Packing Qubits:
Density and Connectivity Across Quantum Computing Modalities

I keep running into a question that sounds simple and turns out to be surprisingly deep: how many qubits can actually fit in a given piece of space, and once they are packed in, who can talk to whom? In this post I work through the packing density of superconducting circuits, semiconductor spin qubits, trapped ions, neutral atoms in optical tweezers, and photonic modes. I put real numbers on the areal density of each platform, I show why the wiring and not the qubit is usually the thing that sets the pitch, and I finish by converting raw physical density into the quantity I actually care about, which is logical qubits per unit area once quantum error correction is paid for.

Five ways of filling a plane with qubits, from sparse to dense: superconducting crosses, a ring of trapped ions, a tweezer array of neutral atoms, a fine crossbar of spin qubits and streams of photonic pulses.

From millimetre-scale superconducting circuits to 100 nm quantum dots: five ways of filling a plane with qubits, and five ways of letting them talk to each other.

1Why Packing Density Is Worth Thinking About

I used to treat qubit count as the only number that mattered. A processor with 1000 qubits beats one with 100, end of story. Then I started paying attention to floor plans, and the picture changed.

Here is the thing that made it click for me. A rotated surface code with distance \(d\) needs

\[ n(d) = 2d^2 - 1 \]

physical qubits to protect a single logical qubit. At a distance of 25, which is a fairly modest target for running a useful algorithm, that is 1249 physical qubits for one logical qubit. If each physical qubit occupies a square millimetre of silicon, one logical qubit swallows about 12.5 square centimetres. A 300 millimetre wafer has roughly 70,000 square millimetres of usable area, so the entire wafer holds around 56 logical qubits. That is a sobering number, and it comes purely from geometry.

So packing density is not a vanity metric. It sets how much refrigerator volume, how much wafer area, how much vacuum chamber, and how much control wiring a fault tolerant machine needs. And density interacts with a second quantity that is just as important: connectivity. A very dense array of qubits that can only reach their four immediate neighbours is a different computational object from a sparse array in which any qubit can be brought next to any other. Density buys you scale. Connectivity buys you code efficiency. Most of the engineering fight in quantum hardware right now is about the exchange rate between those two currencies.

2The Numbers I Will Be Using

Before comparing platforms I need to fix definitions, because it is easy to compare apples with vacuum chambers.

2.1Areal density

For a two dimensional array on a square lattice with centre to centre pitch \(p\), the areal density is

\[ \rho_A \;=\; \frac{N}{A} \;=\; \frac{1}{p^{2}}. \]

A pitch of \(1\) mm gives \(\rho_A = 1\) qubit per square millimetre. A pitch of \(100\) nm gives

\[ \rho_A = \frac{1}{(10^{-4}~\text{mm})^{2}} = 10^{8}~\text{qubits per square millimetre}. \]

Eight orders of magnitude separate those two cases, and both are real platforms in real laboratories today. That gap is the single most striking fact in this whole post.

2.2Two footprints, not one

I find it essential to distinguish two areas:

  • The qubit footprint \(a_q\), meaning the area of the quantum object itself, such as the capacitor pads of a transmon or the confining potential of a quantum dot.
  • The cell footprint \(a_c\), meaning the area actually consumed once couplers, readout structures, control lines and routing layers are included.

The ratio \(a_c / a_q\) is the overhead factor \(f\), and for most solid state platforms \(f\) is somewhere between 3 and 30. A useful way to write it:

\[ \rho_A^{\text{practical}} \;=\; \frac{1}{f\,a_q}. \]

Whenever a press release quotes a qubit size, it is almost always \(a_q\). The number that determines how big the machine gets is \(a_c\).

One superconducting qubit cell. A small cross-shaped qubit sits in the middle of a much larger square cell that also holds couplers, a readout resonator, drive and flux lines and vertical interconnects. aq ac 1 2 3 4 p
  • Qubit footprint \(a_q\): the quantum object itself, here the cross-shaped capacitor of a transmon.
  • 1Tunable couplers to the four neighbours, sitting on the shared cell boundary.
  • 2Readout resonator.
  • 3Drive and flux lines.
  • 4Vertical interconnects to the wiring layers.
Figure 1. The two footprints of one superconducting qubit. The cell footprint \(a_c\) (teal outline) has to hold the qubit and everything that serves it, and on a square lattice it is simply \(a_c = p^2\). With a capacitor of about 0.1 mm² in a cell about a millimetre across (Section 3.1), a transmon has an overhead factor \(f \approx 10\).

2.3Connectivity

I describe connectivity by a graph \(G = (V, E)\) whose vertices are qubits and whose edges are native two qubit interactions. Three properties matter to me:

  • The degree \(\deg(v)\), meaning how many partners a qubit can directly reach. Surface codes need 4. Many quantum low density parity check codes want 6 or more.
  • The interaction range \(r_{\max}\) in physical units, meaning how far apart two coupled qubits can sit.
  • The graph diameter \(D\), meaning the worst case number of hops between two qubits. On an \(L \times L\) grid with \(N = L^2\), the diameter is \(D \approx 2(\sqrt{N} - 1)\).

