Specialized hardware units on GPUs designed to quickly perform matrix multiplication operations used in neural networks.