| (in English) |
We have recently proposed a heuristic algorithm for combinatorial optimization, which is based on the numerical simulation of classical Hamiltonian systems exhibiting bifurcations. We call it “Simulated Bifurcation (SB) algorithm,” and refer to Ising machines based on SB as SB machines (SBMs). Unlike simulated annealing (SA), which is based on sequential update determining whether a spin should be flipped or not, in SB all the positions and all the momenta, respectively, can be updated in parallel, and therefore SB possesses high parallelizability. Harnessing this advantage of SB, we can realize ultrafast and ultralarge Ising machines using conventional digital computers. In fact, we have realized an SBM using a single FPGA for an all-to-all connected 2000-spin Ising problem, which is about ten times faster than a state-of-the-art coherent Ising machine using optical parametric oscillators. We have also realized an all-to-all connected 100000-spin SBM using a GPU cluster, which is about 1000 times faster than optimized SA software performed using a single CPU core. While the SBMs are classical Ising machines, they originate from our proposed quantum computer named “Quantum bifurcation Machine (QbM),” and therefore the SB is related to quantum omputating. Here we present the principle and performance of our SBMs after introducing the QbM and deriving the SB algorithm. |