The diameter is the one that hurts in compilation. If a circuit wants a gate between two randomly chosen qubits on a grid, and the only way to move quantum information is by SWAP gates, the cost scales as

\[ C_{\text{SWAP}} \sim \mathcal{O}(\sqrt{N}), \]

which for \(N = 10^4\) means roughly one hundred extra two qubit gates for every long range interaction. At a two qubit error rate of \(10^{-3}\), one hundred SWAPs contribute about a 10 percent chance of failure on their own, which is often enough to destroy the computation.

Tap or click two qubits

A B
12 hops separate A and B, so 11 SWAP gates must run before the two qubits are neighbours. At an error rate of 10⁻³ per gate, those SWAPs alone fail about 1.1% of the time.
Figure 2. SWAP routing on a 7 × 7 grid with nearest neighbour gates only. The route between opposite corners is the diameter, \(D = 2(\sqrt{49} - 1) = 12\) hops. The same corner to corner route on a 100 × 100 grid (\(N = 10^4\)) is 198 hops long.

3Superconducting Circuits: Big Qubits, Tidy Lattices

3.1Why a transmon is so large

A transmon is a Josephson junction shunted by a capacitor. The junction itself is tiny, roughly \(100 \times 100\) nm. The capacitor is the problem. To reach a charging energy in the right range while keeping the electric field out of lossy dielectric interfaces, designers use large planar electrodes with vacuum or high quality substrate between them. Typical planar transmon capacitor structures occupy of order \(10^{5}~\mu\text{m}^{2}\), which is about \(0.1\) mm² [1].

The analogy I like: the junction is the engine of a car, and the capacitor is the parking lot the car needs in order not to scrape against anything. The engine is small. The parking lot is enormous, and it exists purely to keep the car away from lossy surfaces.

Once couplers, readout resonators, feedlines and vertical interconnects are added, the practical cell pitch on current processors lands near a millimetre, giving an areal density of order

\[ \rho_A^{\text{SC}} \approx 1~\text{qubit per square millimetre} = 10^{2}~\text{qubits per square centimetre}. \]

This is why a 100 qubit superconducting chip is about a centimetre on a side, and it explains why Google's Willow processor with 105 qubits and IBM's Nighthawk processor with 120 qubits both live on dies of roughly that scale [2][3].

A one millimetre superconducting qubit cell drawn to scale, holding two capacitor pads that cover about a tenth of a square millimetre, with a zoomed inset of the Josephson junction, about one hundred nanometres on a side. 1 mm × 1 mm cell capacitor pads ≈ 0.1 mm² 200 μm junction, zoomed in Josephson junction ≈ 100 nm × 100 nm 100 nm
Figure 3. The engine and the parking lot, with the cell and the pads drawn to scale. The Josephson junction (inset) is about 100 nm on a side; the two capacitor pads cover about \(10^{5}~\mu\text{m}^{2}\), some \(10^{7}\) times the area of the junction [1]; and the cell that also holds the couplers and the readout resonator is about a millimetre across.

3.2What the density buys and what it costs

The payoff of a fixed lithographic lattice is regularity. Google's Willow demonstration of a distance 7 surface code used 101 qubits on the 105 qubit device, and showed the logical error rate dropping by a factor of \(\Lambda = 2.14 \pm 0.02\) each time the distance grew by two [3]. That result depends on having a clean, uniform, degree 4 square lattice, which is exactly what lithography is good at producing.

The cost is that the lattice is frozen. IBM's Nighthawk moved from the heavy hexagonal layout to a square lattice with 120 qubits and 218 tunable couplers, giving each interior qubit four neighbours, and that change alone reduced SWAP overhead enough to support noticeably deeper circuits. But a qubit in the top left corner of the chip still cannot interact with a qubit in the bottom right without a long chain of intermediate operations.

3.3Can transmons be shrunk?

Yes, and this is one of my favourite results in the area. Antony and colleagues replaced the planar capacitor with a parallel plate capacitor built from crystalline niobium diselenide electrodes separated by hexagonal boron nitride, and reduced the qubit area by more than a factor of 1000 while preserving the required capacitance [1]. The catch is coherence: they measured a relaxation time \(T_1\) of about \(1.06~\mu\text{s}\), which is roughly two orders of magnitude short of what conventional transmons achieve.

That trade is the whole story of superconducting density in one experiment. Shrinking the capacitor increases the fraction of electric field energy sitting inside dielectric material, and that material hosts two level system defects. Writing the participation ratio of the lossy region as \(\eta\) and its loss tangent as \(\tan\delta\), the quality factor obeys

\[ \frac{1}{Q} \;\approx\; \eta \tan\delta \;+\; \frac{1}{Q_{\text{other}}}, \]

and \(\eta\) grows as the plates come closer together. Density and coherence pull in opposite directions unless the dielectric itself gets better.

