Hardware Overview

Quantum hardware connects an abstract operation to a physical encoding, control pulse, measurement, and operating environment. The useful comparison is how well a complete device performs a specified task: an isolated lifetime, gate fidelity, or qubit count describes only part of that question. This page develops the shared language for comparing processors, simulators, sensors, memories, and network nodes without assigning a universal platform ranking.

Required background. The circuit model defines the intended operations; noise in quantum information distinguishes a target gate from its physical implementation. Helpful background. Generalized measurements supplies outcome probabilities and conditional state changes.

Physical encodings and implemented operations

A device usually has more states than its intended computational system. For a subspace encoding, let

$$V:\mathcal H_{\mathrm{comp}}\longrightarrow\mathcal H_{\mathrm{phys}}, \qquad V^\dagger V=I,\qquad P=VV^\dagger.$$

The isometry $V$ embeds logical coordinates into physical states; $P$ is the projector onto their physical subspace. An encoded qubit might use two atomic levels, two circuit eigenstates, an electron spin, two optical paths, or two oscillator codewords. Subsystem encodings additionally identify degrees of freedom whose state need not be fixed; the subspace formulas here do not describe that extra structure.

A control model can take the form

$$H(t)=H_0+\sum_k u_k(t)H_k.$$

$H_0$ is the drift Hamiltonian, and $u_k(t)$ are available classical controls. Their amplitudes, bandwidth, and simultaneous use are constrained by the device. Coupling to the environment produces a physical channel $\mathcal E_u$, not generally a unitary. A channel on the device alone also presupposes a specified environment preparation without uncontrolled initial system–environment correlations; persistent memory requires a multi-time description.

Suppose the desired operation is $U$ on the computational space. The part of the actual output remaining in that space is

$$\Phi_{\mathrm c}(\rho) =V^\dagger\mathcal E_u(V\rho V^\dagger)V, \qquad s(\rho)=\operatorname{Tr}\Phi_{\mathrm c}(\rho).$$

This is a completely positive, trace-nonincreasing map. Leakage probability is $L(\rho)=1-s(\rho)$ for normalized inputs and trace-preserving physical dynamics. The normalized surviving state $\Phi_{\mathrm c}(\rho)/s(\rho)$ is generally a nonlinear function of $\rho$; it must be reported together with its success probability. A mathematical projection does not mean the device supplies a nondestructive leakage flag. Moreover, leaked population can return later, so composing these projected maps may not reproduce a sequence of physical gates. Wood and Gambetta (2018) distinguish leakage, return, and within-subspace errors systematically.

DiVincenzo (2000) identifies characterized qubits, initialization, sufficiently preserved coherence, universal operations, and measurement as basic computing requirements, with two additional communication requirements. The National Academies (2019) develops the systems implications. A modern architecture must make these capabilities work together: preparing an ancilla, coupling it to data, reading it, and resetting it can involve different noise and timing constraints from an isolated gate experiment.

Readout probabilities and backaction

Effects $M_m\ge0$ with $\sum_m M_m=I$ describe a measurement’s probabilities:

$$p(m\mid\rho)=\operatorname{Tr}(M_m\rho).$$

For state reuse or feedback, specify the quantum instrument $\mathcal I_m$ as well:

$$\operatorname{Tr}\mathcal I_m(\rho)=p(m\mid\rho), \qquad \rho_m=\frac{\mathcal I_m(\rho)}{p(m\mid\rho)}.$$

Identical outcome probabilities do not imply identical output states. For example, $K_m=|m\rangle\langle m|$ performs a projective $Z$ measurement, whereas $L_m=|0\rangle\langle m|$ measures and resets to $|0\rangle$. Both have effects $|m\rangle\langle m|$. Only the first preserves both $Z$ eigenstates, as needed for ideal repeated nondemolition readout. The second can be useful for reset. Whether either is appropriate depends on the circuit.

A binary assignment matrix

$$A=\begin{pmatrix}p(0\mid0)&p(0\mid1)\\ p(1\mid0)&p(1\mid1)\end{pmatrix}$$

describes calibration on basis states. The relation $\mathbf f=A\mathbf p$ for arbitrary inputs holds if the effects are diagonal in that basis, or if the input has already been dephased there. Otherwise off-diagonal elements of $M_m$ make the outcome depend on coherence too. Preparing two reference states alone cannot establish that diagonal model. Assignment fidelity also does not determine backaction, leakage discrimination, preparation errors, or detector latency. Control, readout, and calibration explains how the instrument and the electronics are characterized.

