Parallel Quantum Annealing Techniques

• Parallel Quantum Annealing is the method of embedding multiple independent QUBO or Ising formulations onto disjoint regions of a quantum annealer to enable simultaneous anneals.
• It employs strategies like same-problem replication, independent task execution, and integration in hybrid workflows (e.g., supervised Quantum Boltzmann Machine training) to maximize hardware utilization.
• Empirical studies demonstrate significant throughput gains and improved sample quality, although challenges remain in embedding efficiency and per-instance parameter tuning.

Parallel quantum annealing denotes, in the D-Wave-centered literature, the use of a single quantum annealer to solve multiple independent QUBO or Ising formulations—or multiple isolated copies of one formulation—within one annealing cycle by embedding them onto disjoint regions of the hardware graph (Pelofske et al., 2021). Its immediate motivation is hardware underutilization: minor embedding onto sparse topologies such as Chimera or Pegasus often leaves many qubits idle, especially for small and medium-sized logical problems. The contemporary literature treats this as a throughput problem rather than a new annealing law: one anneal can be made to return several logically independent samples at once, provided the embeddings are disjoint and the coefficient ranges remain compatible. Within that core meaning, the subject has developed along three main lines: same-problem replication to improve success probability, simultaneous execution of many independent tasks, and application-specific integration into hybrid workflows such as decomposition-based optimization and supervised Quantum Boltzmann Machine training (Schuman et al., 18 Jul 2025).

1. Definition and scope

In its strict hardware sense, parallel quantum annealing is not software-level batching, queue parallelism, or post hoc aggregation of separate jobs. It is single-cycle hardware parallelism achieved by placing several embeddings on one chip and annealing them simultaneously (Schuman et al., 18 Jul 2025). The original formulation explicitly allows two modes: solving several independent problems at once, or solving the same problem multiple times in parallel on disjoint embeddings (Pelofske et al., 2021).

A mathematically convenient description uses a block-diagonal composition of independent QUBOs. If the binary variables are concatenated as $X=[X_1\;X_2\;\dots]$, then the combined objective is

$X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$

so the global minimizer is obtained by concatenating the minimizers of the independent subproblems (Bishwas et al., 2024). This construction is central to simultaneous multi-problem execution, because it preserves logical independence while enabling a single physical anneal.

A more refined formulation appears in multitasking quantum annealing, where different problems are embedded into spatially distinct regions of one QPU, but with per-instance scaling, per-instance chain strengths, and per-instance unembedding rather than one global parameterization (Artag et al., 10 Mar 2026). This distinguishes a mature multitasking formulation from simpler packing schemes that merely average parameters across all co-scheduled instances.

The literature also uses the phrase more loosely in adjacent contexts. Repeated short nonadiabatic anneals can be interpreted as parallelizable independent trials (Katsuda et al., 2013), and local-search annealing can be embedded into replica-based schemes analogous to parallel tempering or population annealing (Chancellor, 2016). These are conceptually related, but they are not the same as multi-embedding several QUBOs on one annealer.

2. Physical and mathematical basis

The hardware-level implementations are built on the standard transverse-field Ising annealing Hamiltonian. One representative formulation is

$H(t)=A(t)H_D + B(t)H_P,$

with

$H_D = -\sum_{i=1}^{N}\sigma_i^x$

and

$$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$

Logical QUBO or Ising coefficients are mapped into $h_i$ and $J_{ij}$, then minor-embedded into the hardware graph (Schuman et al., 18 Jul 2025).

For independent multitasking, the idealized Hamiltonian is block diagonal. A representative two-task form is

$$H_{\mathrm{different}} = \begin{bmatrix} H_{\mathrm{MVCP}} & 0 \ 0 & H_{\mathrm{GPP}} \end{bmatrix},$$

for which the combined ground-state energy is additive and the minimum spectral gap is the smaller constituent gap under exact isolation (Artag et al., 10 Mar 2026). This theoretical argument supports the intuition that packing independent problems does not, by itself, make the hardest constituent problem harder. The caveat is equally explicit: actual hardware deviates from perfect block diagonality because of analog noise, calibration error, residual couplings, and integrated control errors.

