Solving computationally demanding combinatorial optimization problems using conventional computing architectures is slow and energy intensive. Quantum computing could improve optimization efficiency but remains at an early stage. Probabilistic computing offers a practical near-term approach to faster optimization through stochastic techniques. Here, we experimentally demonstrate a scalable spin-transfer-torque-magnetic-tunnel-junction based probabilistic processor for efficiently solving all-to-all connected quadratic assignment problems. Our system integrates 144 compact spintronics tunable random number generators with a massively parallel architecture, achieving a high Monte-Carlo sampling throughput of 14.4 million flips per second. We co-design a parallel trial annealing scheme, and the integrated system achieves a 123× speedup with 98.3% energy savings over conventional Gibbs sampling, and a 3.2× speedup with 58.3% energy savings relative to the central processing unit implementation based on a compiled language. We further benchmark performance across graphics processing unit, and D-Wave quantum annealers, showing gains in solution quality, speed, and energy efficiency. Combinatorial optimization challenges conventional computing due to high computational costs. Here, Yang et al. demonstrate a probabilistic processor for quadratic assignment problems, achieving fast and energy-efficient stochastic optimization.
Yang et al. (Mon,) studied this question.