I'm aware (excruciatingly aware...) of that. Our numbers are ~10X better despite a process disadvantage.

And as always, Nvidia is twiddling the numbers, they exclude the cost of memory access when calculating 1 top/w, whereas we include it.