GPU-Accelerated PSO for High-Performance American Option Valuation

Using artificial intelligence tools to evaluate financial derivatives has become increasingly popular. PSO (particle swarm optimization) is one such tool. We present a comprehensive study of PSO for pricing American options on GPUs using OpenCL. PSO is an increasingly popular heuristic for financial parameter search; however, its high computational cost (especially for path-dependent derivatives) poses a challenge. We review PSO-based pricing and survey prior GPU acceleration efforts. We then describe our OpenCL optimization pipeline on an Apple M3 Max GPU (OpenCL 1.2 via PyOpenCL 2024.1). Starting from a NumPy baseline (36.7 s), we apply successive enhancements: an initial GPU offload (8.0 s), restructuring loops (forward/backward) to minimize divergence (2.3 s → 0.95 s), kernel fusion (0.94 s), and explicit SIMD vectorization (float4) (0.25 s). The fully fused float4 kernel achieves 0.246 s, a ~150X speedup over CPU. We analyzed all eight intermediate kernels (named by file), detailing techniques (memory coalescing, branch avoidance, etc.) and their effects on throughput. Our results exceed prior art in speed and vector efficiency, illustrating the power of combined OpenCL strategies.