Interpreting hardware metrics

Metrics for quantum hardware develops the detailed definitions and inference protocols. Three distinctions are enough to organize a first comparison.

Coherence time and circuit duration. In a stationary two-level model with independent Markovian population relaxation and pure dephasing in the energy basis,

$$\frac1{T_2}=\frac1{2T_1}+\frac1{T_\phi}.$$

$T_1$ is the population equilibration time, while $T_2$ measures transverse coherence decay in the specified experiment. Nonexponential noise and pulse-dependent filtering need a fuller description. $T_2/t_g$ is a time ratio, not a guaranteed gate count: driven control, spectators, measurement, reset, and idle intervals sample different parts of the noise.

Average gate fidelity and leakage. With normalized Haar measure and the encoded map above, the unconditional target overlap is

$$F_{\mathrm{avg}} =\int d\psi\, \langle\psi|U^\dagger \Phi_{\mathrm c}(|\psi\rangle\langle\psi|) U|\psi\rangle.$$

This definition counts leakage as lost overlap. For a trace-preserving map within the computational space it is the usual average gate fidelity (Nielsen, 2002). Dividing each surviving output by its trace would define a different, conditional quantity. Neither one number nor one input average determines worst-input error, correlated faults, or error accumulation in a long circuit. Noise and fidelity must be interpreted with the characterization protocol and simultaneous operating context.

Physical and logical performance. A resonator lifetime characterizes a component; a two-qubit gate benchmark characterizes an operation; repeated syndrome extraction tests an encoded system; a chemistry estimate tests a complete task. Logical error depends on the code, circuit, decoder, and physical noise. An illustrative odd-distance scaling fit,

$$p_{\mathrm L}(d)\approx A(p/p_{\mathrm{th}})^{(d+1)/2},$$

is not a universal function of one measured infidelity. Bias, erasure flags, coherence, leakage, and correlations can all change the result. See the surface code and resource estimation for code-specific assumptions and task costs.

Platform mechanisms and tradeoffs

The table summarizes mechanisms, not dated performance records. The linked articles develop each platform; the cited reviews explain the underlying physics. A material substrate, an encoding, and a subsystem role can overlap: for example, a superconducting resonator may hold a bosonic memory.

Platform or encoding	Physical operation	Main architectural tradeoff
Superconducting circuits	Nonlinear circuit levels, microwave drives, engineered coupling, and dispersive readout	Fast local control must coexist with material loss, leakage, frequency crowding, and cryogenic wiring (Krantz et al., 2019).
Trapped ions	Internal states coupled through collective motion; optical or microwave control	Flexible pair access within a register comes with shared-mode constraints; larger systems use multiple zones, transport, or photonic links (Bruzewicz et al., 2019).
Neutral atoms	Trapped internal states, rearrangement, and Rydberg interactions	Reconfigurable geometry and parallel zones must accommodate loading, atom loss, motion, decay, and addressing (Morgado and Whitlock, 2021).
Photonic qubits	Optical modes, interferometers, sources, detectors, and measurement-induced or nonlinear operations	Low-loss transmission supports networks; sources, indistinguishability, switching, loss, and feedforward determine usable output rates (Wang et al., 2020).
Semiconductor spins	Spin resonance, exchange, spin-to-charge conversion, and electrical detection	Small footprints require uniform confinement, low charge noise, controllable couplings, and scalable wiring (Burkard et al., 2023).
Solid-state defects	Electronic and nuclear spins with optical interfaces	Long-lived local registers need efficient, indistinguishable optical emission for useful remote links (Awschalom et al., 2018).
Bosonic encodings	Oscillator codewords with nonlinear control and syndrome extraction	One oscillator provides a large state space, but preparation, loss correction, and ancilla faults set the benefit (Joshi et al., 2021).
Continuous variables	Quadratures, Gaussian transformations, detection, and non-Gaussian resources	Finite squeezing, loss, mode matching, and resource preparation constrain a full computation (Weedbrook et al., 2012).
Topological encodings	Fusion spaces or parity sectors manipulated by braiding or measurement	Protection requires a suitable phase and controlled operations; a spectral signature alone establishes neither (Nayak et al., 2008).

