Source-linked AI summary
A quantum-classical cloud platform optimized for variational hybrid algorithms
Peter J. Karalekas, Nikolas A. Tezak, Eric C. Peterson, Colm A. Ryan, Marcus P. da Silva, Robert S. Smith
TL;DR
Traditional quantum-cloud queues add network and shared-queue overhead, while variational hybrid algorithms require tight coupling between classical and quantum resources. The paper benchmarks runtime bottlenecks and implements parametric compilation and active reset in Rigetti QCS. Aspen-4 measurements report lower runtimes with QCS colocation and a further small-shot reduction with parametric compilation.
Problem
Shared-queue quantum-cloud services add network overhead, while variational hybrid algorithms need tight classical–quantum coupling for near-term quantum computing.
Source: §1 · p. 2Method
Results
T_V fell from 1.0 s to 410 ms and T_Q from 270 µs to 110 µs with QCS colocation on Aspen-4; parametric compilation reduced small-shot T_V to 36 ms.
Source: §5 · p. 12Takeaways & Limitations
Quantum-cloud platform performance should account for the latencies governing access and application runtimes, alongside error rates and qubit count.
Source: §6 · p. 12
Abstract
from arXiv · showhide
In order to support near-term applications of quantum computing, a new compute paradigm has emerged--the quantum-classical cloud--in which quantum computers (QPUs) work in tandem with classical computers (CPUs) via a shared cloud infrastructure. In this work, we enumerate the architectural requirements of a quantum-classical cloud platform, and present a framework for benchmarking its runtime performance. In addition, we walk through two platform-level enhancements, parametric compilation and active qubit reset, that specifically optimize a quantum-classical architecture to support variational hybrid algorithms (VHAs), the most promising applications of near-term quantum hardware. Finally, we show that integrating these two features into the Rigetti Quantum Cloud Services (QCS) platform results in considerable improvements to the latencies that govern algorithm runtime.
1. Introduction
Quantum-cloud platforms extend cloud access to scarce quantum hardware, but shared queues and generic web APIs add runtime overhead and fail to exploit quantum-specific execution. Because variational hybrid algorithms need tight classical–quantum coupling, the paper examines platform bottlenecks and benchmarks latency reductions from QCS enhancements.
1. Introduction
- Quantum-cloud data centers extend Internet-accessible infrastructure to quantum devices to provide access to scarce hardware.The infrastructure includes specialized equipment such as dilution refrigerators and microwave-control racks.Source: §1 · p. 2
- Hybrid cloud queues add network and shared-queue overhead, while traditional web APIs do not adapt to quantum-device computation.The approach provided worldwide public access to quantum resources.Source: §1 · p. 2
- Variational hybrid algorithms use a QPU alongside a classical computer, but require tight coupling that cloud-hosted queues do not provide at superconducting-operation timescales.The paper identifies VHAs as a promising way to use near-term devices across optimization, chemistry, and machine-learning applications.Source: §1 · p. 2
- The paper benchmarks quantum-classical cloud runtime and measures latency reductions from specialized compilation and qubit-reset techniques in Rigetti QCS.It frames the work against the need to improve device error rates and ultimately reach fault tolerance while pursuing applications on current NISQ processors.Source: §1 · pp. 2–3
2. Runtime bottlenecks in the quantum-classical cloud
A quantum-classical cloud combines quantum hardware, control, execution, compilation, and classical compute resources. For variational algorithms, latency arises at both the iteration and shot levels; compilation and passive qubit reset are the largest improvement targets in the described QCS budgets.
2. Runtime bottlenecks in the quantum-classical cloud
- A quantum-classical cloud needs a qubit apparatus, control system, executor, compiler, and access to classical compute.In Rigetti QCS, users work from a data-center QMI; programs are compiled, sent to an executor, loaded onto AWGs, and run on the Aspen-4 processor.Source: §2 · p. 3
- For superconducting-qubit VHAs, colocating classical compute with the QPU enables a low-latency link; QCS places its QMI inside the Berkeley data center.QCS compiles Quil programs to pulse-level instructions, and its Aspen-4 QPU uses Rigetti AWGs to manipulate and read out qubits.Source: §2 · pp. 3–4, Figure 1 · p. 4
- Each VHA iteration step includes QPU communication, result collection, and optimizer work, while each shot runs one quantum program and returns one bitstring.The latency budget separates costs that occur once per variational iteration from those that occur once per quantum shot.Source: §2 · pp. 4–5
- Compilation is the largest per-step latency contributor by at least one order of magnitude, while passive qubit reset already dominates per-shot latency.The paper identifies compilation and reset as the budget components whose improvement would have the largest impact on algorithm runtime.Source: §2 · p. 5, Table 1 · p. 5
3. Optimizing for the variational execution model
The platform enhancements target compilation between VHA iterations and reset between quantum shots. Parametric compilation defers variational parameters to runtime, while active reset uses measurement feedback to reduce reset time on Aspen-4.
