Inside Quantum Processors

A QPU (quantum processing unit) is the chip that holds the qubits. It never works alone. A compiler rewrites your circuit for the chip. Normal electronics turn each step into a physical signal. Readout hardware turns the qubits back into plain bits. Knowing this stack explains most of what a program can and cannot do on real hardware.

A physical qubit is one real piece of hardware. A logical qubit is one error-protected qubit, stored with backup copies spread across many physical qubits. There is no fixed exchange rate between them. The cost depends on the code, its distance, the raw error rate, and how reliable you need it to be. So a physical count and a logical count answer different questions.

A superconducting qubit is a tiny electrical circuit cooled to around 0.015 kelvin. Its two lowest energy levels act as 0 and 1. Its gates are among the fastest of any kind of qubit, taking nanoseconds. But each qubit can only talk to the neighbours it was built next to. And it keeps its quantum state for only tens to hundreds of microseconds. So the number of gates, and the extra moves needed to bring qubits together, decide most of the errors.

A trapped-ion qubit is a single charged atom held in a vacuum by electric fields. Lasers drive its gates. Two ions interact through the way the whole chain of ions wobbles together. Any pair in a chain can interact directly, and shown fidelities are the best of any platform. But the gates run roughly a thousand times slower than superconducting ones.

Neutral-atom machines hold single uncharged atoms in tightly focused laser beams, called optical tweezers. They entangle nearby atoms through the Rydberg blockade. This is a strong effect that only works over a short distance. So the layout of the atoms is itself the map of which qubits can link. Atoms can even be moved during a job. That swaps extra routing gates for time spent moving.

Photonic machines store qubits in single particles of light, called photons. For example, a qubit can be which of two paths a photon takes. They compute with beam splitters, phase shifters and detectors. Photons are fast and do not need a super-cold chip. But photons do not naturally interact, and they are easy to lose. So two-qubit logic only works some of the time, and fighting loss is most of the engineering.

A gate-model QPU runs an ordered circuit of gates and then measures. A quantum annealer is different. It physically settles toward low-energy answers of a scoring function and returns samples. The two kinds of machine give qubits completely different jobs. Annealers also spend many physical qubits on each problem variable. So raw qubit counts across the two kinds cannot be ranked.

A coupling map is a drawing of which qubit pairs can do a two-qubit gate together. Any pair not on the map must be routed with SWAPs, and each SWAP costs three CX gates. One long-distance CX on a line of five qubits becomes ten two-qubit gates. So the layout, called the topology, directly sets how long a circuit runs and how many errors it collects. It is a real limit, not just a diagram.

A native gate set is the short list of steps a device's control electronics can really do. The compiler rewrites textbook gates such as CX into these tuned steps. For example, it can build CX from a CZ with a Hadamard gate on each side. Two versions can give the same math but use different numbers of physical pulses, with different error. The two-qubit pulses cause most of the error.

A fidelity figure means nothing without its method, scope, how it was summed up, its date, and its uncertainty. Randomized benchmarking reports an average error per gate under a set test. It is not your circuit's success rate. Using the best pair of qubits instead of the worst can swing a ten-gate circuit's success from about 98% to about 78% on the same chip. Tuning drifts day by day, so numbers with no date cannot be used.

Read a QPU spec page as a record of evidence. Every number needs a source, date, method, scope, aggregation, and uncertainty. All the fields must describe the same version of the device. "Not publicly disclosed" is a clear finding with a date. It is a different state from a value that is just missing or out of date.

There is no best quantum architecture for everything. A fair comparison first fixes one job. Then it scores each architecture on four things: how well its links fit the job, what the gates compile into, timing, and dated error evidence. Example arithmetic shows a trapped-ion profile winning on success per shot (0.970 vs 0.904 on a ten-gate circuit). Meanwhile a superconducting profile runs thousands of times more shots per second. The ranking depends on the circuit and the evidence, never on one headline number.