Bosonic codewords must be physical, normalizable states. Ideal GKP grid states are mathematical limits; their finite-energy approximations introduce errors that must enter the architecture estimate. Gaussian processing alone does not provide universal quantum computation, even when Gaussian operations are high quality. The necessary non-Gaussian resource may enter through a state, gate, or measurement.

Topological protection is also operation-specific. Ising-anyon/Majorana braiding alone supplies a restricted gate set, so universal computation requires an additional resource. Other anyon models can support universal braiding in principle. In either case, identifying a phase, controlling an encoded state, and demonstrating protected operations are separate physical questions. Dated experimental assessments are collected in Research; a stable description of the mechanism does not imply that a scalable implementation has been demonstrated.

For microscopic derivations, see Josephson physics, ion traps, optical tweezers, Rydberg blockade, and quantized electromagnetic modes.

Connectivity, scheduling, and classical feedback

A graph of native two-qubit interactions says which pairs can couple. It does not say which edges can run together. Shared buses, lasers, couplers, spectral crowding, and readout resources impose scheduling constraints. Pairwise conflicts can be represented by a conflict graph; some power or bandwidth limits constrain whole groups of operations and need more than pairwise edges.

For example, a CNOT between the endpoints of a four-qubit line $a-q_1-q_2-b$ can be implemented by moving the state at $a$ through two SWAPs, applying the gate, and restoring the mapping. With three CNOTs per SWAP, that construction costs thirteen CNOTs. Keeping the new mapping costs seven. Neither is a universal optimum: a different native entangler, measurement-assisted routing, or motion of carriers can change the cost. These are operations on physical computational qubits; encoded logical gates require their own fault-tolerant construction.

For strictly serial preparation, control, measurement, decoding, and feedback,

$$t_{\mathrm{serial}} =t_{\mathrm{prepare}}+t_{\mathrm{control}}+t_{\mathrm{measure}} +t_{\mathrm{decode}}+t_{\mathrm{feedback}}.$$

The execution time of a general schedule follows its dependencies and shared resources. Some decoding can overlap measurement when the required records arrive early enough. Many correction rounds can instead proceed with a tracked Pauli frame while decoding is pipelined. Then distinguish the syndrome-round period, decoder throughput, and latency until an adaptive operation needs the result. A decoder must avoid an ever-growing backlog; it need not always finish each round before the next one starts.

The quantum software stack connects this device description to compilation and execution. Reilly (2015) explains why the quantum–classical interface is a physical scaling constraint.

Infrastructure and inter-module links

Fabrication or assembly changes the available interaction graph through missing sites, parameter variation, and imperfect interfaces. Control fan-out consumes bandwidth, wiring, optical access, and power; multiplexing can introduce crosstalk. Calibration must track drift without invalidating operations that share controls. These mechanisms, rather than fabrication heritage alone, determine scalability. See materials and fabrication and cryogenic and vacuum infrastructure.

A quantum memory must accept, preserve, and return the intended states. Lifetime alone omits write/read efficiency, conditional fidelity, bandwidth, mode capacity, background, and duty cycle. Interconnects and transduction adds mode matching, carrier conversion, noise, and transmission loss. Conversion efficiency by itself cannot establish preservation of quantum correlations (Kurizki et al., 2015).

For independent link attempts with fixed success probability $p$, the number $N$ of attempts through the first success is geometric:

$$\mathbb E[N]=\frac1p,\qquad \Pr(N\le n)=1-(1-p)^n.$$

With one completed attempt every $\tau$, mean generation time is $\tau/p$. That mean is not a deadline guarantee. Memory decay during failed attempts, reset, acknowledgment, multiplexing, and correlated failures change the service supplied to a modular architecture. Compare sustained throughput and task error alongside best-component performance, using the same operating conditions and including rejected attempts.

Exercises

1. Leakage and a conditional state

A physical qutrit implements $W|0\rangle=|0\rangle$ and $W|1\rangle=\sqrt{1-\ell}|1\rangle+\sqrt\ell|2\rangle$, where $0<\ell<1$. The qubit is the span of $|0\rangle,|1\rangle$. Find the computational branch and show why normalizing it is nonlinear.