Cross-section of a planar capacitor: two flat pads on a substrate, with field lines arcing through the vacuum above and the substrate below. vacuum substrate

(a) Planar pads on a substrate

Cross-section of a van der Waals parallel plate capacitor: niobium diselenide, hexagonal boron nitride and niobium diselenide stacked on a substrate, with the field confined to the dielectric. vacuum substrate NbSe2 hBN NbSe2

(b) Van der Waals parallel plate

Figure 4. Density against coherence, in cross-section. (a) Planar pads spread their field through vacuum and substrate, so only a thin sliver of it crosses the lossy surface layers (red) and \(\eta\) stays small. (b) The parallel plate stack stores the same energy in a thin dielectric. In the device of Antony and colleagues the NbSe₂/hBN/NbSe₂ capacitor covers about 109 μm² with 35 nm of hBN, roughly 1400 times smaller than the planar capacitor it replaced, but nearly all of its field now sits inside material that can host two level systems [1].

4Semiconductor Spin Qubits: The Density Champion

4.1One hundred nanometres

A spin qubit is a single electron or hole confined in a gate defined quantum dot in silicon or germanium. The confining potential is set by lithographic gates, and the whole structure is comparable in size to a transistor. Standard crossbar architecture proposals assume barrier and plunger gates of 30 and 40 nm width and a quantum dot pitch of 100 nm [6].

At that pitch the raw areal density is

\[ \rho_A^{\text{spin}} = \frac{1}{(0.1~\mu\text{m})^2} = 100~\text{qubits per square micrometre} = 10^{8}~\text{qubits per square millimetre}. \]

To make that concrete, the same architecture paper estimates that two billion trapped ion qubits in a microwave driven design would need an area larger than \(100~\text{m} \times 100~\text{m}\), two billion superconducting qubits would need roughly \(5~\text{m} \times 5~\text{m}\), while two billion spin qubits could in principle fit inside \(5~\text{mm} \times 5~\text{mm}\) [6]. I find that comparison genuinely startling every time I reread it.

Three nested zooms comparing the area two billion qubits would need: more than 100 by 100 metres for microwave-driven trapped ions, about 5 by 5 metres for superconducting qubits, and less than 5 by 5 millimetres for spin qubits. zoom ×20 zoom ×1000 trapped ions > 100 m × 100 m superconducting ≈ 5 m × 5 m spin qubits < 5 mm × 5 mm
Figure 5. The footprint comparison of the crossbar paper drawn as a chain of zooms [6]: each small marked square is enlarged into the next panel. The side length shrinks twenty-fold from trapped ions to superconducting qubits, and then a thousand-fold again from superconducting qubits to spins.

4.2Why nobody actually achieves that density yet

Because of wiring, which I will come back to in Section 10. Real spin qubit devices are still small. Recent work in planar germanium presents a crossbar chip with tightly pitched spin qubit tiles capable in principle of hosting 1058 hole spin qubits, with measurements demonstrated on a subset of 40 tiles [8]. Sparse architectures deliberately throw away density to make room for local control electronics and shuttling channels, often reducing the effective density by three to five orders of magnitude relative to the raw 100 nm pitch.

4.3Connectivity of spins

The native two qubit interaction is exchange coupling, which arises from wavefunction overlap and therefore falls off exponentially with separation:

\[ J(r) \;\approx\; J_0 \exp\!\left(-\frac{2r}{a_B}\right), \]

where \(a_B\) is an effective Bohr radius of a few nanometres in silicon. Exchange is a genuinely short range interaction: it is strong at 100 nm and effectively zero at 300 nm. So the native graph is a square lattice of degree at most 4, with a very hard cutoff [7].

The escape route is shuttling. Because an electron can be moved coherently along a channel of gates, spin platforms can trade time for connectivity, moving a qubit tens of micrometres to meet a distant partner. This is the same idea trapped ions use, implemented in a solid.

Cross-section of a row of gate-defined quantum dots at 100 nanometre pitch, with 40 nanometre plunger gates and 30 nanometre barrier gates. Below it, the potential landscape: a lowered barrier lets two electron wavefunctions overlap, switching exchange on, while a raised barrier keeps the third dot isolated. 40 nm 30 nm P P P B B 100 nm pitch barrier low: J on barrier high: J ≈ 0 gates well energy
Figure 6. A row of gate defined quantum dots at the 100 nm pitch of the crossbar proposal, with 40 nm plunger (P) and 30 nm barrier (B) gates [6]. Each plunger holds one spin. Exchange needs the wavefunctions to overlap (gold), so lowering the barrier between two dots switches it on and raising it switches it off; nothing couples dots that are not neighbours.

5Trapped Ions: Sparse in Space, Complete in Connectivity

5.1The spacing question

Ions in a linear Coulomb crystal sit a few micrometres apart, typically 3 to 5 μm, set by the balance between Coulomb repulsion and the trapping potential [9]. That sounds dense. It is misleading, because each ion needs a region of trap electrodes, laser access, and often a dedicated operation zone. The chip that supports the ions is centimetre scale.

