In short: Two-qubit entangling gates (CZ, iSWAP, cross-resonance) are the single hardest engineering problem in scaling quantum computers, because their fidelity — today around 99.5–99.9% on the best superconducting and trapped-ion hardware — sets the floor for whether error correction helps or hurts.
Any quantum algorithm worth running needs entanglement, and entanglement only comes from two-qubit gates. You can rotate a single qubit all day with near-perfect microwave or laser pulses, but a circuit built only from single-qubit rotations is classically simulable and useless. The moment two qubits must interact and become correlated, every physical imperfection in the device — stray couplings, calibration drift, control crosstalk, decoherence — shows up at once. This is why two-qubit gate fidelity, not qubit count, is the number that experienced architects watch.
Why one qubit is easy and two are brutal
A single-qubit gate is a rotation on the Bloch sphere. Its worst enemies are amplitude and phase errors in one control line, and these are routinely calibrated to error rates below 0.05% (fidelity > 99.95%). A two-qubit gate is fundamentally different: it must turn on a controlled interaction between two quantum systems for a precise duration and then turn it off cleanly. The interaction Hamiltonian is never perfectly isolated — neighbouring qubits, higher energy levels (the leakage problem), and the control hardware itself all leak into the dynamics.
A useful mental model: the two-qubit gate error budget is roughly the sum of a coherent part (miscalibration, residual ZZ coupling, leakage to the |2⟩ state) and an incoherent part (T₁/T₂ decoherence during the ~20–300 ns the gate is on). On superconducting chips the gate is deliberately kept short precisely to minimise the decoherence term.
The three workhorse gates
Different hardware platforms get their native entangling gate from whatever physical coupling they can control. Three dominate the field:
- CZ (controlled-Z) — the native gate of tunable-frequency superconducting qubits (Google Sycamore, many IBM and academic devices). You briefly tune two qubits into resonance so the |11⟩ state picks up a conditional phase of exactly π. Fast (~20–40 ns) and high-fidelity, but requires flux-tunable qubits and careful avoidance of the |02⟩ leakage level.
- iSWAP / √iSWAP — arises naturally from the exchange (XY) coupling between qubits; it swaps excitations and adds a phase. Google's quantum-supremacy experiment used a tunable fSim family that interpolates between iSWAP and CZ. Attractive because the √iSWAP is a perfect entangler reachable in a single pulse.
- Cross-resonance (CR) — IBM's choice for fixed-frequency transmons. You drive the control qubit at the target qubit's frequency; the always-on coupling converts this into an effective ZX interaction that builds toward a CNOT. No tunable elements means better coherence and reproducibility, at the cost of longer gates (~160–300 ns) and the need to echo away unwanted terms.
Fidelity: the number that decides everything
Fidelity is measured with randomized benchmarking (and increasingly cross-entropy benchmarking), which averages gate error over random circuits so that single-qubit errors and SPAM (state-preparation-and-measurement) errors are separated out. State-of-the-art results as of the mid-2020s:
| Platform | Native gate | Typical 2-qubit fidelity | Gate time |
|---|---|---|---|
| Trapped ions (Ca⁺/Yb⁺) | Mølmer–Sørensen (XX) | 99.9%+ | 10–100 µs |
| Tunable superconducting | CZ / fSim | 99.5–99.8% | 20–40 ns |
| Fixed-freq superconducting | Cross-resonance | 99–99.7% | 160–300 ns |
| Neutral atoms (Rydberg) | CZ (blockade) | 99.5% | ~0.3–1 µs |
Why 99.9% matters so much: the surface code's error-correction threshold is roughly 1% per gate, but to get a useful logical error rate without astronomical qubit overhead you want physical errors well below that — often cited near 0.1%. Each factor-of-ten improvement in two-qubit error can cut the number of physical qubits per logical qubit dramatically.
Where the errors actually come from
In practice an architect chasing the last 0.2% of fidelity fights a short, stubborn list:
- Residual ZZ coupling: an always-on static interaction that dephases idle qubits and distorts gates. Tunable couplers (now standard on advanced superconducting chips) exist largely to zero this out.
- Leakage: transmons are weakly anharmonic, so population escapes to the |2⟩ state. DRAG pulse shaping and careful frequency choices suppress it.
- Crosstalk: driving one gate perturbs spectator qubits. This scales badly and is a core reason scaling is hard, not just "more of the same".
- Calibration drift: qubit frequencies wander, so gates must be re-tuned continuously with automated calibration stacks.
The toolchain here mirrors classical chip design. Just as VLSI engineers lean on Synopsys, Cadence, and Siemens EDA for timing closure and parasitic extraction, quantum control teams use pulse-level frameworks (Qiskit Pulse, Google's Cirq, and optimal-control solvers) to shape the microwave envelopes that implement these gates. The spirit is identical: a physically correct model, closed-loop optimisation, and relentless verification. Teams building the surrounding classical control ASICs and cryo-electronics often partner with silicon design houses such as Avecas for custom VLSI design services.
Why this is the bottleneck to scaling
Here is the uncomfortable arithmetic. A fault-tolerant algorithm may need billions of logical operations. If every logical gate rests on thousands of physical two-qubit gates, and each physical gate has even a 0.5% error, the whole machine is noise unless error correction is working with margin to spare. Scaling to more qubits does not help if adding qubits also adds crosstalk and calibration burden — which it does. That is the real reason the industry obsesses over two-qubit fidelity rather than raw qubit count: it is the exponent, not the coefficient. If you want to see how these entangling gates compose into the full universal gate set, read our companion piece on CNOT and universal gate sets, and for the error-correction side see surface-code error correction.
Want to go from the Bloch sphere to real pulse-level gate calibration with hands-on labs? Explore the CourseTron Quantum Computing course track and build intuition the way working engineers do. For complementary semiconductor and VLSI foundations, our partner ChipXpert VLSI training covers the classical side.
Frequently asked questions
Why is a two-qubit gate harder than a single-qubit gate?
A single-qubit gate controls one isolated system, so errors come from one control line. A two-qubit gate must switch a controlled interaction on and off precisely, and every residual coupling, leakage level, crosstalk path, and decoherence channel contributes error simultaneously — often an order of magnitude more than single-qubit gates.
Which is better: CZ, iSWAP, or cross-resonance?
None is universally best — each is the natural gate for a specific hardware coupling. CZ and fSim suit tunable superconducting qubits and are very fast; cross-resonance suits fixed-frequency transmons and buys better coherence at the cost of speed; iSWAP arises from exchange coupling. The right choice depends on the physical platform.
What two-qubit fidelity is good enough for error correction?
The surface-code threshold is roughly 1% per gate, but practical fault tolerance with reasonable qubit overhead wants physical two-qubit error near 0.1% (99.9% fidelity). The best trapped-ion and superconducting devices are now at or approaching that figure.
How is two-qubit gate fidelity measured?
Primarily by randomized benchmarking, which runs long random Clifford sequences and fits the decay of the survival probability, isolating gate error from state-prep and measurement errors. Cross-entropy benchmarking and gate-set tomography give complementary, more detailed views.