Skip to content
Tech Interview Prep
Technical interview guide

Quantum Hardware Architectures

The competing physical implementations of a qubit (superconducting, trapped-ion, and photonic) and their tradeoffs in coherence time, gate fidelity, connectivity, and scalability.

Read
54 min
Practice MCQs
25
Interview Q&A
25
Edition
v2
Editorial status
Reviewed

Scope: IBM Quantum and Amazon Braket current hardware, metrics, calibration, and device documentation; modality provider guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-2A team builds a superconducting qubit as a plain LC resonator. Spectroscopy shows the 0→1 and 1→2 transitions both sit at 5.0 GHz, so a drive at 5.0 GHz excites both and population leaks out of the computational subspace. Which change fixes this?
MCQ-3A team runs the same two-qubit Mølmer–Sørensen gate (same laser power, same detuning, same trap axial frequency ω_z/2π = 1 MHz) on a 5-ion chain and a 20-ion chain. The 20-ion gate takes 4× longer and its fidelity drops from 99.5% to 97%. Which mechanism explains both observations?
MCQ-4A team traps single ⁸⁷Rb atoms in optical tweezers 3 µm apart and encodes each qubit in the two hyperfine ground states. They need a two-qubit controlled-phase gate with ~99% fidelity and a sub-microsecond duration. Which interaction do they use to entangle the pair?
MCQ-7A team has a 6-atom Rydberg array with global drive H(t) = (Ω(t)/2) Σ_i σ^x_i - Δ(t) Σ_i n_i + Σ_{i<j} V_ij n_i n_j, where V_ij is fixed by atom positions. They want to simulate a 6-qubit Heisenberg model H = Σ_{i<j} J_ij (σ^x_i σ^x_j + σ^y_i σ^y_j + σ^z_i σ^z_j) with independently chosen J_ij. They upload 200 segments of Ω(t) and Δ(t). The measured dynamics do not match the Heisenberg model. Which statement explains why?
MCQ-5A photonic team ports a Gaussian boson sampling experiment to a chip and reports that the device uses 12 qumodes. A reviewer asks what that count actually measures. Which statement is correct?
MCQ-6A routing team models each candidate leg as a binary variable and each pairwise conflict between legs as a penalty. They want to submit the model to a D-Wave quantum annealer. What must the submitted problem be expressed as?
MCQ-23A team benchmarks a neutral-atom array against a superconducting transmon processor on the same circuit. The transmon completes shots at roughly 10 kHz; the neutral-atom array runs at 1-10 Hz. Which stage accounts for the three-to-four-orders-of-magnitude gap in shot rate?
MCQ-16A team benchmarks two quantum processors for a 100-layer QAOA circuit. Processor A reports 2,400 CLOPS on a 20-layer QAOA circuit; Processor B reports 1,200 CLOPS on the same 20-layer circuit. Both use 1,024 shots. The team has not measured CLOPS at 100 layers. They plan to select A because it is twice as fast. What does the CLOPS comparison actually establish for the 100-layer decision?
MCQ-12A 27-qubit transmon device is benchmarked for readout with 20,000 shots prepared in |0> and 20,000 prepared in |1>: 200 of the |0> shots are classified as 1 and 1,800 of the |1> shots are classified as 0. The team confirms the assignment matrix is stable between runs and that gate and state-preparation error are negligible next to readout. A state-preparation circuit is then run for 500,000 shots and 140,000 of them are classified as 1. Which estimate of the circuit's true P(1) follows from the benchmark?
MCQ-10A team must choose one of two transmon qubits to run a circuit of 60 two-qubit (CZ) gates; single-qubit gates and readout are identical on both and can be ignored. Qubit A has T2 = 100 µs and a CZ gate of 500 ns. Qubit B has T2 = 40 µs and a CZ gate of 100 ns. Which qubit should they choose, and why?
MCQ-1A team must run a one-layer 100-qubit QAOA circuit for MaxCut on a random 3-regular graph with 150 edges. Device S is a 100-qubit superconducting processor with heavy-hex connectivity (degree ≤3), median two-qubit gate error 0.8%, and two-qubit gate time 40 ns. Device T is a 100-qubit trapped-ion processor with all-to-all connectivity, median two-qubit gate error 0.3%, and two-qubit gate time 200 µs. Assume routing on S adds one SWAP per edge on average, and each SWAP costs three two-qubit gates. Which conclusion is best supported?
MCQ-8A 30-qubit QAOA circuit contains 900 abstract CNOT gates and is written with abstract gates (RX, RZ, CNOT). It is transpiled for two backends. Backend A is superconducting with native set {RZ, SX, CX}, nearest-neighbor connectivity, and a CX error rate of 0.05%; its compiled circuit reports 1,800 CX, 300 SWAP, and estimated fidelity 0.42. Backend B is trapped-ion with native set {RZ, RX, XX}, all-to-all connectivity, and an XX error rate of 0.03%; its compiled circuit reports 900 XX, 0 SWAP, and estimated fidelity 0.78. The abstract circuit is identical. Which explanation accounts for both the gate-count and fidelity differences?
MCQ-9A five-qubit device's coupling map is a line: edges 0-1, 1-2, 2-3 and 3-4, with no others. A circuit's only gate is a CNOT between two logical qubits, and the chosen initial layout places those two logical qubits on physical qubit 0 and physical qubit 4. The transpiler routes with shortest-path SWAP insertion and is configured to keep the layout it arrives at rather than inserting further SWAPs to undo the routing. How many SWAP gates does it insert before the CNOT can execute?