Quantinuum's Helios is the clearest current example. It holds 98 barium ions on a chip roughly the size of a thumbnail, giving an effective density of order one qubit per square millimetre, which is remarkably close to the superconducting figure despite the completely different physics [10]. And the chip is the small part. The lasers, optics, vacuum system and control electronics occupy cubic metres.

5.2The compensation: everything talks to everything

What trapped ions give up in density they recover entirely in connectivity. Helios uses the quantum charge coupled device architecture, in which ions are physically transported between a storage ring and eight operation zones through a junction. Any pair of qubits can be brought together and entangled, which is genuine all to all connectivity, and the reported average infidelities are \(2.5(1)\times 10^{-5}\) for single qubit gates and \(7.9(2)\times 10^{-4}\) for two qubit gates [10].

In graph language the coupling graph is the complete graph \(K_N\), with

\[ \deg(v) = N - 1, \qquad D = 1, \qquad C_{\text{SWAP}} = 0 . \]

The cost moves out of the gate count and into wall clock time, since transporting an ion takes hundreds of microseconds. That is a much better place for the cost to live, because transport is slow but nearly error free, whereas SWAP chains are fast but noisy.

6Neutral Atoms: Dense and Reconfigurable at the Same Time

6.1Six thousand atoms in a millimetre

Optical tweezer arrays are where the density picture gets genuinely exciting. Manetsch and colleagues trapped more than 6100 neutral caesium atoms in roughly 12,000 tweezer sites, within a region about one millimetre across, while measuring a coherence time of \(12.6(1)\) s [11].

Running the arithmetic on a circular region of diameter 1 mm:

\[ A = \pi \left(\frac{1~\text{mm}}{2}\right)^{2} \approx 0.79~\text{mm}^{2}, \qquad \rho_A^{\text{atoms}} \approx \frac{6100}{0.79} \approx 7.8 \times 10^{3}~\text{qubits per square millimetre}. \]

That is nearly four orders of magnitude denser than a superconducting chip, achieved with no lithography at all, using only interference and focused light. Corresponding tweezer pitches are in the range of 3 to 10 μm.

6.2Blockade sets the natural interaction range

The native two qubit interaction is the Rydberg blockade. When two atoms are excited towards a Rydberg state with Rabi frequency \(\Omega\), their van der Waals interaction \(V(r) = C_6 / r^6\) shifts the doubly excited state out of resonance whenever the atoms are closer than the blockade radius

\[ R_b \;=\; \left(\frac{C_6}{\hbar \Omega}\right)^{1/6}. \]

For typical parameters \(R_b\) is around 5 to 10 μm. The sixth power in the denominator makes this a wonderfully sharp switch: at \(r = R_b/2\) the interaction is 64 times stronger than at \(R_b\), and at \(r = 2R_b\) it is 64 times weaker. Blockade is essentially binary, which is why it makes such a clean gate.

Two atoms and the Rydberg blockade radius, above a log-log plot of the van der Waals interaction against separation. The interaction falls as the sixth power of distance and crosses the drive strength at the blockade radius. Rb r blockade independent 10⁻⁴ 10⁻² 1 10² 10⁴ ¼ ½ 1 2 4 V = ħΩ separation r / Rb interaction V / ħΩ (log scale) 64 ħΩ ħΩ ħΩ / 64
r = 0.50 Rb
V = 64 ħΩ. The doubly excited state is pushed far off resonance, so only one of the two atoms can be excited: blockaded.
Figure 7. Drag the slider to move the second atom. Because \(V = C_6/r^6\), every halving of the separation multiplies the interaction by 64, so on log axes it is a straight line of slope \(-6\) that crosses the drive strength \(\hbar\Omega\) exactly at \(R_b\).

6.3Moving atoms rewrites the graph

Here is the property that makes atom arrays architecturally special. Atoms can be picked up by mobile tweezers and physically transported across the array while remaining in superposition. Bluvstein and colleagues used exactly this to run a logical processor on reconfigurable atom arrays, entangling logical qubits by relocating physical atoms rather than by routing gates through fixed wires [12]. Manetsch and colleagues demonstrated transport across hundreds of micrometres in the 6100 atom array while preserving coherence [11].

So the connectivity graph is not fixed by fabrication. It is programmed, cycle by cycle, by where the tweezers put the atoms. This gives the platform both very high density and effectively nonlocal connectivity, which is a combination no solid state platform currently offers.

7Photonics: A Different Accounting System

Photonic platforms break my framework, and it is instructive to see how.

A photonic qubit is not a stationary object sitting at a lattice site. It is a mode of the electromagnetic field, generated, interfered and measured as it flies past. Asking for its areal density is a little like asking how densely packed the cars on a motorway are: the answer depends entirely on whether you count the cars present at one instant or the cars that pass a point over an hour.

