Plus the architecture improvements. Do we know how many cores per SM? They’ve decoupled int and FP execution units which could give larger increases for certain

Plus the architecture improvements. Do we know how many cores per SM? They’ve decoupled int and FP execution units which could give larger increases for certain kernels (although FP heavy deep learning kernels aren’t likely to benefit as much, they will still get address calculation benefits).