MCQ-19A team ports a 40-qubit variational eigensolver from a simulator to a 127-qubit superconducting device with a heavy-hex coupling map. The compiled circuit's two-qubit gate count is 3.1x the simulator count, and the transpiler reports an average SWAP overhead of 2.4 per logical two-qubit gate. The team's first instinct is to blame the device's coherence times. Which explanation best accounts for the gate-count inflation?
MCQ-14At 09:00 a team compiles a 12-qubit circuit against a calibration snapshot: qubit Q3 has readout error 1.2% and CNOT error 0.8%, and the transpiler places the circuit's two most-used logical qubits on Q3 and Q4. At 15:00 the same circuit runs without recompilation. The afternoon calibration report shows Q3 readout error 3.5% and CNOT error 2.1%, while Q7 now has readout error 1.3% and CNOT error 0.9%. The afternoon job's measured error is several times the morning estimate. What best explains the gap?
MCQ-15A team is estimating the runtime of a quantum simulation that applies 200 layers of 8 parallel CZ gates on a 16-qubit superconducting device. They have two measurements: isolated CZ error = 0.3% per gate from interleaved randomized benchmarking on isolated pairs, and EPLG = 1.1% per gate from a cycle benchmark with 8 parallel CZs. Which error rate should they use in their resource estimate, and why?
MCQ-11A 20-qubit processor runs a workload of 50 layers, each layer applying CZ gates to 10 disjoint pairs (500 CZ gates total). Randomized benchmarking on isolated pairs gives 0.6% error per CZ, predicting 0.994^500 ≈ 5% success for the workload. With all 10 pairs driven in parallel the workload succeeds 0.3% of the time; with the same 500 CZ gates but the pairs driven one at a time, a run about ten times longer, it succeeds 4.5%. What does the 15× shortfall indicate?
MCQ-20Your team attaches a custom DRAG X-gate calibration (160 dt, amplitude 0.32, β = 0.40) to qubit 3 of a 27-qubit backend and measures 99.7% randomized-benchmarking fidelity. Three weeks later the same circuit, with the same attached calibration, runs on the same qubit and RB fidelity has fallen to 97.5%, while the backend's reported error for its default X gate on qubit 3 is unchanged. What is the most likely cause?
MCQ-21A 17-qubit transmon device runs a distance-3 repetition code whose two ancillas each measure one parity check per cycle. Control hardware on this device executes a cycle as a fixed sequence (readout, then reset, then the correction stage) with measured timings of 380 ns for ancilla readout (both ancillas read out in parallel on separate digitizers) and 220 ns for ancilla reset (also in parallel). The correction stage runs on a single shared feed-forward channel that issues one conditional Pauli per 720 ns and cannot accept a second outcome until the first one is issued. The team plans a 1.5 µs cycle and requires every conditional Pauli of a cycle to be issued before the next cycle's readout begins. What is the shortest cycle time this hardware can run?
MCQ-17When implementing dynamic quantum circuits requiring mid-circuit measurement and fast classical feed-forward, which physical hardware bottlenecks primarily constrain round-trip feedback latency in superconducting circuits versus trapped-ion architectures?
MCQ-22A trapped-ion group scales a single linear chain from 20 to 60 ions. Two-qubit gate fidelity falls from 99.5% to 98.0%, and profiling shows algorithm wall-clock time is dominated by 1 ms ion-transport steps between trap zones rather than the 50 µs two-qubit gates. Which bottleneck governs further scaling?
MCQ-18A team's 1,000-physical-qubit machine runs a distance-5 surface code that measures a logical error rate of 1e-3 per cycle. On this device the logical error rate drops by 10× for every increase of the code distance d by 2. The team wants to run as many logical qubits as possible while keeping each logical qubit's error rate at or below 1e-5 per cycle. How many logical qubits can the machine support?
MCQ-13A 5-qubit transmon processor reports a per-Clifford error of 2×10⁻³ from randomized benchmarking. A distance-3 surface-code simulation using that same physical error, with an independent Pauli channel per qubit, predicts 6×10⁻³ logical error per round; the measured value is 3×10⁻². After a 200-Clifford sequence, state tomography finds 1.2% of the population in the transmon's |2> level, and in 30% of the decays out of |2> the neighbouring qubit's state also flips. What explains the gap, and what should the team do?
MCQ-25A team must co-fabricate its qubits with classical control electronics on a standard silicon process line, and the array has to fit within a few square millimetres. They accept that entangling operations will be limited to qubits that are physically close. Which trade-off does this architecture accept?
MCQ-24A team must choose between two accelerators for an LLM serving fleet. Accelerator A wins both vendor microbenchmarks: peak FLOPs on batch-512 matrix multiplies, and streaming memory bandwidth on a 1 GB buffer. The incumbent B wins neither. In production, at batch size 1 with 2k-token prompts in fp16, B serves 30% more tokens/s than A. Which benchmark change would have predicted that result?