Abstract: We develop a GPU-accelerated and compressed hierarchical Schur algorithm for finite-difference frequency-domain photonic device analysis. The compressed hierarchical Schur (CHiS) method is a customized sparse linear system solver which is inspired from common features in photonic structure design, numerical method, and modern HPC architectures. Several numerical tests show significant memory and time saving. When performing GPU acceleration, several modifications are investigated. In particular, proper workload distribution is a critical factor for efficient multi-GPU acceleration. Numerical results using quad-GPU demonstrates over 6X speedup compared to CPU-only simulation, and good multi-GPU scaling with our GPU-accelerated CHiS code.