Xanadu's Aurora machine makes the point concretely. It combines 84 squeezers and 36 photon number resolving detectors across 35 photonic chips housed in four server racks connected by 13 km of optical fibre, and it furnishes 12 physical qubit modes per clock cycle. Yet the cluster state it synthesised was entangled across 86.4 billion modes [13].

Twelve qubits filling four server racks is, in spatial terms, the worst density in this entire post by an enormous margin, something like \(10^{-6}\) qubits per square millimetre of rack footprint. But 86.4 billion entangled modes is a number no stationary platform can approach. The reason is time multiplexing: photonics converts spatial density into temporal density,

\[ N_{\text{effective}} \;=\; N_{\text{spatial}} \times \frac{T_{\text{coherent}}}{\tau_{\text{clock}}}, \]

where \(\tau_{\text{clock}}\) is the pulse period. Because photons do not decohere while travelling through fibre, the effective mode count is enormous even when the instantaneous qubit count is tiny. Connectivity, meanwhile, comes from fibre routing, so it is limited by loss rather than by geometry.

A space-time grid of photonic modes: twelve rows of spatial modes and one column per clock cycle, linked by entangling bonds. Only the current column exists at any instant; measured columns lie to its left. … 12 modes τclock time now
Clock cycle 7: 12 modes exist right now, and the cluster state so far spans 84 modes.
Figure 8. Counting the cars that pass rather than the cars present. Each dot is one mode, each column one clock cycle, and the lines are entangling bonds of the cluster state. Only the highlighted column exists in the machine at any instant, and earlier columns have already been measured. Aurora ran at a 1 MHz clock for two hours: 7.2 billion cycles of 12 modes, which is where the 86.4 billion comes from [13].

8Side by Side Comparison

Table 1 collects the density figures and Table 2 collects the connectivity properties. I have rounded aggressively, because the point is the orders of magnitude rather than the third significant figure.

Table 1. Approximate packing density across modalities. Densities are order of magnitude estimates for the qubit plane only, excluding cryostats, lasers and control racks.
Modality Qubit pitch Areal density
(per mm²)
Largest reported
single array
Set by
Superconducting\(\sim 1\) mm\(\sim 10^{0}\)120 to 156Capacitor size, wiring
Spin qubits (raw)\(\sim 100\) nm\(\sim 10^{8}\)tensLithographic pitch
Spin qubits (with control)1 to 10 μm\(10^{4}\) to \(10^{6}\)\(\sim 10^{3}\) tilesInterconnect
Trapped ions3 to 5 μm in chain\(\sim 10^{0}\)98Trap zones, optics
Neutral atoms3 to 10 μm\(\sim 10^{4}\)6100Optics, laser power
Photonic modesnot applicable\(\sim 10^{-6}\)12 per cycleRack and fibre volume
Areal density of each modality on a logarithmic axis from one qubit per square metre to a hundred million per square millimetre, with the equivalent square-lattice pitch shown above. equivalent pitch p 1 m 1 mm 10 μm 100 nm ×10⁸ Spin qubits raw 100 nm pitch Spin qubits (raw 100 nm pitch): about 10⁸ per mm² Spin qubits with control Spin qubits (with control): 10⁴ to 10⁶ per mm² Neutral atoms 6100-atom array Neutral atoms (6100-atom array): about 7.8 × 10³ per mm² Superconducting ≈ 1 mm cell Superconducting (≈ 1 mm cell): about 1 per mm² Trapped ions chip only Trapped ions (chip only): about 1 per mm² Photonic modes rack footprint Photonic modes (rack footprint): about 10⁻⁶ per mm² 10⁻⁶ 10⁻⁴ 10⁻² 1 10² 10⁴ 10⁶ 10⁸ qubits per mm² (log scale)
Figure 9. The densities of Table 1 on a single logarithmic axis, with the equivalent square lattice pitch \(p = 1/\sqrt{\rho_A}\) along the top. Superconducting circuits and bare quantum dots sit eight orders of magnitude apart; the photonic point counts rack floor space and is there for scale.
Table 2. Connectivity offered by each modality. Degree refers to the native two qubit coupling graph before any routing.
Modality Native degree Interaction range Graph type Long range route
Superconducting3 to 6\(\sim 1\) mmFixed latticeSWAP chains, couplers
Spin qubitsup to 4\(\sim 100\) nmFixed latticeElectron shuttling
Trapped ions\(N-1\)Whole chipComplete graphIon transport
Neutral atoms4 to 6 in blockade5 to 10 μmReconfigurableAtom transport
Photonic modesset by circuitFibre lengthProgrammableFibre switching

The pattern that jumps out at me is an inverse relationship. The two platforms with the highest density, spin qubits and neutral atoms, have the shortest native interaction ranges, measured in nanometres and micrometres respectively. The platform with perfect connectivity, trapped ions, has among the lowest densities. Neutral atoms are the interesting exception, because atom transport lets them buy connectivity back without giving up density, at the price of time.

