Custom instructions added to a processor architecture to accelerate specific tasks like matrix multiplication or neural network operations.