Performance is usually discussed through time-to-solution or closely related throughput measures. In the early PQA work, ensemble TTS is defined from QPU time, unembedding time, anneals per call, and empirical ground-state probability across the packed problems (Pelofske et al., 2021). Later work on supervised QBM training instead reports direct QPU time expenditure and interprets the gain as a sampling-throughput improvement rather than an intrinsic change in learning dynamics (Schuman et al., 18 Jul 2025).

3. Embedding strategies and hardware topology

The main technical difficulty is not the abstract composition of independent QUBOs but their simultaneous minor embedding onto a sparse physical graph. The original Maximum Clique study precomputes as many disjoint clique embeddings $K_N$ as possible on Chimera or Pegasus, using minorminer on the disjoint union of the desired cliques (Pelofske et al., 2021). This is explicitly heuristic and sub-optimal, because it does not optimize spatial separation between embeddings.

Later work places far more emphasis on spatial isolation. In supervised QBM training, Pegasus is partitioned into ten subgraphs with PyMetis; nodes linking one subgraph to another are removed to create buffer zones; and minorminer is run independently inside each subgraph (Schuman et al., 18 Jul 2025). The result is ten physically separated embeddings of the same logical model, so one anneal returns ten samples at once. The stated trade-off is clear: buffer zones reduce inter-instance interference or “leakage,” but they also consume qubits and reduce packing density.

The multitasking formulation generalizes this isolation idea to heterogeneous workloads. It compares dense packing without isolation to an isolation-layer strategy that removes both occupied qubits and their neighbors during iterative embedding search (Artag et al., 10 Mar 2026). It also replaces global chain-strength selection with per-instance rules. For MVCP it uses uniform torque compensation,

$J_{\mathrm{base},\mathrm{MVCP}} = \sqrt{\frac{\sum J_{ij}^{2}}{N_J}}\sqrt{\bar d}, \qquad J_{\mathrm{chain},\mathrm{MTQA}_{\mathrm{MVCP}} = \alpha_{\mathrm{MVCP}} J_{\mathrm{base},\mathrm{MVCP}},$

with $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$0; for GPP it uses

$X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$1

with $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$2 (Artag et al., 10 Mar 2026). The point is not merely improved embedding success; it is preservation of per-instance Hamiltonian scale under co-scheduling.

This embedding literature establishes a recurring design tension. Maximizing the number of packed instances is not obviously optimal on current annealers, because denser packing can degrade sample quality. The practical art of parallel quantum annealing therefore lies in balancing chain length, physical separation, coefficient scaling, and unembedding overhead.

4. Principal application patterns

A first major application is Maximum Clique. The QUBO used throughout the cited work is

$X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$3

with $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$4 and $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$5, so the linear term rewards larger cliques while the quadratic term penalizes selected nonedges (Pelofske et al., 2022). In the direct PQA study, the same logical clique problem is either replicated several times or several independent clique instances are packed together (Pelofske et al., 2021). In the decomposition-based variant, DBK recursively reduces a larger graph into leaf subproblems and then solves each leaf using multiple simultaneous hardware replicas, allowing graphs with up to $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$6 vertices and $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$7 edges to be solved through a hybrid exact-decomposition-plus-quantum-subsolver workflow (Pelofske et al., 2022).

A second application is supervised QBM training for medical image classification. Here parallel annealing is used to mitigate the dominant cost of repeated positive-phase and negative-phase sampling calls during discriminative training (Schuman et al., 18 Jul 2025). The visible layer is split into input units and label units, but the key architectural observation is that clamped input units need not be represented on the QPU. Their contribution is absorbed into effective biases,

$X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$8

so only hidden and label units are embedded. This permits the same logical QBM core to be replicated ten times on Pegasus subgraphs and sampled in parallel. The formulation is especially well suited to supervised discriminative QBMs because the $X^T Q X = \sum_{i=1}^{n} X_i^T Q_i X_i,$9-pixel input remains classical while the embedded quantum core stays small (Schuman et al., 18 Jul 2025).

