase is scrambled at the rate

$$\Gamma_\varphi = \bar n\,\kappa\,\frac{4\chi^2}{\kappa^2+4\chi^2}\qquad(\bar n \ll 1),$$

where \(\kappa\) is the rate at which photons leave the resonator. For a typical \(\kappa/2\pi = 1\) MHz and \(2\chi = \kappa\), an occupation of 0.029 limits the dephasing time to about 11 µs, far below the hundreds of microseconds that good transmons reach. The rate is proportional to \(\bar n\), so every factor of ten in photons is a factor of ten in coherence. In a real setup the readout resonator is also connected to the output line, which is protected differently, and many groups use 60 dB or more on their input lines, together with infrared filters.

Why the output needs amplifiers

The answer from the chip is a microwave tone of about −126 dBm, a few tens of photons per microsecond, whose amplitude or phase differs slightly depending on whether the qubit is in \(|0\rangle\) or \(|1\rangle\). To tell the two apart within a microsecond, the noise added on the way out has to be comparable to that, a photon or two per unit of bandwidth and time. Every amplifier adds noise, characterised by a noise temperature \(T_i\), and a chain of amplifiers with power gains \(G_i\) adds noise according to the Friis formula:

$$T_{\rm sys} = T_1 + \frac{T_2}{G_1} + \frac{T_3}{G_1G_2} + \cdots$$

The first amplifier's noise counts in full; the second's is divided by the gain of the first. So the first amplifier must be as quiet as possible, and must have enough gain to make the rest of the chain irrelevant.

A HEMT amplifier at 4 K has a noise temperature of a few kelvin, say 3 K, and about 40 dB of gain. Followed by a room-temperature amplifier with a noise temperature of about 100 K, the system noise temperature is \(T_{\rm sys} \approx 3 + 100/10^4 \approx 3.01\) K, about nine photons of noise at 7 GHz. A TWPA at the mixing chamber, with 20 dB of gain and an effective noise temperature of about 0.4 K, which includes the losses before it, changes this to \(T_{\rm sys} \approx 0.4 + 3/100 + 100/10^6 \approx 0.43\) K, close to one photon. Quantum mechanics sets a floor: an amplifier that preserves both quadratures of a signal must add at least half a photon of noise, \(hf/2k_B \approx 0.17\) K at 7 GHz. The time needed to reach a given signal-to-noise ratio is proportional to \(T_{\rm sys}\), so the TWPA makes readout about seven times faster at the same power, or allows a weaker readout tone.

—
—
Tsys— kBTsys / hf— Time for equal SNR—
Figure 4. System noise temperature from the Friis formula at a typical readout frequency of 7 GHz, with typical gains of 20 dB for the TWPA and 40 dB for the HEMT, and a room-temperature amplifier at 100 K. The button removes the TWPA from the chain; the faded bar shows the other configuration for comparison. The time needed for a given signal-to-noise ratio scales with \(T_{\rm sys}\) and is given relative to the chain with the TWPA. The dashed line is the quantum limit for a phase-preserving amplifier, \(hf/2k_B\).

Two details of the output line in Figure 1 follow from this. The isolators, circulators with one port terminated, let the signal pass upwards and absorb whatever travels down: the noise of the amplifiers, and the strong pump tone that the TWPA needs, which must not reach the qubit. And any loss before the first amplifier attenuates the signal but not the noise added after it, so it counts directly against \(T_{\rm sys}\). That is why the cable between the mixing chamber and the HEMT is superconducting niobium–titanium, which loses almost nothing and, like all superconductors, conducts little heat.

The sample package

At the bottom of Figure 1, below the mixing-chamber plate, a gold-plated copper cold finger holds the sample package, inside a magnetic shield. The package is a copper box, often gold-plated, sometimes made of aluminium, which becomes superconducting and screens magnetic fields. Inside it a printed circuit board carries coplanar-waveguide traces from coaxial connectors at its edge to the chip in its centre. The chip is glued or clamped into a pocket, and aluminium wire bonds, 25 µm thick, connect the traces on the board to pads on the chip. Many more bonds tie the ground plane of the chip to the ground of the board all around its edge; without them, parts of the chip's ground could ring on their own.

The box itself is a microwave cavity. Its lowest mode, for a pocket of size \(a\times b\) that is not too tall, has the frequency

$$f = \frac{c}{2}\sqrt{\frac{1}{a^2}+\frac{1}{b^2}},$$

17.7 GHz for a 12 mm square pocket and 8.2 GHz for a 26 mm one. A box mode close to a qubit or a resonator couples to it and opens a new path for its energy to leak away, so pockets are kept small, or broken up with posts, to push their modes well above the band from roughly 4 to 8 GHz where qubits and readout resonators typically live. On large chips, where the pocket cannot be small, the same job is done by air bridges and by vias through the silicon.

Two more shields complete the picture. A magnetic shield of high-permeability metal, sometimes with a superconducting one inside it, keeps the field at the chip small, because flux vortices trapped in the superconducting films as they cool through their transition cause loss. And the whole sample space is made light-tight against infrared radiation. A single photon above \(2\Delta/h\), about 90 GHz for aluminium, can break a Cooper pair, and the quasiparticles it creates make qubits relax; this is the reason for the closed shields, for blackened inner surfaces, and for the infrared-absorbing filter on the input line just above the sample. Lift the lid and go down to the chip:

—
—
—
Lowest mode of the empty pocket— —
Figure 5. A sample package of a common kind, drawn at its real size: a 32 mm copper box, a 0.5 mm circuit board and a 6 mm chip. On the chip, line widths, gaps and the metal thickness are enlarged several times so that they can be seen; a real feedline is about 10 µm wide and the metal is 100–200 nm thick. The chip has two transmons, each with a readout resonator coupled to a common feedline and its own drive line. The box-mode frequency is the estimate for an empty rectangular pocket; the silicon chip and the circuit board lower it somewhat. The red sheet shows where the electric field of that mode is strongest.

At the far end of the zoom are the two pads, the junction between them and the meandering resonator of the previous article. Everything described there, the phase across a nanometre of oxide, the anharmonic ladder of levels, the Rabi oscillations, happens in a few square micrometres at the very bottom of this machine. The rest of it, the pulse tube and the helium mixture, the shields, the attenuators and the amplifiers, exists so that this small patch of aluminium sees almost nothing but the pulses we send it: not the heat of the room, not the noise of our electronics, not a stray infrared photon. A quantum computer is, to a large extent, this machine.

Further in my notes