3. Optimizing for the variational execution model
- The enhancements address compilation and reset, which the paper identifies as latency bottlenecks for the variational execution model.For Aspen-4, the active-reset protocol uses three feedback rounds to obtain performance comparable to passive reset.Source: §3 · p. 5, §3.2 · p. 8
3.1. Parametric compilation
- Parametric compilation stores gate parameters in runtime-updatable data memory, allowing binaries to be repatched between VHA iterations without recompilation.Quil DECLARE registers provide the data-memory layout, and the executor fills it from the current iteration’s memory map.Source: §3.1 · p. 6, Figure 2 · p. 7
3.2. Active qubit reset
- Active reset uses measurement results to branch between an idle path and an X-gate feedback path, repeating for the requested number of rounds.The compiler translates RESET into per-qubit control-flow graphs that run before the main circuit.Source: §3.2 · pp. 7–8
- On Aspen-4, active reset takes 9–12 µs, approximately a tenfold improvement over passive reset.Three feedback rounds achieve reset performance comparable to passive reset; the sequence must finish before the main program begins.Source: §3.2 · pp. 8–9
4. A volumetric framework for benchmarking runtime
The framework measures QPU latency across qubit counts, shot counts, and program families, then separates it into variational-step and per-shot components. It uses random phase gadget circuits at a quantum-volume-based qubit count to benchmark workloads relevant to variational hybrid algorithms.
4. A volumetric framework for benchmarking runtime
- QPU latency is the CPU wait between requesting quantum-device execution and receiving results, and depends on qubit count, shots, and program family.Sweeping shot count for fixed qubit count and program family estimates variational-step latency from the single-shot limit and quantum-shot latency from large-shot scaling.Source: §4 · p. 9
- The model decomposes total latency as T(m, n, P) = TV(m, P) + nTQ(m, P), separating step latency from per-shot latency.A linear fit of runtime against shot count estimates TV and TQ.Source: §4 · pp. 9–10
- Simple one-qubit measurement benchmarks miss scaling with qubit count and program complexity, while maximum-device-size benchmarks can be misleading when errors prevent useful full-QPU entanglement.The proposed benchmark instead uses log2 VQ as a near-term estimate of qubits meaningfully usable in a computation.Source: §4 · p. 9
- Random phase gadget (RPG) circuits replace quantum-volume random two-qubit unitaries with randomly parameterized RZ phase gadgets, while retaining permutations and Hadamard layers.The benchmark sets m = d = log2 VQ; Figure 4 notes this choice is arbitrary and intended to simplify the benchmark.Source: §4 · p. 10, Figure 4 · p. 10
- For the optimal platform configuration, the critical shot number separating TV-dominated from TQ-dominated latency is 1,700 shots.Figure 5 reports this crossover for the configuration with parametric compilation and active qubit reset enabled.Source: Figure 5 · p. 11
5. QPU latency results on Quantum Cloud Services
On Aspen-4, QCS’s colocated architecture, parametric compilation, and active qubit reset reduce the benchmarked latency components relative to the first-generation access model. The reported best configuration achieves greater than 27× and 12× improvements in variational-step and quantum-shot latency, respectively.
5. QPU latency results on Quantum Cloud Services
- The comparison favors Forest Web API because its external-user programs averaged 3 qubits, 2 CZ gates, and 14 RX gates, fewer gates than compiled RPG(3) programs.The authors consider the comparison reasonable because Aspen-4 has log2 VQ = 3, while noting the gate-count difference.Source: §5 · pp. 11–12
- TV = 36 ms and TQ = 21 µs in the optimal QCS configuration, improving by greater than 27× and 12×, respectively, over first-generation access.The optimal configuration combines parametric compilation and active qubit reset.Source: §5 · p. 12
- Colocation reduced Aspen-4 latency from Forest Web API values of TV = 1.0 s and TQ = 270 µs to QCS values of TV = 410 ms and TQ = 110 µs.With parametric compilation, TV fell further to 36 ms; active reset then reduced TQ to 21 µs.Source: §5 · p. 12
6. Conclusions
The conclusion argues that cloud-access latency and application runtime should be considered alongside error rates and qubit count as quantum hardware becomes available through cloud platforms.
