2 ms·

block level rather than thread level programming, automatic optimization across hyperparameters, makes it much easier to write fast kernels