The qubit architecture — superconducting transmon, heavy-hex topology

Why transmon, and why now

The transmon qubit is the workhorse of nearly every superconducting quantum effort that has crossed the hundred-qubit threshold. The reason is unromantic: the transmon is a charge qubit that has been deliberately detuned to be insensitive to charge noise. You take a Cooper-pair box, shunt the Josephson junction with a large capacitor, and push the ratio of Josephson energy to charging energy (E_J/E_C) up into the 50–100 range. The energy levels flatten with respect to offset charge, and the dominant decoherence channel — 1/f charge noise on the substrate — drops away. You pay for that with reduced anharmonicity, typically around 200–300 MHz, which sets a hard floor on how fast you can drive single-qubit gates without leaking into the |2⟩ state.

That trade is the right one for a 100-physical-qubit machine in 2026–2027. Transmons are the only superconducting qubit modality with a credible, demonstrated path through the surface code. They fabricate cleanly on silicon or sapphire with aluminium and aluminium-oxide junctions. They couple to coplanar-waveguide resonators for dispersive readout. The control electronics, room-temperature to mixing-chamber wiring, and SDK toolchains are all mature. For a sovereign Irish machine that has to deliver real chemistry workloads to real climate researchers — not just look impressive on a poster — picking the modality with the deepest engineering substrate is the only defensible choice.

Heavy-hex topology and why it beats square lattice for this scale

The Ireland Quantum 100 architecture uses a heavy-hex coupling graph — a hexagonal lattice where each vertex qubit has degree two or three, with additional qubits sitting on the edges acting as syndrome ancillas. This is the topology IBM moved to in 2021 and has stayed with for good reason. A square lattice gives every data qubit four nearest neighbours, which sounds better until you sit down with the frequency-collision arithmetic.

Transmon two-qubit gates — whether you implement cross-resonance, CZ via tunable couplers, or echoed cross-resonance — require careful frequency placement. With degree-four connectivity you have a combinatorial frequency-allocation problem: every qubit must avoid its neighbours' transitions, its neighbours' |1⟩→|2⟩ transitions, and the difference and sum frequencies that drive unwanted ZZ crosstalk. As you scale, fabrication variance in junction critical current — typically 2–3% one-sigma even with good process control — turns frequency collision into the dominant yield-killer.

Heavy-hex caps the worst-case degree at three. The frequency-collision constraint count drops dramatically, the spectator-qubit error during a CNOT drops with it, and the crosstalk you have to characterise during calibration is a tractable graph problem rather than an exponential one. The cost is that the surface code becomes a heavy-hexagonal code with slightly worse threshold (~0.45% versus ~1% for the standard surface code on a square lattice). For a hundred-qubit machine that is going to spend most of its first year running NISQ-era variational algorithms anyway, that threshold gap is not the binding constraint. Calibration stability is.

The cryogenic stack

The processor sits at the mixing-chamber stage of a dilution refrigerator running below 15 mK — well under the 100–200 mK range where thermal photons in the readout resonators start polluting state preparation. The fridge has the standard four intermediate stages: 50 K and 4 K supplied by the pulse tube, then still (~700 mK), cold plate (~100 mK), and the mixing chamber.

Each stage is an attenuation budget problem. Drive lines carrying microwave control pulses come down through a chain of attenuators — typically 20 dB at 4 K, 10 dB at the still, 20 dB at the mixing chamber — to thermalise the noise photons travelling down from room temperature. Output lines run the other way through circulators and through Josephson parametric amplifiers (JPAs) or travelling-wave parametric amplifiers (TWPAs) at the mixing chamber, then high-electron-mobility transistor amplifiers at 4 K. Every additional cable into the fridge is heat load on a stage that has, at best, a few hundred microwatts of cooling power.

This is why scaling beyond a few hundred qubits in a single can is hard, and why we are sizing the fridge generously for the 100-qubit chip rather than trying to push it. The Q4 2026 cryostat install is the longest-lead item in the build.

Control electronics and the classical stack

Each qubit needs an XY drive line for single-qubit rotations, the chip needs flux-bias lines for tunable couplers, and each readout resonator group needs a frequency-multiplexed input/output pair. The room-temperature side is built around RF arbitrary waveform generators in the 4–8 GHz band with sub-nanosecond timing alignment across the system, plus heterodyne digitisers on the readout side. The classical control loop — measure ancilla, decide, condition next pulse — has to close in under a few microseconds to be useful for any future error-correction work.

The SDK surface is deliberately conventional: OpenQASM 3 as the lingua franca, with Qiskit, PennyLane, and Cirq all targeting the backend through a transpiler that knows the heavy-hex coupling map and the native gate set (parametric single-qubit rotations plus a calibrated two-qubit entangler — most likely echoed cross-resonance for the first generation). Researchers should not have to learn a bespoke language to run a VQE on this machine.

Coherence, gates, and what "100 physical qubits" actually means

The realistic targets for first-light in Q1 2027 are T_1 and T_2 in the 100–200 microsecond range, single-qubit gate fidelities above 99.9%, and two-qubit gate fidelities in the 99.0–99.5% band. These are not record-setting numbers — they are numbers consistent with what well-engineered transmon systems have demonstrated in academic and industrial labs over the last three years. We will publish the actual measured numbers when we have them, on the actual chip, with the actual calibration.

One hundred physical qubits is not one hundred logical qubits. It is enough to run meaningful variational quantum eigensolver and quantum approximate optimisation algorithm workloads on chemistry Hamiltonians of genuine climate relevance — small carbon-capture amine systems, perovskite unit cells, lithium-ion cathode fragments — at a problem size that classical methods can verify but