9Three Ways to Wire a Processor

Figure 10 sketches the three architectural strategies I have been describing.

A four by four square lattice of qubits with nearest-neighbour couplings; one interior qubit's four couplings are highlighted. degree 4

(a) Fixed lattice. Superconducting and spin qubits. Degree 4, everything hardwired.

A ring of twelve trapped ions joined by a junction to two gate zones, each holding an ion; one ion travels from the ring through the junction into a gate zone and back.

(b) Storage ring plus gate zones. Trapped ions. Any pair can be transported together.

A five by five array of neutral atoms. One atom leaves its site, which is left open, and is carried along the gaps between rows and columns to sit next to a distant partner.

(c) Reconfigurable array. Neutral atoms. The graph is rewritten by moving atoms.

Figure 10. Three strategies for connecting densely packed qubits. The dashed curve in panel (c) indicates coherent transport of one atom to a distant partner, and the open circle marks the vacated site.

10The Real Limit Is Wiring, Not Physics

Every time I look closely at a density limit, I find an interconnect problem rather than a physics problem. Nothing in quantum mechanics forbids a superconducting chip with a 10 μm pitch. What forbids it is that every qubit needs control lines, and those lines must reach room temperature.

10.1Rent's rule, borrowed from classical chip design

Classical integrated circuit designers use an empirical relation between the number of external terminals \(T\) and the number of internal components \(N\):

\[ T \;=\; t\,N^{r}, \]

where \(t\) is terminals per component and \(r\) is the Rent exponent. For quantum processors with fully individual addressing, every qubit needs its own control and readout path, so \(r = 1\) and the wiring grows linearly with qubit count. This is a disaster at scale: a million qubit machine needs a million or more coaxial paths.

Crossbar architectures attack precisely this exponent. By defining the qubit grid with shared row and column control lines, an \(L \times L\) array needs only order \(L = \sqrt{N}\) lines per direction, giving

\[ T \approx 2\sqrt{N} \quad \Longrightarrow \quad r = 0.5 . \]

For \(N = 10^{6}\) this is the difference between \(10^{6}\) wires and \(2\times 10^{3}\) wires. That single change in exponent is the reason crossbar schemes appear again and again in spin qubit architecture proposals [6][8].

The price is addressability. Shared lines mean qubits sharing a row cannot be driven fully independently, and the compiler must schedule operations to respect that. Density is bought with a loss of parallelism.

Sixteen qubits, each with its own control line routed to a pad at the chip edge: sixteen lines in total.

(a) Individual addressing: 16 qubits, 16 lines

Sixteen qubits at the crossings of four shared row lines and four shared column lines: eight lines in total.

(b) Crossbar: 16 qubits, 8 shared lines

Log-log plot of control lines against qubit count. Individual addressing grows as N, a crossbar as two times the square root of N; at a million qubits that is a million lines against two thousand. A real germanium crossbar chip with 1058 qubit sites and 53 lines sits on the crossbar trend. control lines T 1 10² 10⁴ 10⁶ 10⁸ 10 10² 10⁴ 10⁶ 10⁸ qubits N 10⁶ lines 2 × 10³ lines QARPET: 1058 sites, 53 lines individual, r = 1 crossbar, r = 0.5

(c) Control lines against qubit count: individual addressing, crossbar

Figure 11. Rent's rule for qubits. Individual addressing gives \(T = N\); a crossbar gives \(T = 2\sqrt{N}\), and at \(N = 10^{6}\) that is 500 times fewer lines. A real example sits on the crossbar trend: the germanium chip of Tosato and colleagues addresses 1058 qubit sites with 53 control lines [8].

10.2The thermal budget

There is a second, harsher limit. A dilution refrigerator delivers of order one watt of cooling at the 4 K stage and only tens of microwatts at the 10 mK stage. If each control line dissipates a power \(P_{\text{line}}\) at the coldest stage, the maximum qubit count is bounded by

\[ N_{\max} \;\le\; \frac{P_{\text{cool}}}{P_{\text{line}}} . \]

With \(P_{\text{cool}} \sim 20~\mu\text{W}\) and \(P_{\text{line}} \sim 1~\text{nW}\), this caps the machine near \(2 \times 10^{4}\) qubits per refrigerator, independent of how small the qubits are. This is the strongest argument I know for cryogenic control electronics, and it is also why platforms that operate at room temperature or use optical control face a very different scaling curve.

A dilution refrigerator drawn as six stacked plates from room temperature down to about ten millikelvin, with control lines running down to the qubit chip. About one watt of cooling is available at four kelvin and only tens of microwatts at the coldest plate. 300 K 50 K 4 K ≈ 1 K ≈ 100 mK ≈ 10 mK qubit chip ≈ 1 W of cooling ≈ tens of μW electronics
1 nW
With 1 nW per line and about 20 μW of cooling at the coldest plate, one refrigerator supports at most Nmax ≈ 20,000 qubits.
Figure 12. Each plate of a dilution refrigerator is colder than the one above it and has less cooling power to spend. The control lines (teal) carry heat all the way down, so the tens of microwatts available at the coldest plate cap how many lines, and therefore how many qubits, one refrigerator can serve.