Solution

The projected Kraus operator is $K=\operatorname{diag}(1,\sqrt{1-\ell})$, so $\Phi_{\mathrm c}(\rho)=K\rho K^\dagger$ and $s(\rho)=1-\ell\rho_{11}$. Conditional outputs of the two basis states are unchanged. But the equal mixture has conditional output

$$\frac{\Phi_{\mathrm c}(I/2)}{s(I/2)} =\frac{1}{2-\ell} \begin{pmatrix}1&0\\0&1-\ell\end{pmatrix},$$

which is not their equal mixture. Surviving trials favor the state with higher survival probability. Leakage also changes coherent superpositions; knowing the lost population does not determine all within-subspace errors.

2. Identical probabilities, different repeated readout

For the two instruments $K_m$ and $L_m$ above, start with $|1\rangle$. Compare the first result and an immediate second $Z$ measurement.

Solution

Both first measurements report 1 with probability 1. The $K_m$ instrument leaves $|1\rangle$, so the next result is 1. The $L_m$ instrument resets to $|0\rangle$, so the next result is 0. Equal assignment fidelity can therefore coexist with entirely different state reuse. Nondemolition performance needs state-update or repeated-measurement information, beyond the single-shot POVM probabilities.

3. Classical assignment and a coherent exception

For a diagonal readout model,

$$A=\begin{pmatrix}0.98&0.07\\0.02&0.93\end{pmatrix}, \qquad \mathbf f=\begin{pmatrix}0.616\\0.384\end{pmatrix}.$$

Recover the populations. Then consider instead the effects $M_0=(I+\eta X)/2$, $M_1=I-M_0$ with $0<\eta\le1$. Can their basis-state assignment matrix distinguish $|+\rangle$ and $|-\rangle$?

Solution

For the first model $\det A=0.91$, and $A^{-1}\mathbf f=(0.60,0.40)^T$. For the second, all entries of the assignment matrix are $1/2$. Both $X$ eigenstates also have equal $Z$ populations. Nevertheless,

$$p(0\mid\pm)=\frac{1\pm\eta}{2}.$$

The basis calibration misses a coherence-sensitive measurement. Inverting a classical assignment matrix is justified only after its model is established; it cannot correct arbitrary detector behavior.

4. Routing and restoration

For the four-qubit line described above, count the CNOTs with and without restoring the original state-to-site mapping. Does the cheaper construction produce the same final physical state at each site?

Solution

Restoration uses four SWAPs and one desired CNOT: $4(3)+1=13$. Without restoration the cost is $2(3)+1=7$. The logical action is the same when the compiler tracks the permutation, but the state-to-site mapping differs. Subsequent gates and measurements must use that new mapping.

5. Serial latency and streaming decoding

Preparation, gates, measurement, decoding, and feedback take $1,4,3,2,1\,\mu\mathrm{s}$ respectively. Find the serial time. Under what condition could a complete decoding overlap reduce it to 9 microseconds?

Solution

The serial time is $11\,\mu\mathrm{s}$. If the needed data arrive early enough that all decoding finishes by the end of measurement, the elapsed time is $1+4+3+1=9\,\mu\mathrm{s}$. Knowing only the two durations is insufficient: a two-microsecond calculation that requires the final measurement bit cannot finish during that measurement. Streaming, dependencies, and resource contention determine the possible overlap.

6. Why an erasure flag helps

A distance-$d$ code must correct $t$ arbitrary unlocated qubit errors and $s$ known erased qubits. State a sufficient worst-case distance condition, and explain why imperfect flags alter the comparison.

Solution

The standard distance condition is $2t+s<d$ (Nielsen and Chuang, 2010): distinguishing two candidate unlocated errors can involve $2t$ sites, while the erased locations are already known. A reliable flag therefore supplies useful information. False or missed flags and correlations must be included in the actual decoder model. The inequality concerns a fixed code’s correctability; it is not a universal circuit-level threshold or a guarantee about every decoder.

7. Mean waiting time versus a deadline

An independent link attempt completes every $5\,\mu\mathrm{s}$ with success probability $p=0.002$. Find the mean time and the smallest number of attempts giving at least 95% probability of success.

