The characteristics of discrete fractures, including their small volumes, high permeability, and intricate connectivity, significantly hinder the computational efficiency of reservoir numerical simulations. Conventional reservoir simulators are typically limited to models with a modest number of fractures. This paper introduces a new generation of reservoir simulation software designed for GPUs, leveraging advanced GPU architectures and Single Instruction Multiple Data (SIMD) programming paradigms to enhance computational efficiency and scalability. This enables the simulation of large-scale embedded discrete fracture models on a standalone personal computer (PC). The simulator encompasses core functionalities such as Jacobian matrix construction, linear and nonlinear solvers. To further boost computational efficiency while preserving solution accuracy, hybrid precision computing strategies are employed in critical processes like derivative calculations. Extensive benchmark testing was conducted across various reservoir models, ranging from SPE10 standard cases with differing fracture densities to an ultra-large real gas reservoir model containing a million discrete fractures. Tests were carried out on a parallel architecture equipped with two NVIDIA RTX 4090 GPUs, with performance comparisons made against a 52-core dual-socket Intel Xeon Platinum 8272CL CPU platform. Results indicate that the GPU-based simulator completes simulations of models with millions of fractures in 123 min, markedly outperforming CPU-based systems in both computation time and energy consumption. Thus, GPU-parallelized architectures demonstrate substantial advantages, showcasing the capability to simulate ultra-large-scale embedded discrete fracture models on a standalone system.