Overcoming the Accuracy vs. Performance Trade-off in Oscillator Ising Machines

Coupled oscillator-based Ising machines offer an efficient stochastic search-based approach to solve NP-hard combinatorial optimization problems (Maximum Cut problem considered here) that are computationally expensive to solve using traditional computers. Here, we measure a 600 oscillator-based Ising machine with >29,000 programmable coupling elements. Using this platform, we experimentally reveal a fundamental trade-off between the solution quality and the time-to-compute, which can limit the accuracy & performance of the Ising machine in comparison to digital algorithms. Furthermore, we also show that the solution is highly sensitive to the size and the sparsity of graph. To overcome these limitations, we propose a hybrid approach that uses the oscillator Ising machine to obtain near-optimal solutions (extremely fast), which are then subsequently improved using a simple local-search algorithm with minimum time penalty. Our hybrid approach: (i) produces better quality solutions than stand-alone oscillator-based hardware in a given time; (ii) shows 3-100x improvement in experimentally measured time-to-compute at iso-solution quality, when compared to a digital algorithm (manifold optimization, executed on a 32-core processor with 256GB RAM); (iii) shows minimal sensitivity (<2%) in performance to the input properties of the graph. Our work provides a pathway to overcome the performance vs. quality tradeoff in oscillator Ising machines, and subsequently, accelerate hard combinatorial problems while maintaining high accuracy.