Firefly: A Versatile Experimental Platform for Oscillator-Based Ising Machines

Oscillator-based Ising machines (OIMs) are specialized in solving combinatorial optimization problems, that can be represented as the Ising model. They exploit the interaction of (integrated) electrical oscillators in a configurable network for the computation. Such systems naturally evolve towards a ground state, which forms a solution to the problem quickly and energy efficiently. This work presents the design of our 400 oscillator node chip in a 28nm technology. The focus is on the analog oscillator and coupler circuits, which determine the computing performance. Weighted optimization problems with up to 6-bit resolution can be solved within just 714ns. A comprehensive experimental analysis based on a versatile benchmark set is provided. We discuss the computation process and investigate the impact of multiple factors including the randomness of the initial oscillator phases, the frequency mismatch, the coupling strength, and the locking strength. A small range of parameters like the coupling strength and locking strength exists, which show the highest accuracy. Extensive benchmarks achieve an accuracy compared to the best-known solution of more than 94.5% for problems with equal weights and 89.8% for weighted problems. This emphasizes, that carefully designed oscillator-based Ising machines (OIMs) are not only fast, but can find solutions near the global optimum.