Solution

The mean is 500 attempts, or 2.5 milliseconds. For a 95% success probability,

$$n_{95}=\left\lceil\frac{\ln0.05}{\ln(1-p)}\right\rceil =1497,$$

giving 7.485 milliseconds. At 500 attempts the success probability is only $1-0.998^{500}\simeq0.6325$. Thus comparing memory lifetime only with the mean wait can hide a substantial long-wait tail.

8. Finite-distance logical-error data

Logical failure estimates at distances 3, 5, and 7 are respectively $2.0\times10^{-3}$, $8.0\times10^{-4}$, and $1.1\times10^{-3}$. Do they establish monotonic suppression through distance 7?

Solution

The point estimates are not monotonic: the last increases. Without error bars they establish neither a statistically significant reversal nor monotonic suppression of the underlying probabilities. Compare the same noise, circuit family, and decoder conditions; finite-size effects, leakage, correlations, or changed hardware can alter the trend. One improved distance point does not establish asymptotic below-threshold scaling.

References

• D. D. Awschalom, R. Hanson, J. Wrachtrup, and B. B. Zhou, “Quantum technologies with optically interfaced solid-state spins,” Nature Photonics 12, 516–527 (2018), doi:10.1038/s41566-018-0232-2.
• C. D. Bruzewicz, J. Chiaverini, R. McConnell, and J. M. Sage, “Trapped-ion quantum computing: Progress and challenges,” Applied Physics Reviews 6, 021314 (2019), doi:10.1063/1.5088164.
• G. Burkard, T. D. Ladd, A. Pan, J. M. Nichol, and J. R. Petta, “Semiconductor spin qubits,” Reviews of Modern Physics 95, 025003 (2023), doi:10.1103/RevModPhys.95.025003.
• D. P. DiVincenzo, “The physical implementation of quantum computation,” Fortschritte der Physik 48, 771–783 (2000), doi:10.1002/1521-3978(200009)48:9/11<771::AID-PROP771>3.0.CO;2-E.
• A. Joshi, K. Noh, and Y. Y. Gao, “Quantum information processing with bosonic qubits in circuit QED,” Quantum Science and Technology 6, 033001 (2021), doi:10.1088/2058-9565/abe989.
• P. Krantz et al., “A quantum engineer’s guide to superconducting qubits,” Applied Physics Reviews 6, 021318 (2019), doi:10.1063/1.5089550.
• G. Kurizki et al., “Quantum technologies with hybrid systems,” Proceedings of the National Academy of Sciences 112, 3866–3873 (2015), doi:10.1073/pnas.1419326112.
• M. Morgado and S. Whitlock, “Quantum simulation and computing with Rydberg-interacting qubits,” AVS Quantum Science 3, 023501 (2021), doi:10.1116/5.0036562.
• National Academies of Sciences, Engineering, and Medicine, Quantum Computing: Progress and Prospects (National Academies Press, 2019), doi:10.17226/25196.
• C. Nayak, S. H. Simon, A. Stern, M. Freedman, and S. Das Sarma, “Non-Abelian anyons and topological quantum computation,” Reviews of Modern Physics 80, 1083–1159 (2008), doi:10.1103/RevModPhys.80.1083.
• Michael A. Nielsen, “A simple formula for the average gate fidelity of a quantum dynamical operation,” Physics Letters A 303, 249–252 (2002), doi:10.1016/S0375-9601(02)01272-0.
• Michael A. Nielsen and Isaac L. Chuang, Quantum Computation and Quantum Information, 10th anniversary edition (Cambridge University Press, 2010).
• D. J. Reilly, “Engineering the quantum-classical interface of solid-state qubits,” npj Quantum Information 1, 15011 (2015), doi:10.1038/npjqi.2015.11.
• J. Wang, F. Sciarrino, A. Laing, and M. G. Thompson, “Integrated photonic quantum technologies,” Nature Photonics 14, 273–284 (2020), doi:10.1038/s41566-019-0532-1.
• C. Weedbrook et al., “Gaussian quantum information,” Reviews of Modern Physics 84, 621–669 (2012), doi:10.1103/RevModPhys.84.621.
• Christopher J. Wood and Jay M. Gambetta, “Quantification and characterization of leakage errors,” Physical Review A 97, 032306 (2018), doi:10.1103/PhysRevA.97.032306.