[Paper Review] Physics-inspired optimization for constraint-satisfaction problems using a digital annealer.

Maliheh Aramon, Gili Rosenberg+3 more|arXiv (Cornell University)|Jun 22, 2018

Parallel Computing and Optimization Techniques8 citations

TL;DR

This paper presents the Fujitsu Digital Annealer, a specialized CMOS hardware accelerator that uses a physics-inspired optimization algorithm based on simulated annealing with parallel-trial and dynamic escape mechanisms to solve large-scale fully connected quadratic unconstrained binary optimization (QUBO) problems. It achieves a time-to-solution speedup of roughly two orders of magnitude over single-core simulated annealing and parallel tempering on fully connected spin-glass problems with bimodal and Gaussian couplings.

ABSTRACT

The Fujitsu Digital Annealer (DA) is designed to solve fully connected quadratic unconstrained binary optimization (QUBO) problems. It is implemented on application-specific CMOS hardware and currently solves problems of up to 1024 variables. The DA's algorithm is currently based on simulated annealing; however, it differs from it in its utilization of an efficient parallel-trial scheme and a dynamic escape mechanism. In addition, the DA exploits the massive parallelization that custom application-specific CMOS hardware allows. We compare the performance of the DA to simulated annealing and parallel tempering with isoenergetic cluster moves on two-dimensional and fully connected spin-glass problems with bimodal and Gaussian couplings. These represent the respective limits of sparse versus dense problems, as well as high-degeneracy versus low-degeneracy problems. Our results show that the DA currently exhibits a time-to-solution speedup of roughly two orders of magnitude for fully connected spin-glass problems with bimodal or Gaussian couplings, over the single-core implementations of simulated annealing and parallel tempering Monte Carlo used in this study. The DA does not appear to exhibit a speedup for sparse two-dimensional spin-glass problems, which we explain on theoretical grounds. We also benchmarked an early implementation of the Parallel Tempering DA. Our results suggest an improved scaling over the other algorithms for fully connected problems of average difficulty with bimodal disorder. The next generation of the DA is expected to be able to solve fully connected problems up to 8192 variables in size. This would enable the study of fundamental physics problems and industrial applications that were previously inaccessible using standard computing hardware or special-purpose quantum annealing machines.

Motivation & Objective

• To develop a hardware-accelerated solver for large-scale fully connected QUBO problems beyond the reach of classical algorithms.
• To improve optimization performance on dense, highly degenerate problems using parallelization and dynamic escape mechanisms.
• To benchmark the Digital Annealer against established Monte Carlo methods like simulated annealing and parallel tempering with isoenergetic cluster moves.
• To evaluate the scalability and performance of the Digital Annealer on both sparse (2D) and fully connected (3D) spin-glass models with different disorder types.
• To demonstrate the potential of the Digital Annealer for solving classically intractable problems in physics and industry.

Proposed method

• The Digital Annealer implements a variant of simulated annealing with a parallel-trial scheme that evaluates multiple candidate solutions simultaneously.
• It employs a dynamic escape mechanism to avoid getting trapped in local minima, enhancing convergence to global optima.
• The algorithm is executed on custom application-specific CMOS hardware, enabling massive parallelization across up to 1024 variables.
• Performance is evaluated on two-dimensional and fully connected spin-glass problems with bimodal and Gaussian couplings to represent sparse and dense problem regimes.
• The study compares the Digital Annealer to single-core implementations of simulated annealing and parallel tempering with isoenergetic cluster moves.
• An early version of the Parallel Tempering DA is also benchmarked to assess its scaling behavior on fully connected problems.

Experimental results

Research questions

• RQ1 Does the Digital Annealer achieve a significant speedup over classical single-core Monte Carlo methods on fully connected QUBO problems?
• RQ2 Why does the Digital Annealer not show a speedup on sparse two-dimensional spin-glass problems despite its parallel architecture?
• RQ3 How does the performance of the Parallel Tempering DA variant scale with problem size on fully connected problems with bimodal disorder?
• RQ4 What is the impact of problem structure—specifically sparsity versus full connectivity—on the effectiveness of the Digital Annealer?
• RQ5 Can the Digital Annealer solve problems of industrial and fundamental physics relevance that are infeasible for standard hardware or quantum annealers?

Key findings

• The Digital Annealer achieves a time-to-solution speedup of approximately two orders of magnitude over single-core simulated annealing and parallel tempering on fully connected spin-glass problems with bimodal and Gaussian couplings.
• The speedup is not observed on sparse two-dimensional spin-glass problems, which the authors explain through theoretical analysis of the problem’s energy landscape and algorithmic limitations.
• The early implementation of the Parallel Tempering DA shows improved scaling behavior on fully connected problems of average difficulty with bimodal disorder.
• The current Digital Annealer architecture supports problems of up to 1024 variables, with the next generation expected to scale to 8192 variables.
• The Digital Annealer enables the study of previously inaccessible fundamental physics problems and industrial applications due to its ability to handle large, fully connected QUBO instances.
• The results demonstrate that hardware-specific algorithmic enhancements like parallel-trial and dynamic escape are critical for performance gains on dense optimization problems.