Inside Quantum Processors

A QPU (quantum processing unit) is the chip that physically holds the qubits, and it never works alone: a compiler rewrites your circuit for the chip, classical electronics turn instructions into physical signals, and readout hardware turns quantum states back into ordinary bits. Understanding that stack explains most of what a program can and cannot do on real hardware.

A physical qubit is one actual hardware device; a logical qubit is one error-protected qubit encoded redundantly across many physical qubits. There is no universal exchange rate — the overhead depends on the code, its distance, the raw error rate, and the reliability target — so a physical count and a logical count answer different questions.

A superconducting qubit is a microscopic electrical circuit, cooled to around 0.015 kelvin, whose two lowest energy levels serve as 0 and 1. Gates are among the fastest of any platform (nanoseconds), but each qubit talks only to its fabricated neighbours and stays coherent for tens to hundreds of microseconds, so gate count and routing dominate the error budget.

A trapped-ion qubit is a single charged atom held in vacuum by electromagnetic fields, with gates driven by lasers and two-qubit interactions mediated by the ions' shared motion. Any pair in a chain can interact directly and demonstrated fidelities are the best of any platform, but gates run thousands of times slower than superconducting ones.

Neutral-atom machines hold individual uncharged atoms in focused laser beams (optical tweezers) and entangle nearby atoms through the Rydberg blockade — a strong, distance-limited interaction that makes the array's geometry itself the connectivity graph. Atoms can even be moved mid-computation, trading routing gates for shuttling time.

Photonic machines encode qubits in single particles of light — for example, which of two paths a photon travels — and compute with beam splitters, phase shifters and detectors. Photons are fast and need no deep-cryogenic chip, but they do not naturally interact and are easily lost, so two-qubit logic is probabilistic and loss dominates the engineering.

A gate-model QPU executes an ordered circuit of gates followed by measurement; a quantum annealer physically relaxes toward low-energy configurations of an energy function and returns samples. The two models give qubits entirely different jobs — and annealers spend many physical qubits per logical variable — so raw qubit counts across the two models cannot be ranked.

A coupling map is the graph of qubit pairs that can host a two-qubit gate; any pair off the graph must be routed with SWAPs, and each SWAP costs three CX gates. A single long-range CX on a five-qubit line becomes ten two-qubit gates, so topology directly sets circuit depth and error — it is an executable constraint, not a diagram.

A native gate set is the short list of operations a device's control electronics can physically execute; the compiler rewrites textbook gates such as CX into calibrated primitives, for example CZ sandwiched between Hadamards. Equivalent decompositions match mathematically but differ in physical pulse count and error, and the two-qubit pulses dominate the error budget.

A fidelity figure is meaningless without its method, scope, aggregation, date, and uncertainty: randomized benchmarking reports an average error per gate under a protocol, not your circuit's success rate, and best-pair versus worst-pair scope can swing a ten-gate circuit's success from about 98% to about 78% on the same chip. Calibration drifts day by day, so undated numbers are unusable.

Read a QPU spec page as an evidence record: every number needs a source, date, method, scope, aggregation, and uncertainty, and all fields must describe the same device revision. "Not publicly disclosed" is an explicit, dated finding — a different state from a value that is merely missing or stale.

There is no universally best quantum architecture: a defensible comparison fixes one workload, then scores each architecture on topology fit, native decomposition, timing, and dated error evidence. Illustrative arithmetic shows a trapped-ion archetype winning per-shot success (0.970 vs 0.904 on a ten-gate circuit) while a superconducting archetype runs thousands of times more shots per second — the ranking depends on the circuit and the evidence, never on one headline number.