Fujitsu Digital Annealer boosts industrial solver selection, study finds

A head-to-head benchmark of industrial scheduling problems reveals the potential of the Fujitsu Digital Annealer alongside established quantum platforms from IBM Quantum and D-Wave, according to a new study. Researchers rigorously tested these approaches and a classical solver, using a real-world variant of the Job-Shop Scheduling Problem provided by Siemens AG, focusing on practical utility rather than simply claiming speedups. The work demonstrates that hardware-software co-design is essential for both solution quality and scalability, a finding that highlights the need for tailored formulations alongside quantum hardware. This study serves as a benchmark of current performance and a first step toward development and preparation of future industrial optimization pipelines, suggesting a path toward improved approximations for complex industrial scheduling challenges.

Job-Shop Scheduling Problem as an Industrial Optimization Challenge

The relentless pursuit of optimization in industrial logistics has led researchers to increasingly explore the potential of quantum and quantum-inspired computing, yet practical utility remains a critical hurdle. The work compares three distinct platforms: IBM Quantum, the D-Wave Quantum Annealer, and the Fujitsu Digital Annealer, highlighting a quantum-inspired approach alongside established quantum contenders. The researchers demonstrate that tailoring problem formulations to the specific constraints of each platform is paramount, exposing a trade-off between modelling simplicity, constraint density, and the inherent limitations of the hardware itself. They benchmarked all approaches against an exact classical solver and a Mixed-Integer Linear Programming (MILP) formulation, establishing a rigorous comparative framework that extends beyond mere claims of acceleration and focuses on practical applicability. The study employed two Quadratic Unconstrained Binary Optimization (QUBO) formulations, a Single-Constraint Model and a Multi-Constraint Model, to systematically investigate this interplay.

Results revealed that the number of constraints and the degree to which the formulation aligns with the hardware’s architecture significantly impact both runtime and the quality of the solutions obtained across all tested platforms. This suggests that a deep understanding of the quantum stack, from algorithm design to device-level execution, is vital for achieving meaningful results. The researchers emphasize that hardware performance cannot be separated from modelling choices and device characteristics, advocating for systematic design-space exploration.

The pursuit of quantum solutions for complex industrial problems increasingly focuses on the interplay between algorithm design and underlying hardware. Recent work demonstrates that simply translating a classical optimization problem into a quantum format isn’t sufficient; the method of encoding, specifically the choice of QUBO formulation, profoundly impacts performance. Researchers have been meticulously comparing different QUBO approaches, notably the Single-Constraint Model and the Multi-Constraint Model, to understand how these choices interact with the capabilities of diverse quantum and quantum-inspired platforms. The study detailed a comparison of modelling trade-offs. The Single-Constraint Model prioritizes compactness, aiming for a streamlined representation of the Job-Shop Scheduling Problem. Conversely, the Multi-Constraint Model introduces additional constraints to more accurately reflect the intricate dependencies inherent in scheduling tasks, such as sequencing and resource allocation. This distinction is critical because the number and structure of these constraints directly influence how effectively a given quantum or quantum-inspired device can process the problem.

Fujitsu’s Digital Annealer was included in a rigorous industrial benchmark alongside established quantum players IBM Quantum and D-Wave Quantum Annealer, offering a comparative analysis of these platforms. The choice of models proved critical, revealing that the number and structure of constraints significantly impacted runtime and solution quality across all tested hardware. The study compared the performance of the Fujitsu Digital Annealer with that of gate-based and dedicated quantum annealers, highlighting the nuances within the broader category of non-universal quantum computation. Results indicated that quantum and quantum-inspired optimization can support industrial solver selection, integration in classical workflows, modelling decisions, and early proof-of-concept development, but only when carefully tailored to the specific hardware. The study’s focus on an industrial variant of the JSSP, provided by Siemens AG, allowed for evaluation of end-to-end performance and practical scheduling solutions, moving beyond theoretical speedups to assess real-world applicability.

Establishing reliable classical benchmarks proved crucial for evaluating the performance of quantum and quantum-inspired approaches to the Job-Shop Scheduling Problem. The researchers deliberately developed an exact classical solver for smaller instances, alongside a Mixed-Integer Linear Programming (MILP) model designed for larger, more practical problem sizes. While MILP offers theoretical exactness through techniques like branch-and-bound and branch-and-cut, computational limitations necessitated solving larger instances within a fixed time budget, using the best solution found as an approximation, a pragmatic approach for real-world comparison. This methodology provided a scalable and practically relevant yardstick against which to measure the quantum methods. The team’s work extends beyond simply comparing algorithms; it provides insights into how problem encodings and solver paradigms affect end-to-end performance, offering transferable design considerations for near-term optimization.

While quantum computing often promises speedups, practical gains hinge on a more nuanced interplay between algorithms and the underlying hardware than is commonly appreciated. This comparative analysis goes beyond simply claiming quantum advantages; it focuses on practical utility for industrial applications. This paper is augmented by a reproduction package, which is permanently archived on Zenodo.