COCOSSim is a cycle-accurate simulator designed to evaluate the performance of modern deep-learning workloads on hardware accelerators. The simulator integrates systolic arrays, vector units, and realistic memory models to provide an in-depth analysis of the performance of deep learning models.