11From Physical Density to Logical Density

Now I can put the two halves together, because the quantity I actually want is not qubits per square millimetre but logical qubits per square millimetre.

11.1The conversion

For a code with parameters \([[n, k, d]]\), encoding \(k\) logical qubits into \(n\) physical qubits at distance \(d\), the logical density is

\[ \rho_L \;=\; \rho_A \cdot \frac{k}{n}. \]

For the rotated surface code, \(k = 1\) and \(n = 2d^2 - 1\), so the rate falls quadratically with distance. At \(d = 12\) this gives \(k/n = 1/287\).

Compare that with the bivariate bicycle codes of Bravyi and colleagues, whose \([[144, 12, 12]]\) instance protects 12 logical qubits with 144 physical qubits at the same distance, giving \(k/n = 1/12\) [5]. That is roughly 24 times better logical density at equal distance.

But those codes are not free. They require each qubit to participate in six checks, including some between qubits that are not physical neighbours. On a fixed planar lattice that means long range couplers, which is exactly the capability IBM is prototyping. On an atom array, the same nonlocality comes almost for free by moving atoms [12].

So here is the sentence I would underline if this were on paper: connectivity is what converts physical density into logical density.

11.2A worked comparison

Table 3 works out the area per logical qubit at distance 25 under the surface code, using the practical densities from Table 1.

Table 3. Area required per surface code logical qubit at distance 25, needing \(n = 2(25)^2 - 1 = 1249\) physical qubits. Figures are order of magnitude and count qubit plane area only.
Modality Physical density (per mm²) Area per logical qubit
Superconducting\(1\)\(\sim 1250\) mm² (about 12.5 cm²)
Spin qubits, sparse control\(10^{4}\)\(\sim 0.13\) mm²
Spin qubits, dense crossbar\(10^{6}\)\(\sim 1.3 \times 10^{-3}\) mm²
Neutral atoms\(8 \times 10^{3}\)\(\sim 0.16\) mm²

Two observations follow immediately.

First, the existing 6100 atom tweezer array already contains enough physical qubits for roughly four or five distance 25 surface code logical qubits, in a region the size of a pinhead. It is not yet performing those operations at the required fidelity, but the real estate is there today.

Second, superconducting processors need a different strategy, and the field knows it. Reaching a hundred logical qubits at one square centimetre each means either wafer scale integration, modular chiplets connected by couplers, or higher rate codes that reduce \(n\). All three are being pursued, and the third is the cheapest if the connectivity can be built.

11.3Computing this for any platform

Algorithm 1 is the small calculation I run whenever I read a new hardware announcement and want to know what it means for fault tolerance.

Algorithm 1 Logical qubit density from platform parameters
Require: qubit pitch \(p\), overhead factor \(f\), code parameters \((n, k, d)\), target logical error \(\epsilon_L\)
Ensure: logical density \(\rho_L\) and area per logical qubit \(A_L\)
  1. \(a_c \gets f \cdot p^{2}\)\(\triangleright\) cell footprint including routing
  2. \(\rho_A \gets 1 / a_c\)\(\triangleright\) physical qubits per unit area
  3. if \(d\) is not fixed then
  4. choose smallest \(d\) satisfying \(A(p_{\text{phys}}/p_{\text{th}})^{(d+1)/2} \le \epsilon_L\)
  5. end if
  6. \(R \gets k / n\)\(\triangleright\) code rate
  7. \(\rho_L \gets \rho_A \cdot R\)
  8. \(A_L \gets 1 / \rho_L\)
  9. return \(\rho_L\), \(A_L\)

The threshold condition in line 4 uses the standard surface code scaling

\[ p_L \;\approx\; A\left(\frac{p_{\text{phys}}}{p_{\text{th}}}\right)^{\lfloor (d+1)/2 \rfloor}, \]

which is the usual estimate for the logical error rate per cycle below threshold [4]. The important structural point is that improving physical fidelity reduces the required \(d\), and since \(n\) grows as \(d^2\), a modest fidelity gain produces a quadratic saving in area. Fidelity is, in a real sense, a density technology.

Surface code distance
d = 25
Physical density \(\rho_A\)
1 per mm²
Physical qubits per logical qubit
1,249
at d = 25
Logical density \(\rho_L\)
8 × 10⁻⁴ per mm²
Area per logical qubit \(A_L\)
1,249 mm²
about 12.5 cm²
Logical qubits on a 300 mm wafer
56
Figure 13. Algorithm 1 as a calculator for the surface code (\(k = 1\), \(n = 2d^2 - 1\)). Pick a platform from Table 1 and Table 3 or enter your own pitch and overhead, then fix the distance or let line 4 choose it. The defaults \(A = 0.03\) and \(p_{\text{th}} = 0.57\%\) are the empirical fit of Fowler and colleagues [4], and the wafer count uses the roughly 70,000 mm² of usable area from Section 1.

