New GPU Optimizer Finds All Modes of Complex Functions Faster

A new GPU-native parallel optimizer, named Chisao (Convergence-Halt-Invert-Stick-And-Oscillate), has been developed to tackle the challenging problem of finding all modes within multimodal black-box functions. Unlike existing sequential methods, Chisao leverages the massive parallelism of modern GPUs by processing an entire sample batch simultaneously. Its core innovation lies in a deliberate oscillation cycle between convergence and anticonvergence, which allows it to escape local optima while preserving already identified modes. The algorithm employs an asymmetric structural move: samples that locate true peaks are "frozen" and retained, while others continue exploration through momentum-based anticonvergence and stochastically smoothed gradients. Population diversity is maintained via adaptive reseeding strategies called Repulse Monkey and Golden Rooster. Benchmarking across 42 functions from the Simon Fraser University suite showed Chisao achieving 100% mode recovery, even where CPU-based methods failed at higher dimensions, and delivering up to 39x speedup.