Branch divergence occurs in parallel computing when different execution paths are taken by threads within the same warp, leading to inefficient use of resources. It is particularly problematic in architectures like GPUs, where threads are expected to execute the same instruction simultaneously for optimal performance.