A third pattern is simultaneous multitasking across heterogeneous instances. One study combines Asset Liability Modelling and Traffic Flow Optimization by forming a block-diagonal QUBO and sending it to D-Wave samplers or to the Leap hybrid solver (Bishwas et al., 2024). Another integrates heterogeneous graph problems such as MVCP and GPP with per-instance scaling and chain strength, explicitly positioning this as a refinement over simpler global-parameter PQA (Artag et al., 10 Mar 2026). This suggests that, as the field matured, the question shifted from whether several tasks can be co-scheduled to how task-specific parameterization must be preserved for co-scheduling to remain reliable.

5. Empirical performance and throughput gains

The earliest direct PQA study reports that, for multiple independent Maximum Clique problems solved in parallel, average TTS speedups are about $H(t)=A(t)H_D + B(t)H_P,$0 on D-Wave 2000Q and about $H(t)=A(t)H_D + B(t)H_P,$1 on D-Wave Advantage relative to sequential execution (Pelofske et al., 2021). For a single problem replicated many times on one anneal, the same study reports roughly $H(t)=A(t)H_D + B(t)H_P,$2 average TTS improvement and $H(t)=A(t)H_D + B(t)H_P,$3 average ground-state probability increase on size-$H(t)=A(t)H_D + B(t)H_P,$4 Maximum Clique instances on Advantage (Pelofske et al., 2021). The same work is explicit that per-problem ground-state probability may decrease slightly under parallel execution, but throughput gains dominate.

In decomposition-driven Maximum Clique, the benefit is conditional on cutoff and graph density. The DBK-plus-parallel-QA workflow is reported to outperform the classical FMC solver by up to around two orders of magnitude for high-density graphs and low cutoffs around $H(t)=A(t)H_D + B(t)H_P,$5 or $H(t)=A(t)H_D + B(t)H_P,$6; in a real-time experiment on the first $H(t)=A(t)H_D + B(t)H_P,$7 random graphs, repeated $H(t)=A(t)H_D + B(t)H_P,$8 times each, all $H(t)=A(t)H_D + B(t)H_P,$9 runs found the maximum clique (Pelofske et al., 2022). These results do not show a blanket advantage for annealing; they show that decomposition and parallel replication can create a favorable regime for current hardware.

In supervised QBM training, the central empirical result is different in kind: parallel annealing yields a speed-up of $H_D = -\sum_{i=1}^{N}\sigma_i^x$0 in QPU time expenditure relative to regular sequential QA for a fixed workload of $H_D = -\sum_{i=1}^{N}\sigma_i^x$1 mini-batches, $H_D = -\sum_{i=1}^{N}\sigma_i^x$2 data points each, and $H_D = -\sum_{i=1}^{N}\sigma_i^x$3 samples for both clamped and unclamped phases (Schuman et al., 18 Jul 2025). The paper is careful to identify this as a sampling-cost reduction. It is not a claim that PQA intrinsically reduces the number of epochs to convergence.

The multitasking study on MVCP and GPP shows why parameter handling matters. For MVCP, MTQA with per-instance control and standard single-instance QA both maintain high ground-state probability, above $H_D = -\sum_{i=1}^{N}\sigma_i^x$4 across most sizes up to $H_D = -\sum_{i=1}^{N}\sigma_i^x$5 nodes, whereas the simpler global-parameter PQA baseline drops from $H_D = -\sum_{i=1}^{N}\sigma_i^x$6 at $H_D = -\sum_{i=1}^{N}\sigma_i^x$7 to about $H_D = -\sum_{i=1}^{N}\sigma_i^x$8 at $H_D = -\sum_{i=1}^{N}\sigma_i^x$9 and effectively zero for all $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$0 (Artag et al., 10 Mar 2026). For GPP, by contrast, MTQA, PQA, QA, and SA remain broadly comparable in solution quality, implying that global parameterization can be catastrophic for some classes and tolerable for others.

Across these results, a single pattern recurs: parallel quantum annealing is primarily a hardware-utilization and throughput strategy. Gains usually arise because one programming cycle produces more useful sample streams, not because the annealing physics of each individual logical problem has been fundamentally improved.

