The Inverse Synthetic Aperture Radar (ISAR) technique serves to identify and/or classify targets such as aircraft, ships and ground vehicles. In the context of radar operator training, ISAR simulation is a complex CPU-intensive process since it needs to compute and update a target's range-Doppler signature for constantly varying aspect angles. The main challenge is to generate a reliable ISAR image in real-time. To do so, current simulators employ several approximations. These commonly consist of (1) computing only direct reflections and tuning manually some reflection features by editing the 3D models, (2) coarsening the target mesh to reduce the number of intersected polygons or (3) undersampling the integration time by ignoring intermediate aspect angles. In this work, we present a solution where we implement a modified visual ray tracing method as an analogy to simulate the radar wave scattering. The objective is to improve real-time simulation of the ISAR imagery for the purpose of radar operator training. The method presented here has the ability to compute multiple reflections. Consequently, intense flashes produced by corner reflectors are naturally depicted. The method also allows using complex 3D models directly without offline preprocessing or manual modifications of the models. Real time is achieved by implementing a ray tracing-based algorithm written in CUDA running on GPU. This takes advantage of the massive number of parallel threads that can run on current GPUs. Moreover, the general-purpose programming model supported by CUDA offers a more flexible implementation with more appropriate data structures. The proposed solution can generate realistic ISAR images of a ship at sea including effects of multiple reflections on a mid-range graphics card. Since this solution does not require preprocessing or manual alteration of the 3D models to add scatterers, it makes training on realistic ISAR simulation more accessible.