This thesis presents an approach to optimizing the computation of soft shadows from area lights. The light source is sampled uniformly by traversing shadow rays as packets through an N-tree. This data structure stores an additional line space for every node. A line space stores precomputed information about geometry inside of shafts from one to another side of the node. This visibility information is used to terminate a ray. Additionally the graphics processing unit (short GPU) is used to speed up the computations through parallelism. The scene is rendered with OpenGL and the shadow value is computed on the GPU for each pixel. Evaluating the implementation shows a performance gain of 86% by comparison to the CPU, if using the GPU implementation. Using the line space instead of triangle intersections also increases the performance. The implementation provides good scaling with an increasing amount of triangles and has no visual disadvantages for many rays.