6. Conclusions
- Colocation, parametric compilation, and active qubit reset considerably improve latency over first-generation quantum-cloud offerings, though they are only some of many potential platform optimizations.The authors identify cloud-access latency and resulting application runtimes as critical platform-performance considerations alongside error rates and qubit count.Source: §6 · p. 12
Appendix A. Implementing hybrid algorithms with parametric compilation
Parametric compilation can package the parameterized ansatz and measurement operations of a variational hybrid algorithm into one runtime-patched program.
Appendix A. Implementing hybrid algorithms with parametric compilation
- A VHA can be expressed as one parametric binary with runtime-patched RZ arguments, encapsulating its ansatz and arbitrary-basis measurement.The same program can also encapsulate readout-error symmetrization.Source: Appendix A · p. 13
Appendix A.1. Readout error symmetrization and mitigation
Readout symmetrization balances measurement errors by averaging the error rates for reported 0s and 1s. Calibrating the symmetrized readout then permits rescaling Pauli expectations to remove measurement bias, though this increases uncertainty and is not scalable.
Appendix A.1. Readout error symmetrization and mitigation
- The Max-Cut QAOA Experiment uses one ZZ measurement setting and compares symmetrized data with data rescaled using pyQuil calibration.The data sweep 100 γ values on Aspen-4; the symmetrized data are not corrected until the rescaling step.Source: Figure A1 · p. 14
- Flipping bits before measurement and flipping the outcomes afterward makes the readout errors symmetric, replacing the two error rates with their arithmetic mean.The procedure corresponds to a simple form of twirling and can be implemented by randomly choosing which bits to flip across measurements.Source: Appendix A.1 · p. 14
- Calibration measures the symmetrized expectation of Z-observable tensor products on the ground state, and mitigation divides later expectations by the corresponding λ to remove measurement bias.Different tensor products may require different λ values because of qubit-dependent signal-to-noise ratios and correlated readout errors.Source: Appendix A.1 · p. 15
- Rescaling removes bias but increases uncertainty because |λ| < 1, making this mitigation procedure non-scalable despite its effectiveness for small qubit counts.The appendix notes that more sophisticated symmetrization using orthogonal arrays is beyond its scope.Source: Appendix A.1 · pp. 14–15
- pyQuil’s Experiment framework packages the circuit program and measurement settings while specifying shots, active reset, and readout symmetrization and calibration.The appendix uses it to implement readout symmetrization and correction for a two-qubit Max-Cut QAOA program.Source: Appendix A.1 · p. 15
Appendix A.2. Bell state tomography
Single-qubit rotations before z-basis measurement enable measurements in other bases, and runtime-adjustable rotation parameters let one binary run the basis settings needed for Bell-state tomography. The readout-corrected reconstruction yields a Bell-state fidelity of 99.35%.
Appendix A.2. Bell state tomography
- Single-qubit rotations before z-basis measurement allow the QPU to measure in a different basis.An arbitrary single-qubit gate can be decomposed into an Euler-angle sequence.Source: Appendix A.2 · p. 15
- 99.35% Bell state fidelity is calculated by comparing the ideal and readout-corrected estimated density matrices.The estimate uses linear inversion on experimental data, and the comparison uses the fidelity function from forest-benchmarking.Source: Appendix A.2 · p. 16
- Runtime-adjustable Euler-decomposed rotations before measurement let one parametric binary perform a collection of measurements.The rotation’s Rz arguments can be changed at runtime, and the technique can be combined with readout symmetrization and correction.Source: Appendix A.2 · p. 16
Appendix A.3. The variational quantum eigensolver
The VQE example estimates the hydrogen molecule’s ground-state energy by preparing a parameterized UCC ansatz, measuring Hamiltonian terms, and feeding the resulting expectation value to a classical minimizer. The minimizer updates the variational parameter until convergence.
Appendix A.3. The variational quantum eigensolver
- VQE prepares a parameterized UCC ansatz from a Hartree-Fock reference state using the variational parameter θ.The paper identifies VQE as a leading variational hybrid algorithm for quantum chemistry and applies it to the H2 ground-state energy.Source: Appendix A.3 · pp. 17–18
- The H2 Hamiltonian is expressed as a sum of Pauli-basis expectation values with coefficients, including Z, ZZ, YY, and XX terms.Its stated form is H = g01 + g1Z0 + g2Z1 + g3Z0Z1 + g4Y0Y1 + g5X0X1.Source: Appendix A.3 · p. 18
- The expectation value ⟨H⟩ is fed to a classical minimizer, which selects the next θ until convergence.The example uses parametric compilation and pyQuil’s Experiment framework to simulate the H2 ground-state energy.Source: Appendix A.3 · p. 18