With cuDNN, some expert had to sit down and write an optimized kernel for your ops. With TC, the idea is you write the simple, tensor comprehension for your op,

With cuDNN, some expert had to sit down and write an optimized kernel for your ops. With TC, the idea is you write the simple, tensor comprehension for your op, and then throw the autotuner at it, and get something that approaches or even beats the hand-tuned example. The diagrams in the paper give a sense for some of these situations.