12What I Take Away From This

Working through these numbers changed how I read hardware announcements.

Raw density spans eight orders of magnitude across platforms, from about one qubit per square millimetre in superconducting circuits and ion traps up to a hundred million per square millimetre for bare quantum dots. That spread is far larger than the spread in qubit counts, gate fidelities, or coherence times.

Density and connectivity trade against each other almost everywhere. Short range interactions permit tight packing but produce shallow, local coupling graphs. Long range or transported interactions produce rich graphs but force qubits apart to make room for the machinery that moves them.

Neutral atoms are the current outlier, achieving roughly \(10^4\) qubits per square millimetre while retaining reconfigurable connectivity through atom transport. That combination is why I watch that platform closely.

The binding constraint is almost never the qubit. It is the wiring, the routing, and the thermal budget. Any innovation that lowers the Rent exponent or moves control electronics into the cold is worth more to density than a smaller qubit.

Finally, the only density that matters at the end is logical density, and connectivity is the exchange rate. A platform that can implement high rate quantum low density parity check codes converts its physical qubits into logical qubits roughly an order of magnitude more efficiently than one restricted to the surface code. That is why so much effort is now going into long range couplers on chips and into atom movement in arrays. The race is not simply to pack more qubits into a millimetre. It is to pack more qubits into a millimetre and still let them reach each other.

References

  1. A. Antony, M. V. Gustafsson, G. J. Ribeill, M. Ware, A. Rajendran, L. C. G. Govia, T. A. Ohki, T. Taniguchi, K. Watanabe, J. Hone, and K. C. Fong, "Miniaturizing Transmon Qubits Using van der Waals Materials," Nano Letters 21, 10122 (2021); arXiv:2109.02824.
  2. P. Krantz, M. Kjaergaard, F. Yan, T. P. Orlando, S. Gustavsson, and W. D. Oliver, "A quantum engineer's guide to superconducting qubits," Applied Physics Reviews 6, 021318 (2019); arXiv:1904.06560.
  3. Google Quantum AI and Collaborators, "Quantum error correction below the surface code threshold," Nature 638, 920 (2025); arXiv:2408.13687.
  4. A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, "Surface codes: Towards practical large-scale quantum computation," Physical Review A 86, 032324 (2012); arXiv:1208.0928.
  5. S. Bravyi, A. W. Cross, J. M. Gambetta, D. Maslov, P. Rall, and T. J. Yoder, "High-threshold and low-overhead fault-tolerant quantum memory," Nature 627, 778 (2024); arXiv:2308.07915.
  6. R. Li, L. Petit, D. P. Franke, J. P. Dehollain, J. Helsen, M. Steudtner, N. K. Thomas, Z. R. Yoscovits, K. J. Singh, S. Wehner, L. M. K. Vandersypen, J. S. Clarke, and M. Veldhorst, "A crossbar network for silicon quantum dot qubits," Science Advances 4, eaar3960 (2018); arXiv:1711.03807.
  7. G. Burkard, T. D. Ladd, A. Pan, J. M. Nichol, and J. R. Petta, "Semiconductor spin qubits," Reviews of Modern Physics 95, 025003 (2023); arXiv:2112.08863.
  8. A. Tosato, A. Elsayed, F. Poggiali, L. E. A. Stehouwer, D. Costa, K. L. Hudson, D. Degli Esposti, and G. Scappucci, "A crossbar chip for benchmarking semiconductor spin qubits," Nature Electronics 9, 324 (2026); arXiv:2504.05460.
  9. C. D. Bruzewicz, J. Chiaverini, R. McConnell, and J. M. Sage, "Trapped-ion quantum computing: Progress and challenges," Applied Physics Reviews 6, 021314 (2019); arXiv:1904.04178.
  10. A. Ransford et al. (Quantinuum), "Helios: A 98-qubit trapped-ion quantum computer," arXiv:2511.05465 (2025); published as "A 98-qubit trapped-ion quantum computer with all-to-all connectivity," Nature 655, 81 (2026).
  11. H. J. Manetsch, G. Nomura, E. Bataille, X. Lv, K. H. Leung, and M. Endres, "A tweezer array with 6,100 highly coherent atomic qubits," Nature 647, 60 (2025); arXiv:2403.12021.
  12. D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter, J. P. Bonilla Ataides, N. Maskara, I. Cong, X. Gao, P. Sales Rodriguez, T. Karolyshyn, G. Semeghini, M. J. Gullans, M. Greiner, V. Vuletić, and M. D. Lukin, "Logical quantum processor based on reconfigurable atom arrays," Nature 626, 58 (2024); arXiv:2312.03982.
  13. H. Aghaee Rad et al. (Xanadu Quantum Technologies), "Scaling and networking a modular photonic quantum computer," Nature 638, 912 (2025).
← Back to Blog