Acceleration strategy for the multicanonical Monte Carlo simulations in the two-dimensional Blume-Capel model

Abstract

We developed the code of the GPU multicanonical method for the two-dimensional Blume-Capel model. For the test of the validity of the code, we investigate the first order transition region, T = 0.5 , and the tricritical transition region , T = 0.609. We could show that the phase transition value of ∆ is matched well with previous literature’s results to the fourth decimal place in the first and the tricritical transition region. It only took 113 minutes for estimating the maximum point of the magnetic susceptibility, where the lattice size is 32 by 32. The speed comparison with GPU and CPU showed that using GPU is much more faster. The gap between the convergence time of using CPUs and that of using a GPU is exponentially increasing. The spin flip time of a GPU device is ten times faster than that of multi-core CPUs. We also find the efficient way estimating the initial weight in the Blume-Capel model. The final weight from the estimated initial weight is reliable and converged rapidly.