Comment by ap4

It is apples-to-apples. The code from the article runs a chosen kernel 50 times in a loop. In the case of CuBLAS it runs the function cublasGemmEx() 50 times.