6. Related extensions, indirect uses, and quantum-inspired analogues

A broader conceptual reading of parallel quantum annealing appears in nonadiabatic repetition. A short anneal of duration $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$1 has success probability $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$2; repeating $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$3 independent trials in parallel yields success probability $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$4, so the sequential cost $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$5 can be reinterpreted as wall-clock time $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$6 with parallel width $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$7 (Katsuda et al., 2013). This is not multi-embedding on one QPU, but it supplies a theoretical template for trading coherent anneal duration against independent parallel trials.

A different extension treats the annealer as a local-search primitive inside replica-based metaheuristics. Reverse or partial anneals around user-selected classical states can be organized into quantum analogues of parallel tempering and population annealing, where the parallelism is algorithmic rather than hardware-simultaneous (Chancellor, 2016). This line of work suggests that “parallel quantum annealing” can also mean ensemble control over many replica states, even when the device executes calls sequentially.

Several papers are explicitly adjacent but not direct contributions to hardware PQA. Discretised quantum annealing on gate-model hardware exposes natural parallelism through independent shots, parallel truncation-depth sweeps, and gate-level parallel scheduling, but it does not define a formal parallel annealing method (Bhave et al., 2023). Classical, quantum-inspired methods such as Parallel Quasi-Quantum Annealing couple many GPU-resident replicas through a diversity term, while Simulated Bifurcation Quantum Annealing introduces inter-replica couplings inspired by the Suzuki–Trotter mapping of a transverse-field Ising model (Ichikawa et al., 2024); (Pawłowski et al., 1 Apr 2026). These methods are relevant as analogues and baselines, not as direct realizations of quantum annealing on annealing hardware.

7. Limitations, controversies, and open questions

The most persistent limitation is embedding. Parallel annealing only works when several logical problems can be minor-embedded simultaneously, and the required chains can be long even for a single dense problem. This is why the decomposition literature keeps leaf problems small, why QBM studies restrict the free quantum core to at most $$H_P = -\sum_{i=1}^{N} h_i \sigma_i^z -\sum_{\substack{i,j=1\ i<j}}^{N} J_{ij}\sigma_i^z\sigma_j^z .$$8 fully connected logical variables, and why direct QA for power-grid partitioning becomes embedding-limited well below the size of realistic transmission networks (Pelofske et al., 2022); (Schuman et al., 18 Jul 2025); (Hartmann et al., 2024).

A second limitation is parameter heterogeneity. When several co-scheduled problems differ strongly in coefficient scale or embedding characteristics, global scaling and one averaged chain strength can distort the effective Hamiltonians. The ALM-plus-TFO study finds that many normalization schemes fail to improve the quality of pure-QPU parallel runs, while the multitasking study shows that global-parameter PQA can collapse entirely on MVCP (Bishwas et al., 2024); (Artag et al., 10 Mar 2026). This controversy is not about whether multiple instances fit on the chip; it is about whether they remain well posed once jointly parameterized.

A third limitation is the tension between isolation and capacity. Buffer zones, isolation layers, and disjoint subgraph constructions are introduced to reduce leakage, residual couplings, and integrated control errors, but they necessarily waste qubits (Schuman et al., 18 Jul 2025). This suggests that future gains may depend as much on hardware topology and calibration as on annealing methodology.

A fourth issue is evidentiary. Some papers argue for isolation strategies through design rationale and prior warnings about leakage but do not provide exhaustive sample-quality metrics against denser packings (Schuman et al., 18 Jul 2025). The block-diagonal spectral arguments for multitasking are likewise conditional on exact problem isolation and are supported only on small idealized systems (Artag et al., 10 Mar 2026). This suggests that the central open questions are experimental: how far packing density can be increased before interference becomes dominant, how embedding position affects success probability, and how per-instance calibration should scale with larger and more heterogeneous workloads.

Finally, parallel quantum annealing should not be conflated with quantum advantage. The direct literature consistently frames it as a practical method for increasing samples per anneal, amortizing control overhead, and exploiting otherwise idle hardware. Its strongest current role is architectural and operational: it turns a sparse, moderately sized annealer from a one-problem-at-a-time device into a multitasking or replicated-sampling device, with performance determined by embedding quality, parameter control, and the structure of the application workload.